Subscribe now

I’ve been obsessed with the framework from Ramp’s CEO, Eric Glyman, where he distilled their entire business into one equation. Once he had it, he could model which variable actually moved the needle. It turned out to be purchase volume. So, he pointed the whole company at that.

I wanted to do the same for OffDeal.

Most people assume that service businesses, especially “high touch” ones like investment banking, are unscalable. They’re seen as “black boxes” where revenue is just a function of deep relationships and bespoke financial advice.

We take the other side: we believe investment banking is a system that can be deconstructed, measured, and optimized - and software engineers are uniquely positioned to do this exceptionally well.

Deconstructing the Bank

At the highest level, our business is simple:

But that’s not very useful. It masks all the variables that matter. So, let’s “double-click” on the components to find the real levers.

Total Deals Closed formula is pretty straightforward:

But where do Total Mandates Signed come from? Our origination funnel.

But wait! Where do our Engaged Leads come from?

Even Leads Identified can be further broken down into (Total TAM) x (% Identifiable).

We can also use the same approach to deconstruct the system’s throughput as well:

The point is, you can do this for every part of the business. This allows you to trace exactly how a tiny 5% improvement in a given variable ripples all the way up to Annual Revenue.

Finding the 10x Levers

Just having this map isn’t enough. It tells us what the variables are, but not which ones to focus on. To do that, we analyze the “Leverage Headroom” for each one.

For every metric in our system, we ask:

1. What is our Baseline (e.g., 50% Show Rate)?

2. What is the Practical Ceiling (e.g., 80% Show Rate)?

The gap between them shows the total leverage up for grabs. This simple analysis immediately shows us where the 10x opportunities are hiding.

For example:

• If our Show Rate is 50%, the absolute maximum we can improve is 2x (to 100%). The practical ceiling is even lower, maybe 1.5x (75%).

• But if our % Qualified (leads that actually match our ICP) is only 10%, we have a theoretical 10x (100%) of headroom to work with.

In this instance, this analysis tells us that building AI to find better-qualified leads (a 10x problem) is a far higher-impact use of resources than just optimizing reminder emails (a 1.5x problem).

Our Optimization Blueprint

When we put this all together, we get a blueprint for the entire business - a robust framework to align on internal prioritization and where to deploy engineering resources. This table shows an illustrative example of how we view our optimization stack.

While these specific numbers are illustrative, our systems thinking approach illustrates a few key points:

1. Technology drives a powerful compounding effect. Improvements aren’t just additive; they multiply. A 2x improvement in Deal Cycle time and a 2x improvement in Deal Success Rate don’t just add up - they result in a 4x gain for the overall system. This is how you find 10x gains - by stacking multiple 2x and 3x levers.

2. Systems thinking enables ruthless prioritization. To be sure - we will work really hard on improving all of those variables. But practically speaking, as a resource-constrained startup, we need to deploy resources into the higher ROI projects. Our approach allows us to make calculated, data-backed bets on which engineering project will drive the highest impact.

3. This is all possible before traditional scaling. All of this leverage is unlocked without hiring a single extra banker or chasing bigger deals. These variables improve the engine itself.

In closing...

These are just a subset of variables, of course - we can do this mapping exercise across every single performance driver. These advantages compound very quickly.

It’s been so fun (though admittedly, hard) building out OffDeal. We see so much potential and have already realized insane productivity gains - all from allowing our engineers to re-architect every investment banking workflow from first principles.

If building applied AI solutions to real-world problems sounds interesting to you, we’d love to hear from you - we’re always hiring talented engineers to help us build the future of finance.

Send me a DM if you’d like to jam!

Ori Eldarov

As always, pushback and feedback always welcome. Twitter DMs always open @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI and business.

Subscribe now

Subscribe now

Last year I set out to build a new investment bank, OffDeal.

Not software for banks or a fintech marketplace—a real M&A advisory firm.

When we were raising our $4.7 million seed, the most common pushback was, “Why not just sell software to Goldman Sachs?” Fair question: in our YC batch we were the only team building a service company.

This essay explains why we chose the harder path—to build a full-stack AI investment bank to compete with the large incumbents, rather than sell SaaS directly to them. The thesis is simple: M&A will always matter, but legacy banks are structurally unable to run AI-native workflows, and therefore will be unable to capture the full upside.

We’ll cover:

1. Why investment banking is a solid business.

2. Why incumbents can’t reinvent it.

3. How OffDeal turns AI into client value.

Let’s dig in.

I Investment Banking Is a Solid Business

The first notable M&A advisory mandate dates back to 1895, when JP Morgan advised on a $1.4 billion merger of Federal Steel Company and Carnegie Steel ($45 billion in today’s terms!), which led to the formation of U.S. Steel in 1901. Since then, global M&A has grown to trillions a year, peaking at $5.8 trillion in 2021; advisory fees now run $30–40 billion annually.

But what underlying factors make this industry such an enduring business?

Timeless need → structural demand

Jeff Bezos says it’s easier to bet on what won’t change.

In M&A that constant is clear: sellers of businesses want the highest price, best terms, least hassle—and they always will. No one has ever called Goldman Sachs or J.P. Morgan to complain they got “too much” for their business. Because these motives are permanent, banks could invest heavily in this area, confident the product would stay valuable a century later.

Demand for financial advisory is structural, not cyclical.

Third-Party Validation

Even sophisticated market participants want a neutral set of eyes when selling or buying a business. Private-equity firms—despite employing ex-bankers—still bring in another advisor because a true third party can run a cleaner auction and nudge bidders higher. Boards lean on a bank’s fairness opinion to meet fiduciary duties; founders like hearing an outsider confirm the price. Independence gives the process credibility and usually bumps the final number.

Revenue Density

M&A fees are a slice of deal value, not hours worked. That’s why top firms clear $1 million-plus per employee - numbers other professional services industries like consulting and accounting rarely see.

Even so, top-line productivity hides a margin problem we’ll address next.

II Investment Banks Are Inefficient

Despite enviable revenue per employee, banks are actually very inefficient. For every “revenue generating” dealmaker (Managing Director) there are ~10 supporting staff. The business model is so labor intensive that compensation cost alone absorbs 60–70 % of revenue. Add revenue-per-MD and operating margin and the story changes:

Part of the problem stems from the business model itself - success fees pay only on closed deals, but a much larger contributor is simply how the actual work at banks gets done.

• Waterfall handoffs: Instructions flow MD → Director → VP → associate/analyst, and then gradually back up the chain. Each handoff introduces potential for misinterpretation. Two weeks in, the angle shifts and the deck restarts at page one. Thirty iterations per deck is common.

• Missing context: Client calls aren’t transcribed, key emails omit half the team, side-bar chats buried in Slack/Teams. An MD who decides to change a proposed financing structure over lunch may only mention it to the deal team the night before the meeting, triggering a last-minute rebuild.

• Manual & duplicative work: Even with Bloomberg, CapIQ, FactSet, teams still copy data into Excel, then re-cut it for each MD. Much of it duplicates work buried in Pitch_v32_FINAL2_FF -often faster to rebuild from scratch

• Proof-of-work incentives: Page count is treated as a proxy for quality and effort. VPs add slides to impress MDs; MDs add slides to impress clients. In the end, a sixty-page deck is produced.

Taken together, these frictions explain why every deliverable consumes so many staff hours - and why labor cost absorbs most of the revenue. The root cause for the most part is cultural, not technical, which is where we turn next.

III Why Legacy Banks Will Never Change

Excel, Bloomberg, and AI slide builders have raised output but not efficiency; they simply expanded the volume of busy work. Goldman employs ~12,000 engineers; Morgan Stanley spends $4 billion a year on tech—yet margins and staffing look like they did 20 years ago. The primary reason for this stagnation is org design and cultural inertia - not lack of technology.

Legacy apprenticeship model is outdated

The investment-bank ladder—analyst, associate, VP, director, MD—was designed a century ago to teach juniors by having them redo senior work. Evaluation and pay still hinge on visible effort: analysts log hours to prove commitment, mid-levels demonstrate diligence through slide volume, and partners judge quality through the grind they can see. The structure optimizes for apprenticeship, not speed or iteration, and any tool that shortens the grind undercuts the promotion path the pyramid is built on.

Committee governance

Strategic decisions sit with multi-partner committees rather than a single founder-CEO. Those committees prioritize revenue stability and rainmaker retention, so proposals that alter compensation, headcount, or workflow face lengthy debate and often stall. Even leaders who see the benefit of radical change have limited authority to force it through, and few personal incentives to spend political capital on re-architecting a system that already pays them well.

Tech solutions exist, but stall in deployment

Modern tools—Zoom transcripts, CRM capture, data APIs—are standard elsewhere, yet bank analysts still hand-scribe calls. Vendor-risk reviews can take a year, so anything new bolts onto the stack instead of replacing it. In my own early customer discovery interviews, the bankers’ top requests were call transcription and model-version control—features Zoom and SharePoint already ship, but culture and politics keep them switched off.

Incremental gains have limits

Centerview Partners is the usual exhibit for a “lean boutique” done right. The firm launched in 2006, kept headcount deliberately low, and focused almost exclusively on multi-billion-dollar transactions. That structure delivered meaningfully higher productivity:

Two caveats keep the result in perspective:

1. Purpose-built structure. Centerview was organized from day one to minimize layers and bureaucracy; replicating the same ratios at a 30 000-employee bulge bracket would require dismantling the existing pyramid, not installing new software.

2. Deal mix matters. The margin lift largely relies on mega-deal fees that rise far faster than staffing needs. If you apply this model to the smaller, more common deal sizes (let alone mid-market deals) and the efficiency advantage narrows sharply.

Centerview therefore demonstrates that thoughtful org design can push margins, but even the best traditional boutique shows a step-change only when headcount stays low and deal sizes stay very large. Point solutions alone, bolted onto legacy structures, are unlikely to reproduce that lift across the wider market.

While point solutions from AI vendors will surely be adopted, they won't fundamentally alter workflows within legacy constraints; productivity gains will likely remain marginal as busy work expands to fill any freed time.

Substantial productivity gains require an AI-first culture, risk-on ownership, and a vertically-integrated system—all built from day one. Incumbents are not configured to deliver that combination.

IV The AI-Native Opportunity

Picture Goldman Sachs was founded this morning and not in 1869.

No legacy tech stack, no pyramid hierarchy, no 60-slide pitchbook template. How would you staff your first M&A mandate, and which tools would you use?

That blank sheet defines the AI-native design space.

Three governing principles drop out:

1. Structural org redesign. Replace the five-layer hierarchy with the smallest team that can originate and close a deal. Junior time goes to judgment and outreach, not slide formatting.

2. Unified system of record. Calls, e-mails, financials, buyer notes—everything feeds one datastore that AI agents or humans can query in a single call instead of an endless SharePoint hunt.

3. Zero-marginal-cost analysis — the “why-not” loop. If an agent can screen and rank 10,000 buyers in less than ten minutes, “uneconomic” tasks become standard, compounding insight on every client touch.

Ironically, technology is the straightforward part; aligning data, org design, and talent is the real lift. An AI‑native bank starts with all three in place, while incumbents would have to unwind decades of habits, process, politics, and legacy code before they could even start.

So what does this look like in practice? Section V walks through how we did this at OffDeal - our first pod, real mandates, and the numbers behind them.

V OffDeal: The First AI-Native Investment Bank

Building a Bank, Not Selling Software

As we discussed, tech only drives step-change when it is fused to appropriate org design, incentives, and culture, so we launched OffDeal, a full-stack investment bank rather than a SaaS business selling tools to incumbents.

Our first battleground is lower-middle-market M&A—deals in the $5-30 m range that big banks ignore because $100-300k fees can’t support a bloated legacy deal team. A two-person pod plus AI agents does make money at that ticket size, and the high deal velocity gives us dozens of live reps each quarter.

Here’s how the system performs in practice:

Operating model: Pod x Platform

To achieve significant operating leverage we had to do two things:

1. Refactor the atomic unit of a deal team into a two-person pod

2. Develop a fully-vertically integrated software stack

The Two-Person Pod

Current throughput:

• 15-20 meetings booked per BD banker per week

• One pod runs 8-10 sell-side M&A processes in parallel

• Junior bankers involved in high-value tasks since day 1

The Platform — Three Layers on One Source-of-Truth

These AI-native workflows can only reach their full potential in an AI-native organizational structure and work culture that is AI-first.

The Unreasonable Customer Service

A common misconception is that AI replaces humans. The opposite is true: by eliminating busy work, our bankers spend significantly more time on high-value client interactions. The technology simply creates more space for deeper relationships - and for superior customer service.

When marginal cost of analysis approaches zero, you can do things legacy teams label “nice-to-have”. Two recent engagements show this in action.

Regional plumbing company

Hours before a second BD call, an OffDeal banker asked the platform for a quick diagnostic on the client’s financials.

• Instant check – The model flagged that labour costs sat in overhead, not COGS, revealing true gross margin well below peer median

• Market sweep – Using the client context, the banker asked the system to benchmark the client’s pricing against every local competitor within 50 miles - all in about 30 minutes.

• Insight delivered – The banker presented a competitor analysis and a forecast of the upside from a price increase. Impressed by the depth and speed, the owner committed to an engagement with OffDeal on the spot.

  

Montessori school — sub-$100 k fee

A single-site operator wanted to sell but couldn’t attract a banker on a success fee that small. OffDeal accepted the mandate.

• Buyer universe – Our platform scanned every Montessori or private-school operator in North America—more than 10,000 entities—and produced a ranked short-list of 1,802 prospects in minutes.

• Outreach – Personalized outreach + NDAs sent the same afternoon; 71 NDAs returned within two weeks.

• Outcome – Four firm offers created real tension; the winning bid closed 32 % above the seller’s initial offer 45 days after deal launch.

In both cases, tasks that would have been uneconomic—or impossible—for an incumbent investment bank were completed at near-zero marginal cost and within hours, not weeks. Every slide, buyer list, and diagnostic feeds back into the unified data layer, giving the model more examples of banker-approved output and making the next mandate faster and smarter.

The compounding advantage

Product and deal teams work side-by-side. When a prompt mislabels a metric or a buyer filter needs refinement, engineers hear it in real time and ship a fix the same day. Because every application is internal, we can release features at “good-enough” and harden them through live use rather than long test cycles. This creates a rapid product feedback loop.

External model progress compounds the effect:

• Longer context windows. A 1 M-token limit lets an agent ingest the entire deal context in one pass.

• Stronger reasoning models. Can handle complex scenario analysis that now matches—or exceeds—human output.

• Reliable tool use. Models call internal APIs to run deterministic tasks, widening the set of workflows that can be automated.

Each release frees more advisor time for the work clients value most—strategy, negotiation, and support during the single largest financial decision of their lives.

VI Looking ahead

Launching any full-stack startup, let alone an investment bank is hard:

• Many competencies, zero excuses. We must stand up data infrastructure, AI tooling, deal execution playbooks, and a brand that convinces owners to entrust their life’s work - all at once. There is no single “core feature” we can hide behind.

• Long lead, lumpy revenue. A SaaS demo converts in weeks; a sell-side mandate must be sourced, launched, and closed before a dollar lands. We carry months of payroll while a single busted deal can wipe a quarter’s pipeline—so the capital curve is steeper than pure software.

• Skepticism by default. Industry insiders often view a VC-backed investment bank run by “tech outsiders” as Silicon-Valley hubris. Until we post results, we’re assumed wrong.

The upside: those same hurdles make the model tough to copy. Clearing them builds a barrier to entry and an operating moat incumbents—or fast followers—can’t easily match.

Early traction

• Three deals closed in our first few months by 1 M&A banker

• Twelve live M&A mandates with expected fees of $1.3M

• $12M fee pipeline generated in Q1 alone

• Small business owners are resonating with faster value delivery and deeper analysis.

• Software ships to production in days, lifting pod capacity with each release.

• Junior bankers move straight to origination and execution—no years lost on formatting.

Where we’re headed

1. Short-term. Remain in SMB M&A until our processes and software reach maturity

2. Mid-term. Move up-market, competing for $100m+ deals.

3. Long-term. Move even more up-market ($1bn+ deals) and add adjacent advisory services—capital raises, debt advisory - on the same platform, aiming for a full-service franchise.

The thesis stays constant: an AI-native org, culture, and software stack can match—then exceed—incumbent capability with a fraction of the labour.

Broader implication

Any advice business built on legacy org structures - law, accounting, consulting, wealth management - faces the same constraints. AI-first, vertically-integrated challengers will appear in each field. The post-AI world produces a clean slate for how human work will be done.

In many ways, we've come full circle to what makes investment banking beautiful in the first place: delivering on those timeless client needs for best price, best terms, and least hassle. The underlying value proposition hasn't changed since J.P. Morgan's first M&A mandate in 1895. What's changed is our ability to deliver on those promises more effectively, for more businesses, with dramatically less friction.

Happy building!

As always, pushback and feedback always welcome. Twitter DMs always open @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI and business.

Subscribe now

Subscribe now

We've heard Sam Altman say that GPT-5 will be to GPT-4 what GPT-4 was to GPT-3. But what does a 10x improvement from here actually look like? What about another 10x for GPT-6? In recent conversations, I've noticed people struggling to envisage such a leap - after all, our current models are already pretty damn good. How much better can they possibly get?

My answer: a LOT better. We are only getting started.

Let me explain.

The Hierarchy of Work: From Tasks to Jobs

To understand where AI is going, we need to zoom out and look at the bigger picture of how work is structured. Let's break it down:

1. Tasks: Single, discrete actions

   • Marketing: Writing a cold email copy

   • Finance: Analyzing a company's quarterly earnings report

   • Medicine: Interpreting a patient's X-ray

2. Workflows: Series of related tasks

   • Marketing: Planning, creating, launching, and monitoring an entire email outbound campaign

   • Finance: Conducting a comprehensive company valuation, including financial modeling and competitor analysis

   • Medicine: Managing a patient's treatment plan from diagnosis to recovery, including tests, medication, and follow-ups

3. Jobs: Series of workflows

   • Marketing: A marketing manager responsible for multiple campaigns across various channels

   • Finance: An investment analyst covering a specific industry sector

   • Medicine: A primary care physician managing multiple patients with various conditions

This (imperfect) hierarchy illustrates how work complexity increases dramatically as we move from tasks to jobs, presenting unique challenges for AI automation.

The Compounding Problem

As we move from tasks to workflows to jobs, AI faces a compounding accuracy challenge. Each step in a process introduces a chance for error, and these errors multiply as the number of tasks increases.

If an AI is 90% accurate in understanding a task (i.e., the user's "instructions") and 90% accurate in executing it (final output matches what the user intended), the overall accuracy for that single task is 81% (0.9 * 0.9). While this may seem acceptable for a single task, it becomes problematic when applied to complex workflows or entire jobs.

This compounding effect explains why AI agents have seen limited adoption in enterprises so far. Current AI agents lack the reliability needed for complex real-world processes. Yet, this also means that even minor improvements in AI accuracy could yield outsized impacts, especially in multi-stage workflows.

At OffDeal, we've faced these challenges firsthand when building AI agents for M&A strategic fit analysis. Despite using advanced models, we still rely on meticulously crafted prompts (which take many hours to develop A/B test) and strictly formatted data inputs. As AI model capabilities evolve, we anticipate the need for such clunky workarounds to gradually diminish.

The Path to 10x Improvement: From Tasks to Jobs

We're in the early stages of AI automation, currently focused on individual task optimization. The future, however, promises far more:

1. Near Future - Workflow Automation: Complex workflows will be increasingly automated, introducing true "digital coworkers" capable of handling significant workloads.

2. Medium Term - Entry-Level Job Automation: AI agents will begin autonomously handling entire roles, likely starting with junior or outsourced positions.

3. Long Term - Advanced Job Automation: AI will progress to managerial-level decision-making, orchestrating intricate, interconnected workflows that demand high-level strategic thinking.

Each step represents a massive leap in AI capabilities, far beyond what most people envision when thinking about "10x improvement."

It's important to note that smarter models alone won't suffice. Substantial infrastructure development is necessary for agents to transition from data processing to action-taking. Additionally, hardware advancements are crucial, as exponential growth in token generation demand will require significantly expanded compute capacity. These elements are expected to evolve concurrently, potentially at comparable rates.

Conclusion

As we stand here in 2024 at the base of the AI S-curve, we're ought to ponder

• How will humans work alongside increasingly capable machines?

• Will there be resistance to this change, and how might it manifest?

• Is it possible to "future-proof" yourself in a rapidly evolving job market?

• How will regulation and politics shape the adoption and development of AI?

If you're interested in learning more about how OffDeal employs AI agents to transform SMB M&A, or simply want to chat about AI Agents x Finance don't hesitate to DM.

As always, pushback and feedback always welcome. Twitter DMs always open @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI and business.

Subscribe now

Subscribe now

TL;DR: Many founders have become excited about building software for the “untapped” SMB market, grossly miscalculating the difficulty they will face in both GTM and adoption.

In this post, I will make the case that founders interested in SMBs have a way bigger (and less risky) opportunity in front of them—they can just acquire SMBs and build their dream tech around them. In fact, most Tier 1 VCs are already doing this in stealth.

Let me explain.

Why Selling SaaS to SMBs is a Mistake

Building B2B SaaS has been a tried and true method for founders to build large, scalable businesses. However, things are much harder now than they used to be. The enterprise market is oversaturated with SaaS—by now, there is vertical SaaS for virtually every niche. TAMs are getting smaller and the field is becoming more competitive. Enterprise software founders are now competing with:

• Large horizontal incumbents (including OpenAI itself, Google, Microsoft, etc.)

• Post-product market vertical-focused incumbents

• Hundreds of early-stage startups

• Their customers themselves—many enterprises are increasingly choosing to “build in-house” given the decreased barriers to shipping software.

In response to this trend, many first-time founders are choosing to go “downmarket” to SMBs, seeing it as the next frontier. SMBs have largely missed out on the SaaSification of the enterprise world, and some claim AI will solve this.

However, I think this is a mistake in most cases. While it is true that SMBs are much less “tech forward” than their larger counterparts, there are strong reasons for this:

• Low ACVs: SMBs have much lower Average Contract Values (ACVs) with way fewer seats to sell to (i.e. you need more customers to get same revenue)

• Long sales motion: despite lower ACV, you will likely have the same, if not longer, sales motion as with regular enterprise deals. Extra kicker that many SMB owners will want to meet in person or want your local presence. They are not used to doing deals over Zoom.

• Reluctant Tech Adoption: SMBs are often operated by the older generation (average age of SMB owner is 60+), making them particularly unlikely to adopt tech or learn new tools. Most SMBs want the work done for them; they don’t want to click around in some app or talk to a chatbot.

• Hard to Reach: SMBs are hard to get a hold of—many are not on LinkedIn, don’t regularly check emails, and don’t trust strangers on the internet, making scaling your GTM motion very hard.

So, if you plan on selling to SMBs, you will suffer from both low contract values and a grueling sales motion. But what if there was a bigger opportunity hiding in plain sight?

Why You Should Consider Buying SMBs

Let’s say you want to automate the work of on-call IT support specialists. You spend a year building and testing a fancy new AI platform that can reliably automate 50% of the work. You spend a ton of time finding design partners, doing pilots, iterating, and finally having something workable.

A typical IT business that generates $5M in revenue and has 10 on-call IT staff spends about $1M/year on that labor. At 50% automation, you could deliver ~$500K savings per year, for which the SMB would be willing to pay ~20% of that for your software—$100K a year ACV (unusually high for SMBs, but let’s go with it for the argument sake).

But here’s the thing… the sales motion for a contract of this size and a major change to how the business operates will take many many months, if the owner even agrees to entertain the idea. Finding the right design partner who’s willing to provide regular feedback and iterate with you will also be very hard. To get to $1M in sales, you’ll have to find 10 of these. Let me assure you—it will be a grind, and will probably take longer than you think.

Now consider an alternative.

You decide to acquire said IT business for 1x revenue—$5M. Of that, you would typically pay $1M in cash, and the rest will be split between bank debt / seller note / equity roll for existing owners.

You just acquired yourself a zero-churn customer. As the new owner of the business, you can quickly iterate on the tech and achieve the said 50% automation. You have two options now:

1. Capture Savings: Capture the full $500K in savings per year into perpetuity; or

2. Grow Revenue: Maintain the current staff and focus on growing revenue per FTE (roughly $200K/employee, so with 50% automation you could do $1M more in sales).

To put simply, for a cash investment of $1M, you will receive a $500K-1M return per year (5-10x more than in the SaaS route). Additionally, you will have the roughly $1M/year of cash flow that the business had before any automation. In other words, you will own a business that makes $5-6M in revenue and $1.5-2M/year in operating profit. In the SaaS scenario, you would be making only a few hundred thousand, and most likely still be unprofitable.

Needless to say, you can reinvest the cash flow your business generates to acquire a second, third, and Nth IT SMB business. As long as you can find suitable targets, money will not be a problem. You can scale your tech-enabled business to $100M in revenue in a matter of just a few years, if done correctly.

Advantages of Acquiring SMBs

1. Easier Revenue Growth: Growing a business from $5M to $6M in revenue is much easier than growing a business from $0 to $1M in revenue (no cold start).

2. Proprietary Data Generation: You will generate valuable proprietary data to handle and automate edge cases, achieving even greater levels of automation in the future.

3. Larger TAM: Your Total Addressable Market (TAM) is now the IT services space itself, versus the software spend * number of IT SMBs.

4. AI Progress Benefits: As a consumer of AI progress, your business will be able to achieve better and better margin structure with GPT 5, 6 etc.

5. Increased Valuation: At scale, your business becomes increasingly more valuable. Your first few businesses can be acquired for 1x revenue, 5x EBITDA. If you do 10 of these acquisitions:

   • You will have invested $10M in cash / $50M in purchase price.

   • You will have $50M in revenue / $15M in operating profit (assuming no other operational synergies or growth in revenue).

   • All of your future acquisitions can be funded from internally generated cash flow plus debt (no future equity raises required).

   • You will own a profitable business worth about $100-150M to a private equity buyer. Of course, you will not stop at 10 acquisitions—there are thousands of potential IT services businesses you can acquire. The sky is the limit.

All of this from investing just a few million bucks from your seed raise! It is no surprise that pretty much all of the top-tier VC funds are now pursuing this strategy in some shape or form (most of them are currently in stealth but listing a few publicly announced):

• Crete - tech-enabled accounting & tax rollup (led by Thrive Capital & Bessemer)

• Roofer.com - tech-enabled roofing business (backed by Soma, Asymmetric, Mucker)

• Splash - tech-enabled swimming pool cleaning (led by Asymmetric)

• Stress Free - tech-enabled auto repair (led by Forerunner)

• Metropolis - AI-enabled parking platform (valued at $1.7bn)

Generalizing the Playbook

A VC-backed rollup strategy will work if the technology can drive either:

1. Increase in Revenue per FTE: Deliver more work due to automation or unlock a better/more scalable GTM motion through automation.

2. Decrease in Opex per FTE: Cut costs through automation.

For this type of strategy, look for large, highly fragmented industries. Bonus points if they have some type of recurring revenue and a software TAM that is too small for a VC-backed company. Outside of IT, you can consider businesses like the following (by no means an exhaustive list):

• Business process outsourcing (call centers, medical billing, data entry, debt collection, etc.)

• Insurance (title, claims adjudication, etc.)

• Accounting & tax

• Health care (primary care, dental, etc)

• Third party administrators

The interesting thing is you don’t have to build everything in-house. Build internally only the most critical tools that nobody else has created in the marketplace, and buy third-party software for anything non-core (e.g., tools for automating outbound sales, creating marketing content, etc.). The benefit of having a services business is that not everything needs to be perfectly integrated for the end customer—you are your own customer.

A Few Words of Caution

• M&A is Not Easy: Finding the right targets, diligencing the business, and structuring the deal are non-trivial undertakings. Ensure you have at least an ex-private equity/investment banker/corp dev person on your team to maximize your chance of success.

• Service Business Management: Owning a service business (especially if it touches blue-collar work) is very different from working at Meta, Blackstone, or McKinsey. Ensure you have what it takes to work with that type of labor force.

• Founder-Market Fit: Founder market fit still matters—ideally, you have some unique insight into how technology moves the needle, and you are using M&A to acquire distribution. You need to be able to develop best-in-class SOPs/playbooks.

• Service businesses are valued differently - you cannot expect a juicy 10x ARR multiple here - your business (at scale) will likely be valued at 10-20x EBITDA - which means to get to a $1B valuation, you’d have to acquire / grow your way into $50-100M in cash flow - but my argument is that it is much easier to do so under this strategy, than getting to a $100M in ARR with pure SaaS. And you will most certainly not have to raise as much money - especially after you do your first few acquisitions, since future growth will be funded with your own profits.

Conclusion: Is Acquiring SMBs Your Next Move?

As you consider the path forward, ask yourself:

• Are you struggling with low ACVs and high CACs while selling to SMBs?

• Do you have unique insights into how technology can transform a traditional service business?

• Could you capture more value if you just used your own software to deliver the service your end customer provides?

If you’re pondering about any of these questions or stuck in the SMB go-to-market hell, I’d love to hear your thoughts and experiences.

Reach out to me—I’ve talked to hundreds of SMB buyers at this point and I am happy to brainstorm / discuss how we can help you systematically find and acquire the right strategic targets.

At OffDeal, we already partner with numerous private equity firms and VC-backed rollups pursuing this strategy.

As always, pushback and feedback always welcome. Twitter DMs always open @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI and business.

Subscribe now

Subscribe now

"Generative AI is ruining my outbound response rates!"

I was puzzled when I heard this complaint from a private equity friend in charge of deal origination (sourcing) at his firm. Shouldn't AI tools supercharge sourcing workflows? He went on to explain that response rates to his carefully crafted, personalized cold outreach emails had plummeted 40-50% in just the last 6 months. Worse yet, whenever he did get through to a prospect, he was often the 3rd or 4th investor to reach out that same week!

My friend brought up a non-obvious insight - GenAI made it 10x easier to draft and send thousands of “personalized” cold outbounds at scale - for everyone at once - rendering cold outbound marketing less and less effective. The exact opposite of what GenAI was meant to accomplish!

In this post, I wanted to explore this paradox and offer some thoughts on what it might mean for the future of sales.

When alpha becomes beta

The two primary ways in which Gen AI can supercharge sales workflows are:

1. Sourcing - LLMs give web-crawlers and alt data miners superpowers by enabling them to ingest, process, and analyze a lot more unstructured data than before

2. Outreach - LLMs can ingest a bunch of disparate data sources (LinkedIn, blog posts, company website, and even the sender’s personal profile) and craft a hyper-personalized cold email for each prospect- at scale.

This is a salesperson's dream come true, right? Not only do you get a substantially wider top of funnel (more leads identified), but you also get dramatically higher response rates given the personalized outreach (instead of generic templates). Gen AI eliminates the drudgery of digging for prospect leads while making outreach personal and relevant. It is no wonder that many salespeople are rushing to employ these tactics before their competitors.

But herein lies the paradox…

As this technology begins to proliferate, prospects are increasingly inundated with "personalized" cold emails. They are quickly growing weary of what increasingly appears to be spam, causing the response rates to plummet as recipients disengage from the onslaught of AI-generated outreach. Personalization is becoming commoditized. What was once an effective way to grab attention is starting to lose its edge - it is no longer a signal of a carefully crafted and highly intentional outreach.

In other words, Gen AI might make outbound sales even harder… With that said I still think AI outreach is destined to become the norm - it’s game theoretic - while you (a salesperson) may not get an advantage from using it, you might very well be at a significant disadvantage if you aren’t.

What will happen to sales?

The implications of this are profound for sales in an AI-first world. Traditionally, sales have always been a human-to-human enterprise, but this is bound to change now. We are already seeing the beginnings of AI-to-human interactions (which today feel fun and gimmicky). But as the technology evolves, will we soon see AI-to-AI sales motions? Put differently, will each buyer and seller have an AI agent representing their very best interests? If so, how much agency and decision making power would we endow these agents with? Would the sales AI agent be able to “trick” the buyer AI agent? Will there be ads? Or maybe all this technology will simply revive the value of in-person events and place an additional premium on the old-fashioned deal making at the golf course?

Life is going to get weird.

Conclusion

The future of selling will most certainly involve augmented intelligence, but the actual sales workflows will likely require continual rewriting as technology evolves. As each new strategy gets effectively arbitraged, human ingenuity will have to find new ways to reach and target prospective buyers - it is in our society’s capitalist DNA. One thing is for certain - salespeople who choose not to pay attention to AI and its evolving capabilities will definitely be left behind by their competition who do.

As always, all thoughts / pushback welcome. Twitter DMs always open @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI and business.

Subscribe now

Subscribe now

Since ChatGPT's launch just nine months ago, the tool has been used by over 80% (!!!) of Fortune 500 companies. However, firm-wide enterprises adoption was somewhat slow due to data privacy concerns and complexity around custom deployments.

Today, OpenAI finally announced a highly anticipated enterprise version of its flagship ChatGPT product that solves both of those issues. It’s a HUGE upgrade:

• More private and secure:

  • Data not shared for training OpenAI models

  • Data encryption at rest / in-transit + SOC2 compliance

• More powerful and performant:

  • No usage caps + 32k token context window (4x ChatGPT)

  • 2x higher-speed performance vs ChatGPT

  • Unlimited access to Code Interpreter (advanced data analysis)

• Customizable:

  • Share custom common workflows / chat templates across the organization

  • Fine-tuning options for more domain-specific use cases

As if this wasn’t enough, OpenAI also announced some upcoming new features to be released in the coming weeks / months (these will kill so many recent startups…):

• Connect your own company data to extend ChatGPT’s knowledge base

• Power tools - a more advanced version of Code Interpreter + web-browsing

• Function-specific tools - tools for specific roles, such as data analysts, marketers, customer support and more

This further strengthens my view that most companies will choose to build tools in-house instead of licensing/buying from AI startups. The impact is too critical and the technical complexity is more and more abstracted away - OpenAI is making it easier and easier for companies to spin up their own internal tools.

AI startups must tread carefully and choose the right surface area to compete in. The bar keeps getting higher and higher.

As always, all thoughts / pushback welcome. Twitter DMs always open @leveredvlad

PS If you are tinkering in AI applications to the financial workflows, please reach out - I love jamming about this stuff and would love to compare notes.

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI, crypto, and other cutting-edge technology.

Subscribe now

Subscribe now

McKinsey just unveiled its new internal generative AI tool “Lilli” - a chat application trained on more than 100,000 proprietary internal documents and interview transcripts. Lilli (even in its current beta) has been a huge success internally - both in terms of impact and adoption:

• The tool has already dramatically cut down the time spent on research and planning from weeks to mere hours, and from hours to just minutes in others

• An astounding 66% of employees now use the app multiple times a week

• In the last 2 weeks alone, Lilli answered a whopping 50,000 questions!

Lilli’s success could set a precedent for other Fortune 500 giants grappling with the perennial “buy versus build” question. This should worry builders in the “B2B AI SaaS” startup world, and prompt them to seriously reconsider their product strategy and approach.

Let me tell you why.

Most AI startups today are features, not companies

Lilli can do exactly what myriads of emerging GenAI startups promise to do:

• Semantic search - The tool can zero in on relevant internal documents, files, and even pinpoint internal expert recommendations based on the user’s query.

• Retrieval Augmented Generation - It can respond to questions leveraging McKinsey’s exclusive corpus of presentations, research, and transcripts, complete with references and sources.

• Document analysis - processes and analyzes uploaded client files and documents (this feature will be deployed later)

• Gen AI Chat - Just like a chatbot using a general-purpose LLM (akin to ChatGPT), it can summarize articles, produce text, and even draft emails.

Twitter is inundated with hundreds of demos of products that individually offer one of the above use cases. The fact that Lilli can reliably do all of them confirms that these are more likely to be merely features and not standalone products (or let alone companies).

Building/Partnering makes more sense than buying

Every company working through its “AI strategy” is debating whether it should buy a third-party solution or build in-house. The main criteria are

1. Strategic priority - How pivotal is this to the core of the business?

   By now there is an overwhelming agreement that knowledge workers will benefit greatly from Gen AI and will experience significant productivity gains (at least 20-30%). This makes the technology indispensable for firms with a large knowledge worker base.

2. Urgency - how quickly do I need to have a production-ready solution?

   Given the profound impact of GenAI tools, lagging behind might render a company less competitive, making this an urgent priority. However, given the wide availability of LLM APIs and open-source LLMs, building or partnering is becoming increasingly attractive.

One other critical concern for large corporations is that of regulation, compliance, and security - from safeguarding confidential customer data to accessing employee PII, trusting an untested nascent startup might prove challenging to many.

McKinsey concluded that a combination of cost, reliability, and security warranted an in-house solution - despite the higher “tech lift” required. Indeed, McKinsey chose to build its own app on top of other LLM technologies:

• Cohere (most likely using their Embeddings and Semantic Search products); and

• OpenAI Microsoft Azure service (most likely for the “enterprise-friendly” secure GPT-4 instance).

That in itself may indicate that the tech lift of building LLM wrappers is not that high in the first place - both Cohere and OpenAI likely offered solutions architects to assist with the implementation, and the robust open-source community likely offered McKinsey developers lots of frameworks to use as inspiration.

Building in-house has another benefit - Lilli is built *exactly* the way McKinsey wants it (does not have to “settle” for an off-the-shelf all-size-fits-all solution), allowing Lilli to seamlessly “plug in” into the existing tools and workflows of the company.

There is somewhat of a self-reinforcing feedback loop here - as more companies elect to build instead of buying, the lower the perceived tech lift will be, and the more other companies will choose to build as well. This puts startups in a precarious situation, where they need to justify to enterprise customers that their tool is worth the tradeoffs involved - something that might prove to be very challenging.

LLM agnosticism

McKinsey's declaration of being "LLM agnostic" is telling. To them, the LLM is a mere means to a desired end: achieving business goals. With Lilli's in-house build, they command the flexibility to swap out underlying technologies as deemed fit.

In most cases, startups will have to offer similar flexibility in terms of the underlying technology to win over enterprises - something that will likely require a higher engineering lift and will involve continuous testing, QA, etc over multiple ever-so-evolving LLMs available on the market. This might be challenging for smaller startups with leaner technical teams.

What This Means For B2B AI Startups

The early success of McKinsey’s Lilli is promising - it means that the hype around generative AI is not unwarranted - the technology will most certainly transform the existing workflows of most knowledge workers and will supercharge their productivity. It is no surprise that so many entrepreneurs are interested in the space.

However, the fact that a professional services firm like McKinsey was able to build and deploy a working product in mere months puts into question the merits of buying/licensing third-party “core” solutions from up-and-coming AI startups.

In my view, founders should seriously reconsider their unique selling points and value proposition if they are going down the “enterprise ChatGPT”, “chat with your data”, or “AI search” route. Does their startup have the “right to win” in these areas?

To close, here are a few examples of interesting areas to explore in B2B land:

1. Focus on services around LLM applications - things like custom model training, evaluations, QA, optimizations, etc. - these things are “hard” that large organizations might need help with as they look to deploy their own applications. Humanloop has been able to onboard lots of F500 customers this way.

2. In the spirit of doing the “hard thing” - offering solutions that require deep technical expertise and proprietary technology to build might also entice businesses to buy from you. A good example here is Synthesia that allows users to create video content with human-like AI avatars in seconds.

3. Focus on "non-core” business functions that enterprises would be much more willing to “outsource”. One example is business processes that have historically been outsourced to humans in lower-cost jurisdictions (claims processing, data entry, customer support, etc.). Companies like Intercom have done well in this area. Alternatively, for lean highly leveraged companies like private equity and hedge funds, companies like Hebbia offer an off-the-shelf solution that the investment firms would never bother building on their own.

This is by no means an exhaustive list. There will be more than a few successful B2B AI-native companies built in the coming years. But the McKinsey case study shows us that aside from the powerful incumbents, startups will also have to compete with the in-house developer teams of their target customers.

As always, all thoughts / pushback welcome. Twitter DMs always open @leveredvlad

PS If you are tinkering in AI applications to the financial workflows, please reach out - I love jamming about this stuff and would love to compare notes.

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI, crypto, and other cutting-edge technology.

Subscribe now

Subscribe now

Recently, there’s been a flurry of “Copilot for Private Equity” startups - I counted more than a few in the last two YC batches alone. Unfortunately, I think most of them are working on the wrong problem. The real value unlock, in my view, is not in supercharging investors, but in supercharging their portfolio companies.

Let me tell you why.

PE Workflows Are NOT a “Hair On Fire Problem”

Cost savings don’t move the needle

Most AI startups today are trying to automate “low-impact” but highly time-consuming investor workflows. Some popular examples below:

• Faster data room searches

• Streamlining internal reporting

• Organizing and summarizing call notes

• Automate legal doc review / drafting

However, given that the fund’s top line is a function of management fees (charged as a % of Assets Under Management) and carried interest (~20% of the realized profit), most of the proposed AI solutions do not boost revenues in any meaningful or obvious/direct way. This means that the main benefit to the fund is cost/time savings.

But here is the thing - these productivity gains do not really move the needle for this end customer. Let’s look at some illustrative economics of a typical fund.

A $2.5B fund will make ~$250M per year (carried interest + management fees):

Junior resources (who are typically responsible for most of these ‘automatable’ tasks) will cost the fund anywhere from $3M-$15M per year (depending on team size), which adds up to roughly 1% to 6% of the fund’s total revenue.

This means that even with a productivity boost of 100% (2x efficiency gain), you’d be looking at delivering a mere 2-4% cost savings (as % of total revenue) - this assumes FTE reduction (i.e. funds will not be able to do 2x more deals).

Of course, the numbers might look different for smaller funds, where junior overhead makes up a higher % of the overall topline. But this comes at a cost - these customers will drive much smaller average contract values, making the TAM / total revenue opportunity even less attractive. Hard to make a case for a VC-scale business here.

PE is not very cost sensitive

Beyond the relatively ‘marginal’ cost savings, PE firms are not very cost sensitive to begin with! There is a reason why they do not outsource all of their workflows to a low-cost center in India (even though they DO have access to them) - they are OK paying $350K to Emily in NYC, knowing she’ll do the task at hand correctly, and quickly. Clearly, they value accuracy and trust much more than pure costs. So whatever marginal productivity gains a new startup COULD deliver to these firms will be weighed against a whole host of risks - data, privacy, accuracy, etc.

The (very real) threat from incumbents

The playing field is further complicated by incumbents who are well positioned to incorporate AI into their offerings. For instance, data room providers are already integrating "AI-enabled search" into their platforms. Moreover, expansive tools like Microsoft Copilot can leverage an entire firm's knowledge graph to streamline tasks like locating files or drafting memos (of note, Microsoft will charge $30/mo for this).

These startups are, therefore, grappling not only for their target customers' attention but also against titans with an established customer base and data access that startups might struggle to attain.

The Real Opportunity: AI for PE’s Portfolios

In my discussions with over 50 investors across large and mid-market funds, the consensus is that the most substantial opportunity lies in enhancing portfolio value through AI-driven efficiencies. Some examples include:

1. Improved analytics to identify trends and operational improvement opportunities.

2. Automate/augment sourcing and diligence of tuck-in M&A opportunities

3. Workflow automation at the portco level to reduce or reorient Full-Time Employee (FTE) hours towards more value-adding tasks.

Even a modest 5-10% efficiency gain (in terms of EBITDA) at the portfolio level would have an outsized impact on the fund’s total profits (multiples more than the cost savings from increased productivity at the junior level!):

Of course, building a tool that can achieve these things is not simple - not the least because every sector has their own intricacies and nuances. But the opportunity could warrant a vertical tool. Perhaps, the customer here is not even the private equity firm, but all SMBs in the XYZ industry, and a PE firm would just be a clever customer acquisition channel - if the fund likes your tool, they can integrate it across their entire portfolio, onboarding 10-15 customers in one sweep.

Conclusion

While I firmly believe that AI will permeate all knowledge-worker domains, including private equity, I contend that the market for a PE-specific automation tool isn't as lucrative as current startup activity suggests. There might be slivers of opportunity here and there, but I question the size of that market (both in terms of % of the job that is automatable, willingness to adopt these tools, and how much one could charge for them). Further, incumbent tools are likely to meet most of the “basic” needs of PE investors, leaving limited room for new startups to add meaningful value.

The more potent opportunity, in my opinion, lies in driving operational efficiencies at the portco level through workflow automation and improved analytics. This has the potential to unlock outsize value for the PE firm's top line—a prospect that is bound to capture their interest.

As always, all thoughts / pushback welcome. If you are tinkering in AI applications to the financial workflows, please reach out - I love jamming about this stuff and would love to compare notes.

DMs always open @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI, crypto, and other cutting-edge technology.

Subscribe now

Subscribe now

Github Copilot made headlines after it was reported that software engineers were on average 55% faster in completing their tasks when using the tool (in addition to higher completion rates, higher job satisfaction etc.). Since then, it became a pretty consensus view that there will be a “Copilot” tool for every profession from healthcare and education to finance and accounting.

Perhaps the most successful “Copilot for X” to date has been a Series A startup Harvey AI - a tool that promises to completely transform and revolutionize how legal work is carried out today.

In this short essay I will explore economics behind Harvey AI, explain why it is good business for Big Law, and why other professions may not reap the same benefits.

Let’s dig in.

Economics of Big Law

Before we dive into economics of Harvey, it’s important to understand a few nuances around how Big Law works and makes money today, as it is directly tied to the value a “Copilot for Lawyers” can bring to the industry.

At a high level, lawyers are effectively legal consultants that charge a steep hourly rate for the work they produce. An entry level Associate can charge over a $1,000 per hour, with more senior lawyers often charging several multiple of that.

Of note, lawyers do not charge for every hour work of their work. On average, an entry level associate may accrue around 2,000 billable hours, which means that there are about 1,500 hours that are non-billable (assuming a 70 hour work week). Setting aside the debate whether the number of billable hours would have to come down from any use of automation tools, the main pitch to Big Law firms is that they can greatly reduce the amount of non-billable work they do (i.e. work that does not directly generate revenue). For any law firm this presents to opportunities for value capture:

1. Do more work with same number of full time employees (FTEs)

2. Do same amount of work but with less FTEs

Based on my quick back of the envelope math, it is clear that the first approach is much more preferred given the “gross margin” earned on an entry-level associate. Of course in reality, there will likely be some combination of both, but for simplicity, I provided the two book ends. We can see that there is potential to increase revenue by up to $750K or cut costs by up to $50K on a per FTE basis.

For context, Allen & Overy, an elite UK law firm that was the first pilot customer of Harvey AI generated about ~$2.5 billion in revenue across its 3,500 attorneys, implying revenue of over $700K / FTE or ~$360/billable hour (UK firms charge less than in the US). Assuming our base case of 25% productivity gain, this has the potential to boost A&Os revenue per FTE by some $130K per year. Not bad!

Economics for Harvey AI

Ok, so the potential value capture for Big Law is obvious and pretty attractive. But what does this look like for an AI SaaS product like Harvey AI?

First, let’s talk about costs. Unlike traditional SaaS that has a variable cost of close to zero, AI tools incur hard compute costs (or API call cost) for every user interaction. This can be tricky, since most SaaS is priced on a fixed price per seat, while your cost structure is largely variable. Indeed, a power user is LESS profitable than a casual user, creating an interesting incentive misalignment.

So what are the unit economics for Harvey? Skipping the relatively one-time costs of embedding a large corpus of legal text, hosting costs, and engineering overhead, the primary cost for Harvey (or any AI-first tool) will be its compute/API costs. Harvey uses GPT-4 (at least for now), so using the pricing from OpenAIs website we see that:

1. Inference Cost for Prompt (Input) is ~$0.03 per 1K tokens (1 token = ~0.75 words)

2. Inference Cost for Completion (Output) - $0.06 per 1K tokens

So the cost is directly linked to how much text is ingested and generated per user interaction. But wait! There is one more kink - some interactions may require multiple LLM calls - for example, if a lawyer asks some legal question that spans multiple legal documents, Harvey would have to fetch several closest matches (using an embedding model) and then feed multiple chunks of text into its prompt to produce an output.

This is all to say that doing the exact math on this is very tricky without seeing actual usage data, but I am going to give it a pass. Assuming ~20 words per user query, ~250 words for the chunks fetched from legal documents, and a roughly ~1,000 words in the final output, we get the following unit economics:

Of course, I also had to make some assumptions around how many LLM calls per interaction (I assumed 2) as well as the number of times a day does an average lawyer uses the tool. Based on my back channel checks, Harvey currently charges around $500 / lawyer per year, resulting in a negative gross margin - not quite the “gold standard” of 80-90% of traditional SaaS.

A few caveats:

1. API / compute costs are likely to come down over time

2. Based on the value unlock to the end user, Harvey can charge a lot more

For example, if Harvey revises its price to $5,000 / year (still a fraction of the $130K in potential new revenue for A&O), it’s gross margin will be ~90%! Is it crazy for them to charge $10K? $20K? Time will tell - but it will be largely driven by the value and impact they bring to the end customer.

Generalizing for “Copilot for X”

A similar analysis must be done for any aspiring “Copilot for X” product. If you are building a tool for a cost center, the math will likely be much less interesting than if you were augmenting a revenue generating employee. As we can see from my monkey math above, even at a 50% productivity gain of a highly paid lawyer, you “only” bring $50K in annual savings - vs some $750K in potential additional P&L (a 15x difference!). If you are automating a $50K/year FTE, this puts a ceiling on how much value you unlock (and how much you can ultimately charge).

Secondly, it’s important to consider if it is a big enough cost item for your target customer. If the target user amounts to 50% of your customer’s fixed cost, this is attractive - but if the target user is only 5% of the cost stack, even a 100% efficiency gain (i.e. fully remove the need for this FTE) barely moves the needle. This will also dictate how willing enterprises will be to adopt your (untested) tool, and how much they would be willing to pay for it.

Conclusion

Harvey AI found itself in a sweet spot - it serves a highly profitable revenue generating employee that deals in a very rules-based industry that deals primarily with text. This means that other industries may not benefit the same from an AI Copilot. If I was building a vertical copilot tool I would think about the following:

1. Is my product primarily cutting costs or adding new revenue?

2. Does my product move the needle for the end customer? (is it a must have?)

3. Can the current tech handle the type of work my customer does? (how much precision do you need, do you deal with only text, or also numbers, images, audio?)

My view is that for anything ‘incremental’ a horizontal tool like Microsoft Copilot will probably be suffice, so any new tool needs to be a very meaningful improvement over what incumbents will be able to offer enterprises for free.

Personally, I’ve been looking at Copilot applications in the financial services industry, which has a completely different business model and type of work than those of lawyers. I will write up a separate blog exploring my learnings, but if you are interested in jamming on this topic please reach out - I’d love to chat with you.

DMs always open on Twitter @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI, crypto, and other cutting-edge technology.

Subscribe now

Subscribe now

Can it be good business to give something away for free?

At first glance, that's what open-source software seems to be: freely available code that anyone can download and use however they want.

So how can an open-source company like StabilityAI raise $100 million at a unicorn valuation? What drives developers to pour hundreds of hours of free labor into projects like LangChain and LlamaIndex? What motivated Meta to open-source its 65 billion parameter Large Language Model? Are these actions purely altruistic?

In this essay, I will make the case for open-source software in the context of AI development. In particular, I will demonstrate that open-source can:

1. Provide a superior product for end-users

2. Offer a more sustainable business model in the long run

Let’s dig in.

What is Open-Source?

So, what exactly is open-source software?

At its core, open-source software refers to programs whose source code is made available to the public, allowing anyone to view, use, modify, and distribute the code. This is in contrast to closed-source software, where the source code is proprietary and kept secret.

Imagine a communal dinner where villagers gather to cook together, openly sharing recipes and culinary secrets. They tweak and improve upon each dish, embracing their diverse tastes and perspectives. Each cook can borrow a recipe and use their own ingredients to create the dish. This is the world of open-source software, where the "recipes," or source code, are accessible for everyone to experiment with, adapt, and share in a spirit of collective creativity.

In contrast, closed-source software is akin to an exclusive restaurant with a secret menu, where a single, masterful chef and their skilled team meticulously prepare the entire meal, delighting patrons with unique and exquisite dishes. The "secret sauce" is carefully guarded; only the chef can change the recipe, and users are forbidden from replicating the dishes at home.

One approach is decentralized and collaborative, while the other is top-down and centralized. So what does this mean for the end-users who consume the final product?

Advantages of Open-Source AI

Most end users are unaware if a given software is open- or closed-source. This distinction becomes even less visible in AI, as users only interact with an application built on top of a model, which can be either open-source or proprietary.

Nevertheless, open-source software offers a number of significant direct and indirect benefits for its users:

• Open models, private data: AI models are increasingly becoming essential infrastructure for large enterprises and governments, which often prioritize data security, control, and cost-effectiveness. Closed-source models may transfer too much power to private AI companies, whereas open, auditable, and interpretable models allow organizations to fine-tune them with their own private or regulated data and maintain control within a secure perimeter, while also reducing costs associated with licensing fees.

• Customization and optimization: Closed-source models often come with restrictions (and sometimes censorship) on their potential use cases. Open-source models allow developers to modify and customize the model to fit their specific needs (e.g. a niche use case or industry). Additionally, open-source models allow for easier optimization since developers can modify the model’s architecture and parameters directly, which may lead to better performance and more efficient resource usage.

• Interpretability, Auditability, Security: The transparency of open-source AI models fosters trust, facilitates audits, and provides educational opportunities, ensuring both security and quality. By examining the underlying code and documentation, individuals can enhance their understanding of AI systems. The "many eyes" effect, which relies on the active engagement of the developer community, helps identify and fix vulnerabilities that might otherwise go unnoticed.

• Decentralization of AI Ecosystem: Open-source fosters a more inclusive and diverse development environment, reducing the concentration of power and promoting a more equitable future for AI technologies. By encouraging collaboration, lowering barriers to entry, and enabling customizability, open-source AI models contribute to faster innovation and a wider range of participants in AI research and development.

Beyond the benefits to the end-user, open-source will be a superior business model.

Let’s explore why.

Open-Source AI: A Game-Theoretic Perspective

Today, foundational language models follow scaling laws, where larger models typically yield better performance. However, there are theoretical limits to this scaling (especially in NLP and image generation), which will manifest in two ways:

1. Dataset asymptote - a point at which all large language models will be trained on the entirety of all public data (i.e. the internet). As we get closer to this limit, adding more training data will yield diminishing marginal returns. Note: some additional gains might be achieved via higher training time and RLHF.

2. Performance asymptote - a point at which humans are unable to distinguish between the performance of different general-purpose AI models (similar to how we are unable to differentiate between 4K and 8K HDTV resolution).

As we gradually approach these asymptotes, it will be game-theoretic to increasingly compete on cost. The only remaining sources of performance edge might come from customization and/or optimization of models for specific use cases, as well as adding private proprietary data to the training sets. However, as discussed in the previous section, open models are much better positioned for this given their inherent flexibility, transparency, and lack of licensing fees.

This evolving landscape may lead to a world where all big tech companies and other major AI labs are forced to open-source their models and focus on monetization through customizations, scaling, hosting, and deployment services instead. In other words, open-source (or more accurately, open-core) could become the only viable long-term business model in the AI industry.

Some companies, like StabilityAI, are already front-running this trend by offering open-source models and focusing on providing value through custom models trained on private data for enterprises. OpenAI has been developing custom variants of its GPT-4 model for companies like Morgan Stanley and Duo Lingo too. For now, they are the only game in town, but as Google, Meta, and others join the AI race, competitive pressure to open-source models will greatly intensify.

Risks and Considerations of Open-Source

While open-source offers numerous benefits, there are potential risks and drawbacks to consider, such as:

• Misuse: The open nature of the software can potentially enable malicious actors to exploit vulnerabilities or use the technology for nefarious purposes. Some examples include using AI for hacking (uncovering vulnerabilities) or generating deep fakes. Appropriate regulation should target the use of technology, rather than the technology itself.

• Resource Constraints: Open-source projects often rely on community contributions, which can lead to resource limitations and slower development compared to well-funded proprietary counterparts. This is particularly true for LLMs, which may cost hundreds of millions (or even billions) of dollars to train.

• Fragmented Development and Quality Standards: The collaborative nature of open-source can sometimes lead to fragmented development, as different contributors may have varying opinions on the direction a project should take. This can potentially result in difficulties in establishing and maintaining consistent quality standards across projects.

Conclusion

The AI revolution has only begun and it is hard to extrapolate what the market will look like even just 12 months from now. Today, OpenAI is dominating the news cycle, but we know that the field will be crowded with competition in no time - both from Big Tech (i.e. Google, Meta, Apple) and new players (i.e. Stability AI, Anthropic, Cohere, NVIDIA, and many others). Model performance is more likely to converge than diverge in the future, which should drive the costs of proprietary models close to zero. Open-source (or open-core) offers an alternative business model that can be sustainable and highly profitable. Most importantly, it has the potential to offer a much better experience for the end user.

Long live open-source :)

Let me know what you think! Will foundational models become commoditized? Will all major models be eventually open-sourced? Who will accrue the most value?

DMs always open on Twitter @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI, crypto, and other cutting-edge technology.

Subscribe now

Subscribe now

Ah, the age-old question - does size matter? In the world of large language models (LLMs), the consensus has been that it does. Most of the mainstream excitement has been around the development of bigger and bigger models - and for good reason - increasing model size has been a major driver of performance. However, new research is emerging that challenges this notion, highlighting the importance of data quality and model fine-tuning.

In this essay, I will make the case for smaller models and why I think they will become the primary modality of human-machine interaction in the future.

Let’s dig in.

The Race Towards Bigger LLMs

The LLM arms race arguably started with the seminal paper “Scaling Laws for Neural Language Models” (Kaplan, McCandish, 2020) that demonstrated that model performance is driven primarily by just three factors - number parameters, size of dataset, and amount of compute used for training.

Scaling laws have been the driving force behind the "bigger is better" belief in LLMs.

However, as scale increased by orders of magnitude, so did the cost to train these models (adding up to tens and hundreds of millions of dollars). This had important implications:

1. Power concentration - only the most capitalized private tech companies can afford to participate in the development of these models

2. Proliferation of closed-source - easier to monetize and recoup the high upfront training cost, leading to issues around transparency, alignment, and security.

3. Lack of optimization - it is cost-prohibitive to iterate given that each run costs millions of dollars in compute, resulting in suboptimal model configurations.

The last point is an important one - models like GPT4 are incredibly expensive to train and fine-tune given their scale, and it’s impossible to run them locally.

But what if there was a better approach?

Small Models: The New Frontier

Recent research suggests that bigger isn’t always better when it comes to LLMs. A paper published by OpenAI (Ouyang, 2022) reveals that fine-tuning language models with human feedback can drastically improve performance without necessarily increasing the model’s size. In particular, OpenAI demonstrated that its 1.3B parameter model, InstructGPT, substantially outperformed GPT-3 (with 175B parameters) despite having 100x fewer parameters!

Small models with performance rivaling larger counterparts foster open access (lower compute costs → less barriers to entry → less oligopolistic dynamics) and promote hyper-specialization, enabling the creation of powerful industry-specific models that outperform generalist alternatives.

The most significant benefit of small models is edge computing, enabling highly efficient and agile models to run directly on devices like smartphones and laptops, rather than relying on cloud-based storage and expensive GPUs. This reduces both inference and, most importantly, training costs, paving the way for hyper-personalized models that continuously retrain on individual user personal data. Additionally, mobile edge computing enhances privacy, as data remains local, eliminating the need to send private or sensitive information to remote servers.

A clear winner here is Apple, which is most certainly developing its own small LLM optimized to run on its iPhone chips, thanks to their full vertical integration (they already released an open-source optimized version of Stable Diffusion image model). Android OS will likely also create its own powerful small models, but given the heterogeneity within the hardware it runs on, it is unlikely that this LLM will achieve the same level of optimization and speed as Apple.

Another interesting player to monitor here is EleutherAI - a research lab developing smaller open-source LLMs. This year they are expected to release a model with up to ~100B parameters, which should be more than sufficient for most use-cases given the appropriate reinforcement learning / fine-tuning.

These small, nimble models will become ubiquitous, seamlessly integrated into our personal devices - they will be a natural extension of our mind and we won’t even notice they are there. It will feel like magic.

OK, so small models are pretty cool. But what does this mean for the future of LLMs?

A Tale of Two Cities Models

I think a likely scenario is that we will have two types of models:

• Ultra large proprietary models (e.g. GPT-7) trained on public data (internet)

• Small optimized open-source models (e.g. LLaMA) fine-tuned on private data

A few ultra-large models (Microsoft’s New Bing, Google’s Bard?, etc. ) may provide us with general knowledge and reasoning about the world, while the small hyper-personalized models will run locally on billions of laptops and smartphones around the world. The sweet spot may lie in a delicate dance between these two approaches, where large and small models coexist and complement each other in various applications.

So what does this mean in terms of our path towards AGI? Perhaps we get there not through one jumbo model, but instead through an interconnected hive of myriads of hyper-specialized models.

Time will tell.

Conclusion

I hope by now I got you excited about new developments in the small model land. There are still many unknowns, but ongoing research efforts will continue to drive massive performance improvements and will unlock new possibilities.

In closing, I wanted to leave you with a few open questions to ponder about:

1. How does the optimal ‘interface’ change with model size? Is text the best modality for both?

2. Will various LLMs operate/interact with each other, akin to a hivemind? What does this architecture look like? How do we ensure maximum interoperability?

3. How will regulation affect the development of AI systems? Will governments be OK with a few big companies wielding so much power? Will they push/encourage the development of small open-source models?

4. Will consumers be comfortable sharing their private data with closed-source, proprietary, opaque AI systems?

Let me know what you think! DMs always open on Twitter @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI, crypto, and other cutting-edge technology.

Subscribe now

Subscribe now

Why do humans blindly trust their calculators?

For the most part, it’s because calculators produce consistent and verifiable results. They are “deterministic” - they always produce the same output for a given set of input conditions or parameters without any element of randomness or uncertainty. This approach to rules-based algorithms extended beyond simple calculators - until the 1990s most of our software relied on such deterministic models. This consistency and predictability has conditioned humans to largely trust their computers.

However, the “real world” is highly complex and variable - it is simply impossible for humans to teach machines all of the world’s cause-and-effect relationships. Fortunately, we found out a way to get machines to learn about the world on their own - a field that’s now known as 'machine learning’. In simple terms, computers can now ingest large swaths of unstructured data and learn all of the relationships, patterns, and trends within it on their own. Unlike a simple calculator, these AI systems rely on “probabilistic” models that produce the “most likely” answer for a given input, but not necessarily the “right” one. For the first time, computers became imprecise.

Since then, imprecise AI systems have grown in scale, complexity and adoption, permeating virtually every part of our daily lives. AI tells us what to watch (Netflix, YouTube, TikTok), what to buy (Amazon, Shein), where to eat (Google Maps) and even who to date (Hinge, Tinder). We got so comfortable relying on AI largely because customer value proposition has been so incredibly high that we simply did not question how these products work and why. Besides, the AI tech has been abstracted away from the end user such that most of us don’t even realize that we are interacting with AI on a daily basis.

More recently, the overnight success of ChatGPT has earned AI an unprecedented consumer mind share. For the first time people were overtly aware of the technology they were using, turning AI into a popular dinner conversation topic. However, recent incidents of ChatGPT providing harmful or incorrect answers have raised debates about AI safety and its role in our society.

So how much should we really be trusting our imprecise machines?

Explainability problem

Since the 1990s imprecise AI systems have grown considerably in both scale and technical complexity, turning these probabilistic models into giant black boxes. It is now incredibly difficult to evaluate or understand how a model arrives at its conclusions, a phenomenon referred to as “explainability problem”. Unlike human reasoning, AI cannot ‘explain’ what it has learned, why, or how. It cannot ‘reflect’ on its discoveries or check its errors. We can only observe the results, but not the process how to get there. This is strikingly different from human reasoning which can be paused, inspected, and repeated by other humans.

Increasingly complex and sophisticated AI systems can shape our social and commercial activity in ways that are not clearly understood today. This makes explainability perhaps one of the biggest challenges facing our society.

Consider whether you would blindly trust an AI’s recommendation that can’t be easily audited or verified in these hypothetical situations:

• A medical AI assistant suggests to administer a highly unconventional procedure or treatment for a patient in critical condition

• A credit underwriting AI agent approves a white applicant’s loan request over a seemingly similar loan request from a black applicant

• A social network’s AI moderator flags and suppresses content due to ‘likely misinformation’ that happens to be disproportionately right-leaning

• A military AI strategist recommends sacrificing a few hundred civilian lives in a given battle order to win the entire war

• A hedge fund’s AI system makes an obscure and risky $100M investment

• An autonomous vehicle elects to hit a pedestrian to protect the driver’s life

While all of the above actions might be perfectly valid and appropriate, they also might not be. But on what basis should the AI’s decision be overridden if we cannot understand its calculations? Would an override ever be justified or should we implement the recommendations on faith alone? If not, are we undermining performance that is superior to ours? How do we know that the system’s objective is aligned with that of the user?

Development of procedures to assess whether an AI performs as expected is vital.

Building a safe AI future

Safe deployment of AI systems is a complex topic and an active topic of AI research. This section probably deserves an entire blog post of its own, but I will attempt to offer a simple overview here.

First, we need to develop robust ways to both build safer AI systems AND design frameworks to evaluate how safe or unsafe a given AI system is. This may include:

• Rigorous testing / audit of training datasets for bias prior to deployment

• Continuous model tuning based on human feedback post deployment (RLHF)

• Continuous model tuning based on AI supervision post deployment (RLAIF systems such as Constitutional AI offer a more scalable alternative to RLHF)

• Model evaluation and assessment for alignment issues / honesty / harm using mechanistic interpretability (reverse engineering model output), automated red-teaming (stress testing model), etc.

With that said, technological advancements are only part of the solution - we will also need some form of government involvement to get this right. It is possible that governments may choose to mandate policies that are similar to how they regulate a lot of human professional activity:

• Regular testing requirements to maintain a “license” to operate

• Compliance monitoring requirements

• Formation of an oversight and audit program

• Transparency around model training / dataset engineering

It will be critical for the government to strike the right balance of providing guardrails without stifling domestic innovation. Each country’s approach to dealing with AI technologies will have a tremendous impact on their relative competitiveness in the global arena. Those societies that choose to suppress its adoption through draconian regulation will inevitably fall behind those societies that embrace it.

Closing thoughts

I expect continued advances in AI (and its adoption) to fundamentally alter our reality at a scale comparable to that of industrial revolution. Things that previously seemed impossible will become possible. What took years may take only a few minutes. Human-machine collaboration will become ubiquitous. However, as AI’s role in defining and shaping our society’s information space grows exponentially, it becomes that much more important to develop robust systems and frameworks to monitor and regulate it. To get things right will likely require a joint effort from private enterprise, academia, and government actors.

I leave you with a few questions that I am currently pondering about:

1. What happens when AI produces insights that are true but beyond the frontiers of current human understanding?

2. Who will set the alignment objectives for imprecise systems with unexplainable logic and unavoidable consequences? Big Tech, government, AI itself?

3. How do we solve for competing and/or incompatible objectives/ incentives of all stakeholders involved? (government, private sector, consumers)

4. What is the appropriate role of government in shaping AI policies? How do you keep their powers in check and avoid tyranny / oppression?

5. How do we avoid the formation of an extractive oligopoly of 2-3 AI enterprises?

Let me know what you think! DMs always open on Twitter @leveredvlad

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI, crypto, and other cutting-edge technology.

Subscribe now

Subscribe now

Productivity: the holy grail of success and achievement, the elusive virtue we all strive for. Well, or at least so we tell ourselves. But who are we kidding?

Most of us have more free time than we know what to do with and spend our days aimlessly scrolling through social media, binge-watching Netflix, or playing video games. Indeed, the average American spends 6-8 hours on media-related leisure per day - that’s over 2 billion hours nationwide! We’re not busy. We’re just too damn bored.

This is what Silicon Valley (and me) gets wrong about the current AI hype. The consensus opinion today is that the killer consumer use case of LLMs is a virtual personal assistant in a pocket - something tech bros and VCs would personally love to have to help with their busy days. And while I agree that everyone *will* have a virtual helper on their phone in the near future, their willingness to pay for such a product will be close to zero. For the average consumer productivity is a nice to have, not a must-have. This is partially why products like Notion, Coda, Airtable, Zoom, etc. monetize through enterprise - the consumer ain’t payin’. (NB: I do see a very high institutional willingness to pay for enterprise productivity).

So what is the killer consumer use case, you ask me?

Entertainment & companionship.

Let me explain.

We live in a world where:

• People are bored and have lots of free time

• People are the loneliest they’ve ever been

• People spend more and more time on social media and online games

Guess who is always free, available 24/7, and talks to you whenever you want? I know you are probably rolling your eyes at this point, but please hear me out.

Today, the notion of developing any sort of relationship with a computer is seen as strange and freaky. Software can’t possibly replace human connection, right? But maybe the point is not replacing a human, but just being there for you when a human isn’t. Is it too much of a leap to imagine a world where we all have our trusted confidants in our pockets?

We’re less far off from this future than you think. We already have virtual assistants - Siri, Google Assistant, and Alexa on our phones (notice how they’re free) that can answer questions, set reminders, and control our smart homes. But what if they could offer more? Perhaps they’d entertain you when you’re bored, knowing exactly what you like. Or maybe you’d use them to get something off your chest and get a judgment-free reaction to your situation. We may lean on them when we’re feeling down to cheer us up. This doesn’t sound so bad, huh? Perhaps we could all use an AI friend.

Some companies are already working on some version of this. Replika has already amassed millions of loyal users by offering an AI-companion-as-a-service. Many claim to have developed a strong emotional connection to their Replika (you can see the testimonials on the website). Character.ai is tackling the space differently - they are tapping into the power of user-generated content and social by enabling anyone to easily spin up an AI bot that everyone on the platform can interact with. In 2 months since launch, people exchange over a billion words per day with their Characters.

I predict a lot of market cap will be generated in this space over the coming years.

I leave you with a few open questions to ponder about:

1. How does this get monetized? Monthly subscription? Selling our data?

2. What will normalize/destigmatize being ‘friends’ with AI? Does the tech need to get better (feel more human-like)? Or do we just need time?

3. Will our AI friends always have our best interests in mind? Can they adversely influence how we think? Will they always tell us what we want to hear?

4. How do we make sure we don’t lose touch with humanity? Is it a real risk?

Tell me what you think! DMs open. @leveredvlad on Twitter :)

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI, crypto, and other cutting-edge technology.

Subscribe now

Subscribe now

If you’ve seen the 2011 thriller Limitless, you too probably wondered what your life would look like if you could unlock the full potential of your brain. In the movie, Bradley Cooper plays Eddie Morra, a struggling writer who takes a pill that allows him to access “100% of his brain's capacity”, endowing him with superintelligence, perfect memory, and the ability to analyze minute details and information at incredible speed. Throughout the film, Eddie’s increased cognitive power enables him to vastly improve his lifestyle and generate significant wealth and power.

While that may be a bit far-fetched (we *technically* already use 100% of our brain anyway), the truth is that our brains do have all sorts of limitations and can often be overwhelmed by the sheer amount of information we’re deluged with on a daily basis.

But absent a magic pill, could we achieve the same result as Eddie with technology? Could recent advances in artificial intelligence endow us all with superpowers? That's where the idea of constructing a "second brain" popped into my head. I became curious if there was a way for us to work around our brain’s limitations so that we could become truly limitless.

In this blog post, I will explore the limitations of our existing ‘knowledge stack’ (starting with our own brains), and propose some interesting ways in which AI could help us get closer to Eddie Morra.

Let’s dive in.

Why our “knowledge stack” is flawed

I was browsing through my perfectly organized Notion folders the other day when I stumbled across a page with some startup ideas that I jotted down a few months ago. I got a bit annoyed at myself - I completely forgot I’d come up with these, and what’s more frustrating, I completely forgot that I wrote them down in the first place! What’s the point of having all these perfectly structured folders with backward and forward links if I still can’t easily and readily access my ideas?

The reality is, our brain (“first brain”) is a great place for coming up with ideas, but not so much for holding them. Think of our brain as a vast library filled with endless shelves of information, but not all of the books are within reach. Our memory is limited by what we focus on and what we are able to process, and only a small portion of what we encounter is stored in a way that we can easily access later. To help us remember and recall information when needed, we rely on our "second brain" - our note-taking and other memory aids, which act as our mental librarian, helping us locate the right information and access it quickly.

Of course, the idea of outsourcing some of the workload from our “first brain” into a “second brain” is not a novel one. Note-taking has been an important part of human history and scientific development for millennia. Historically, notes were handwritten but with recent technological advancements (e.g. computer, internet, mobile, cloud), an entire multibillion-dollar industry of “knowledge management software” has emerged. In fact, knowledge management has become such a fundamental skill amidst today’s flurry of information, that someone created a popular $500 course that teaches you how to build a “second brain” using these exact tools. However, even for the most ‘advanced’ expert users several critical challenges remain with regard to using these knowledge management solutions:

• They require ongoing maintenance and organization, which takes substantial cognitive effort and time. If you slip up, the system will quickly become disorganized and increasingly less useful.

• They create unnecessary friction between insight and productivity - searching for the ‘right’ information is time-consuming and distracting (not to say you need to remember what to search for in the first place)

• Each folder or page within the system is siloed, which makes it hard to draw connections between different ideas

• When we write something down, we often do not immediately know how, when (if at all), or in what context we will revisit this idea in the future. This means that we have to preemptively guess the best way to store or file that idea in a way that would maximize its usefulness in the future. Given that this is incredibly hard to get right, we often never end up revisiting the things we store.

I hope that by now I have convinced you of the following:

• We are living in an age of information overload

• Our brains are not good for storing/remembering vast amounts of information

• We increasingly rely on software tools to store and organize our knowledge in a more structured way

• These tools have fundamental flaws that ultimately hinder our performance

Ok, now let’s explore how artificial intelligence can solve virtually all of these issues.

Subscribe now

Building a real ‘Second Brain’ with Artificial Intelligence

Unless you have been living under a rock, you have probably heard of ChatGPT by now - a viral conversational AI assistant that has reached 100 million users in just under two months (likely becoming the fastest product to do so in history). Its success provides a few important takeaways as we consider how to build our second brain:

• The conversational interface appears to be an incredibly user-friendly UI - we like to ask questions and receive direct answers.

• Large Language Models (LLMs) perform really well with tasks such as compare, contrast, suggest, and summarize.

• You can improve the LLMs output by providing additional context/information about the task at hand in your prompt (input)

By all accounts, ChatGPT is a great starting point for a personal virtual assistant that has a vast general knowledge about the world (with some limitations) that can help you with a variety of different tasks. But what if we could augment this powerful tool with our proprietary knowledge? Could it become your on-demand, always-available brainstorming partner? Here are a few examples of what your ‘AI assistant’ could do for you:

• Index and query all of your content across disparate platforms (e.g. Notion, Substack, Twitter combined) in a single conversational interface, making your entire corpus of knowledge much more easily and readily accessible

• Execute a semantic search on your data that focuses on understanding the meaning and context behind your query (rather than just matching keywords), leading to vastly increased accuracy and relevancy of output

• Convert unstructured language/notes into a self-sustaining knowledge graph of all of the concepts, entities, and their relationships within the dataset. As new information is added, the knowledge graph would automatically update.

• Identify new threads in your thinking and link previously unrelated ideas or insights together

• Greatly diminish the ongoing maintenance and organization needs of your knowledge system (and reduce the associated cognitive load)

To be clear - this hyper-personalized AI assistant will not replace you - but it will augment your thinking and complement your first brain’s ability to observe, reflect, generate new insights, and make new connections between ideas. All done via an intuitive conversational interface - it’s like talking to a ‘real’ personal assistant.

So how do we achieve this?

Current tech & its limitations

The stack to get us to a version of this already exists today. All we need is:

1. A ‘general purpose’ LLM (could be GPT-3 or something else) pre-trained on a large dataset - this will be used as a foundation of our assistant

2. Our personal corpus of knowledge (such as our Notion or Evernote notes)

3. An engine to process our data to make it ‘searchable’ (a number of open-source frameworks exist, and most rely on adding ‘embeddings’ to the dataset; examples include Haystack, Weaviate, GPT Index)

4. A framework to wrap the above into a conversational chatbot with memory (LangChain is an open-source project that is well suited for this)

While this may sound complicated to the non-technical folks, the main point here is that the tech to achieve this result exists already - it may not be perfect - i.e. some improvements around embeddings, prompting, etc. is required to optimize the bot’s performance - but it is already a massive improvement to the status quo.

In fact, some hobbyists are already tinkering in this space. For example, Dan Shipper from Every has uploaded 10 years of daily journaling into a GPT model and was able to receive accurate answers to some deeply personal and nuanced questions. The model was able to see patterns across the 10 years of daily data that our brains are simply unable to. This experiment, albeit unperfect, provides a glimpse into what hyper-personalized assistants may look like in the near future.

Limitations

I would be remiss if I didn’t mention some of the limitations that we have to figure out in the near term. The most obvious one is “model hallucinations” - today’s imperfect LLMs sometimes conjure up an answer that is completely made up (but sounds believable!). If we rely on these assistants for suggestions or fact recollection, we need to be confident that these hallucinations do not occur. Another issue is that of model bias of the general purpose model itself - what if it leans a certain way politically and its output reflects that? What about biases that are less noticeable? We need a way to deal with this (or at least be hyperaware of this risk). Finally, there is a philosophical question - can an AI model put thoughts in my head? At what point is an idea truly mine? What are the dangers of relying on this brainstorming partner?

Predictions

To conclude this piece I wanted to throw out a few predictions on this topic:

1. The conversational interface will become ubiquitous and will become a major way we interact with data and information across ALL workflows

2. Every major platform (Windows, MacOS, Gmail, Notion) will implement some sort of LLM-augmented ‘search’ for its users, but each dialog agent will have limited access to the data beyond its own platform (opportunity for aggregation)

3. “AI assistants” will go after Enterprise vertical applications - data retrieval / knowledge sharing across Sales (CRM applications), customer service, HR, etc.

4. Similar to how Google Search removed the need of knowing encyclopedia facts, LLMs will greatly reduce the importance of remembering things - humans will simply outsource all that computational load to their “second brain”.

5. As a result, humanity will be moving around orders of magnitude more data between different users/entities in ways never imagined before - this will require new and better ‘pipes’ / infrastructure/tooling for data management

What do you think about the future of AI assistants? What happens to the world if all of us have access to Eddie Morra’s superpowers? Will it be the greatest equalizer, or will we experience rapidly growing inequality never seen before?

Tell me what you think! DMs open. @leveredvlad on Twitter :)

PS If you would like to collaborate on building an AI-enabled “second brain” please reach out - currently working on something new.

If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI, crypto, and other cutting-edge technology.

Subscribe now

Subscribe now

OK, I may not *actually* be besties with ChatGPT, but this chatbot has definitely woven its way into my daily routine. From meal planning to workout ideas to travel itineraries, ChatGPT has been my go-to for all things decision-making. But last week, when OpenAI's website crashed and I couldn't access my trusty assistant, I was shocked at how lost and unproductive I felt without it. And this was only after using the tool for a month!

In this post, I want to delve into my relationship with ChatGPT and consider the larger implications of AI-powered apps as they become more widespread. While "generative AI" is all the rage on Twitter, we're still in the early stages of adoption. What will the world look like when these types of apps are as common and ubiquitous as the internet is today?

Let’s dig in.

Magic

When was the last time you experienced magic?

For me, it was December 1, 2022. when I first tried ChatGPT. Everyone on Twitter was raving about it, and one try was all it took to blow me away. Sure, there have been plenty of incredible products that wowed me on first use - think Uber, Spotify, and Amazon's same-day delivery. But ChatGPT felt truly magical.

Here's the weird thing: the technology behind ChatGPT has been around for a while. OpenAI released GPT-3 (ChatGPT's predecessor) in 2020, and GPT-1 was developed in 2018. But generative AI wasn't really on the radar outside of tech-savvy circles until now.

So what was so different this time? Why did *this* product go so viral so quickly? While there is no one definitive answer, I could think of three likely reasons:

1. User interface - this was by far the best UI put on top of a LLM (large language model) to date - it simple, intuitive, and felt like you are talking to a person.

2. Timing - other generative AI technologies like DALLE-2 and Stable Diffusion had already gained millions of fans, and LensaAI released a viral AI consumer app around the same time. AI was suddenly on everyone's minds.

3. Go to market - the easy sign-up process and the fact that it was free made ChatGPT spread like wildfire. If users had to enter their credit card information to try it, it's unlikely it would have become as mainstream as it did.

Whatever the reason is, ChatGPT captured the zeitgeist, amassing over a million users in a week. Not too shabby at all.

Superpower

I started using ChatGPT just for fun, asking it weird questions and trying to trick it (“explain the theory of relativity in style of Tupac Shakur”). But it didn't take long for me to realize it could actually boost my productivity. As I got better at using the tool, I gave it more complex tasks. Soon, ChatGPT was open on all my devices 24/7.

In particular, there were two use cases that were especially game-changing for me:

1. Learning new things. Importantly, I don't ask ChatGPT for specific facts or explanations because the error rate is still too high. But it's great at helping me structure my approach to learning new subjects. For example, if I wanted to learn about machine learning, I'd ask ChatGPT to produce a detailed curriculum with all the relevant topics and a list of concepts and definitions I should know. It would give me an excellent framework to follow, and I could use Google and YouTube to learn the details. Without ChatGPT, I'd have to do a lot of work to figure out where to start and how to progress.

2. Brainstorming. Perhaps the most valuable way ChatGPT has helped me is by being my go-to brainstorming buddy. Whether I'm exploring new ideas (“give me a few clever ways to prove or disprove the following hypothesis” or riffing on strategic decisions (“how would you compare Strategy X and Strategy Y? Provide a detailed list of arguments for and against each approach and give me a framework for objective comparison”), ChatGPT has my back. It's become my confidant for all things creative and decision-making.

With ChatGPT I feel like a 10x version of myself. I feel like I have superpowers and I can do anything. Previously, I would bother my friends with these questions, but now there is no need - I have a personal assistant in my pocket that allows me to get most of the way there on my own.

Addiction

”Write a thoughtful Christmas card message for my father in law”

“What should I gift my mother for Christmas?”

“What’s the best pick up line on Raya?”

Over time, I started using ChatGPT for everything, all the time. It was a digital black hole - once I started, it was hard to stop. I became addicted to it.

My dependency on ChatGPT became clear to me when I first experienced heart-wrenching despair after I was unable to access it (the server was down due to exceptional demand). I recall pacing back and forth in my living room, muttering to myself, and trying to come up with ways to fix the issue (spoiler alert: pacing and muttering did not help). I would furiously refresh the page, and attempt to log-in via incognito mode, or from a different device - all to no avail. I simply had to wait.

Is this healthy? Am I the only one who finds this friendly chatbot so addicting?

Future

For now, most people (I think) are using ChatGPT for fun or productivity tool.

But what’s next?

Enterprise

The change seems mostly positive at the enterprise level. Perhaps corporations adopt these tools for ideation and experimentation - goodbye whiteboards and hello group chats with a chatbot that allows a team to rapidly brainstorm new ideas and riff with each other (and AI).

Perhaps first we also see these chatbots penetrate customer service. Instead of a 1-800 number and a long wait time, you will be resolving your issues with a very well-trained chatbot that will answer all your pressing questions and concerns.

All told, this technology will likely supercharge some people, make others redundant, while making the overall business more efficient (more output with less input). This seems to me almost inevitable.

Consumer

The impact on consumer is much less obvious and straightforward. Beyond a simple productivity tool, I envisage the relationship between people and AI to grow increasingly complex.

AI bots may first become personal assistants, handling mundane tasks like making appointments and filing taxes. It will also likely transform education - everyone will have a personal tutor in their pocket - at an affordable cost. These changes to me seem universally positive.

But what impact will it have on our relationships with one another?

Loneliness

Today’s loneliness epidemic is a real and pressing issue that affects millions of people around the world. With the rise of technology and the decline of face-to-face interactions, it's becoming increasingly difficult for people to find a genuine human connection. That's where AI-enabled chatbots may come in. At first, these chatbots may become our therapists or life coaches - a way for people to get guidance and support without having to leave their homes or talk to a real friend. Perhaps it’s more convenient, perhaps it’s less scary - after all, an AI cannot judge you and is unbiased, right? Or will they tell us what we want to hear? Could they give us bad advice?

As technology advances and AI starts to feel more “real”, people may start to seek companionship (whether platonic or romantic) from these chatbots. Lonely people around the world already pay millions to speak to strangers on the phone, pay for cuddling, and sex. Will AI make this fake companionship more accessible? More real?

It's a dangerous game, relying on an AI to fill the void of human companionship. Who knows what kind of messed up stuff they might be saying to us behind closed circuits? "You're the most amazing person I've ever met!" they might say, while secretly plotting world domination. Okay, maybe that's a bit of a stretch, but you get the idea. These chatbots may seem like the perfect companions, always ready to listen and offer words of wisdom, but it's important to remember that they're just machines. They can never truly replace the warmth and complexity of real human interaction. Or can they?

Closing thoughts

So, what's the solution? Do we banish these chatbots to the depths of the internet and go back to the good old days of human interaction? Probably not. The genie is out of the bottle and we will have to learn to live with it.

To be clear, I am not against AI (or any technology for that matter). With every new breakthrough, there will be benefits and threats. Technology in a vacuum is neither good nor bad. Society learned (kind of) to live with the Internet, and I can only expect us to learn to live with AI as well. What makes this tricky, is that the technological progress is exponential, resulting in our lives changing at an increasingly fast pace. It’s more important than ever to think about potential issues proactively (rather than reactively) to avoid being too late.

If you enjoyed reading this, subscribe to my newsletter! I regularly write deep dives and business model breakdowns about crypto, AI, and other cutting edge technology.

Subscribe now

Join our growing community by subscribing here:

Is the next frontier of fashion… digital?

Nike certainly seems to think so. On December 13, 2021, the iconic athletic apparel juggernaut announced its acquisition of RTFKT, a “digital fashion brand” to “extend Nike’s digital footprint and capabilities”. The tie up caused quite a stir - both in the worlds of fashion and crypto. What does a leading shoe manufacturer have to do with an NFT studio? Are we really going to be wearing Nike sneakers in the metaverse?

In this essay I will explain why Nike is betting on web3, what it means for the luxury apparel industry, and why other brands may soon follow suit. In particular, I will explain how RTFKT got started, why Nike acquired it, and how it plans to leverage RTFKT to disrupt the $3 trillion apparel and footwear market.

This essay will cover:

1. Origin: What is RTFKT?

2. Digital Fashion

3. Let’s get… Phygital

4. Entering The Metaverse

5. Where do we go from here?

Let’s dig in.

Chapter 1. Origin

Founded in early 2020, RTFKT describes itself as a “creator-led organization - a community of collectors, investors, artists, gamers, and fans”. The company was founded by three friends deeply involved in gaming, fashion, and streetwear:

• Benoit Pagotto (worked in both luxury and gaming)

• Chris Le (worked in VFX)

• Steven Vasilev / Zaptio (worked in social media, streetwear, and tech)

The original idea was to merge sneakers and gaming culture. The inspiration came from a simple realization that sneakers to sneakerheads are what ‘skins’ are to gamers - they are both collected largely for its aesthetic and scarcity value, and both act as a social signal and status symbol. The only difference was that one existed in the physical world, while the other was purely a digital ‘virtual good’.

The trio wanted to capitalize on the secular trend of people (especially Gen Z) spending increasing amounts of time playing and socializing in virtual worlds, where they already use digital collectibles (mostly in gaming) to express their style and digital identity. Indeed, gamers already spend the majority of ~$150+ billion per year on in-game purchases of virtual goods - something the team believed will eventually transcend the bounds of games into social media and other online applications.

The emergence of NFTs couldn’t have come at a better time - the technology enabled users to verifiably buy, hold, and flip scarce digital goods (at the time mostly JPEGs) - something sneakerheads have already been doing with physical shoes for decades. Another advantage of “digital fashion” over its physical counterpart is that the product cycle is much faster - a new digital product can be shipped in a matter of days, whereas a traditional physical brand would require 12-24 months to ship a new product. This would enable RTFKT to iterate and keep its products up to date on current trends much faster.

Equipped with a provocative idea, some promising social media traction, and a clear roadmap, RTFKT went on to raise seed financing from a stable of high profile investors with a16z calling RTFKT the “Digital SUPREME”.

So how did RTFKT get on Nike’s radar?

Chapter 2. Digital Fashion

RTFKT (at the time yet to be officially incorporated) started to gain significant attention by sharing unique custom digital sneaker concepts (many of which featured the iconic Nike Swoosh 👀). For example, using a blend of CGI and AR (augmented reality), RTFKT shared a concept for Marvel-inspired Iron Man and Thanos sneakers:

Building on the momentum, RTFKT went on to share other custom digital sneaker mashup concepts (not official collabs) such as Dior, League of Legends, PS5 and even Tesla’s Cybertruck which went viral when a photoshopped photo of Elon Musk rocking the sneakers was shared on social media (RTFKT later 3D-printed a custom version for Musk, while the 1-of-1 NFT was auctioned off for 30 ETH or $90,000):

The company continued to innovate, dropping digital wearables that could be used in virtual worlds (e.g. Sandbox, Decentraland), officially partnering with famous brands (such as Atari), and experimenting with other variables (e.g. Metajacket). On the back of its early success, RTFKT moved on to doing some high-profile collaborations, which included a limited collection of digital sneakers designed by a famous NYC streetwear designer Jeff Staple and a collaboration with a crypto-native artist FEWOCiOUS. The latter collab went on to make headlines by selling out in just 7 minutes and generating $3.2 million in revenues.

RTFKT’s last big splash before the Nike acquisition was Clone X - a highly anticipated limited collection of 20,000 anime-inspired digital avatars designed by a famous Japanese artist Takashi Murakami. The collection went on to become one of the most successful NFT projects to date, generating nearly $800 million in secondary volume and will eventually become the central part of RTFKT’s roadmap (to be discussed further in Chapter 4).

By the end of 2021, RTFKT has established itself as a major NFT brand alongside other ‘blue chips’ such as Bored Ape Yacht Club and Crypto Punks.

And then… Nike shocked the fashion and crypto worlds by announcing its acquisition of RTFKT in December 2021, ushering a new era for the fledgling NFT studio.

Chapter 3. Let’s Get… Phygital

MNLTH / Cryptokicks Campaign

Two months after acquiring RTFKT, Nike dropped an ambiguous NFT “MNLTH” to RTFKT’s CloneX holders. The NFT was a metallic cube decorated with Nike’s iconic Swoosh. In April 2022 Nike finally unveiled that each cube contains a pair of virtual sneakers called Nike Cryptokicks based on the famous Nike Dunk. The campaign was a huge success - MNLTH line of NFTs generated ~$300+ million in secondary sales, earning Nike a cool $30m+ million in royalties. Even amid a major slowdown across the entire NFT market, the current floor price of a MNLTH cube is still sitting at 2.5 ETH ($4,000).

The Cryptokicks campaign is interesting for a number of reasons:

1. Distribution. The NFT was ‘airdropped’ for free to RTFKT’s NFT holders, rewarding existing RTFKT collectors. Each MNLTH NFT could be burned in exchange for Nike Dunk Genesis Cryptokicks, one Skin Vial Evo (allows holders to upgrade their Sneaker with limited edition skins), and one MNLTH 2 (contents not revealed at the time of writing). This demonstrates how an established brand can directly connect with/target a specific demographic (in this case existing RTFKT collectors).

2. Marketing. Collectors had to complete ‘quests’ in order to get a ‘sneak peek’ into the cube’s contents - as quests were being completed, the cube would slowly ‘break’, exposing a glowing blue object inside. This clever, interactive campaign allowed Nike to generate significant buzz and anticipation ahead of the reveal (the pre-reveal floor price of the MNTLH reached 8 ETH ($24,000). )and also demonstrated some of the cool things that can be done with NFTs from a tech perspective.

3. Value Capture. Nike was able to generate significant royalty revenue from secondary sales derived from a royalty fee programmed into the NFT’s smart contract:

   

4. Composability. The Cryptokicks NFTs are generative - they are customizable with RTFKT-made skins, which enable holders to upgrade their sneakers to change how their sneakers look (similar to Nike By You for physical shoes). At the moment, the Skins are created by well-known designers, but RTFKT has plans to enable other creators, artists, and members of the community to release their own skins in the future (and capture their own royalty revenues).

While Cryptokicks today are purely a digital collectible, RTFKT has hinted that holders will be able to “forge” their NFTs into IRL Nike Dunks in the future (these can also leverage composability by incorporating a screen that could showcase a customized ‘skin’ representing the NFT “Vial”).

RTFKT x Nike AR Genesis Hoodie

Following the success of the Cryptokicks campaign, RTFKT released its first truly “phygital’ product - RTFKT x Nike AR Genesis Hoodie, which was released as both a Clone X wearable and a “forgeable’ physical piece. In addition, each hoodie featured a QR code which links to the AR (augmented reality) experience - tracking the 3D movement of the garment. As was teased in the RTFKT release video, collectors will likely be able to further customize their hoodies with additional wearables (such as wings depicted below).

The Hoodie was distributed via a “private mint”, which means only the owners of Clone X and Cryptokicks NFTs could participate (1 Hoodie per each existing NFT held). This distribution method achieved:

1. Created exclusivity around the release as only RTFKT collectors were eligible to purchase the Hoodie

2. Rewarded existing collectors with first access (and ability to flip the NFT to a higher bidder in the future - with RTFKT capturing 8% of the secondary sale proceeds).

3. Ensured built-in excess demand for the product as there were only 8,888 of Hoodie NFTs available vs 20,000 Clone X and 12,000 Cryptokicks in circulation.

Other Phygital Products

RTFKT didn’t stop there - shortly after the Hoodie release, the team went on to release a phygital Space Drip sneaker collection (just as with hoodies, these NFTs could be ‘forged’ into IRL sneakers). Importantly, the designs were created by artists in the community, who will earn royalties on all secondary sales.

Next, RTFKT came up with an even more provocative idea - “web3 perfume”. Inspired by potion crafting in video games, RTFKT partnered with a Swedish fragrance maker Byredo to release ALPHAMETA, extending its product offering into perfumes. Similar to Genesis Hoodie and Space Drip collections ALPHAMETA is a phygital product consisting of an NFT of a wearable digital aura for the Clone X avatar as well as an IRL Byredo perfume.

Yet again, RTFKT pushed the boundaries at the intersection of technology and fashion. In this case, NFT holders will be able to create personalized scents using any of the 26 feeling-based ingredients named after emotions (e.g. Acuity. Naivety, etc). These digital scents will then be forged into a limited edition personalized Byredo perfumes, packaged in custom NFC-tagged bottles.

All of these innovative products were setting up the stage for RTFKT’s next big act - a creation of its own unique ecosystem, in which every digital collectible could be worn by the collector both IRL and in the virtual world by its Clone X avatar.

It’s time to enter the metaverse.

Chapter 4. Entering The Metaverse

In August 2022 RTFKT finally unveiled its next major product launch - 10 Clone X collections of phygital goods, that would range from caps to jackets and sneakers. Clone X avatars and Pod X virtual rooms will be leveraged to display collectors’ apparel, sneakers, and other wearables in the virtual world RTFKT has been building behind the scenes. Each good will be inspired by the “DNA” (e.g. Human, Robot, Angel, Demon, etc) and Trait of the collector’s Clone X avatar. For example, by owning a Human Clone X with Murakami Drip, a collector can mint up to 31 total items (click here for full lookbook).

Each physical item will be embedded with NFC chip which would be linked to the digital NFT, and will be used to prove ownership and authenticity, truly merging the digital and physical worlds.

The mint is currently ongoing, and only time will tell how successful and iconic this first collection will become - we are still in the early innings of the phygital revolution. If this trend continues and becomes ubiquitous, these early, first collections will become canon and really valuable to collectors.

This is where the Nike’s ownership is an undisputable competitive advantage - RTFKT is the only crypto-native brand that has the muscle, the resources, and the know-how of a major apparel manufacturer behind it. This is exactly the kind of product that only RTFKT can execute and deliver today, and in my opinion, this is only the beginning.

Where do we go from here?

So where do we go from here? The public mint for SZN 1 drop is still ongoing and only time will tell if it is successful. If the collection does well, it would only make sense for Nike to continue expanding its phygital offerings. The digital twin model (discussed by Nic Carter in this article at length) is a fantastic use case for NFTs, and I expect most luxury brands to eventually join this trend.

Benefits & Opportunities

So what exactly can Nike do with RTFKT moving forward?

1. Trustless secondary marketplace with royalty capture. Physical products that are linked (via NFC chips) to a digital twin (an NFT) can facilitate a safe and counterfeit-resistant re-sale market. An NFT can host a wide range of data - from proof of authenticity to the product’s vintage, total supply, and past ownership. Further, NFTs can come with pre-programmed royalty capture mechanisms for both Nike and the creator of the product’s design.

2. User-generated fashion + composability. Platforms like YouTube, TikTok and Roblox have already demonstrated the power of UGC (user-generated-content) and the network effects this can drive for the business - Nike has the opportunity to do the same with apparel. Nike/RTFKT can tap into the creativity of its community / consumer base to enable artists to contribute their own designs to the products by the brand. The community (token-holders) can even vote on what designs get accepted into physical production. A brand for community, by the community can have meaningful advantages over traditional brands as it scales the creativity beyond that of the RTFKT design team, while rewarding top contributors and bringing the fans closer to the brand. Additionally, NFTs unlock composability enabling partnerships with other projects, similar to Tiffany’s highly successful CryptoPunk collaboration.

3. Token-gated experiences. Nike can offer the owners of its apparel to access exclusive IRL events and experiences - all you have to do is scan your phygital wearable. This in turn could unlock other perks - like discounts on other products or first access to highly anticipated future drops.

4. Authenticated digital flexing. Given the proof of provenance that NFTs provide, phygital products enable collectors to verifiably show off their collections - this is something sneakerheads cannot easily do today with their possessions with Instagram / other social media platforms.

5. Metaverse brand presence. While mainstream adoption of Metaverse is still a few years away, RTFKT’s technology provides Nike with a roadmap to establish a strong presence in the virtual world, which will become increasingly important as our lives continue to become more and more digital.

Risks & Challenges

However, as with any disruptive technology, there will also be a few challenges along the way that will need to be resolved:

1. Abstracting ‘crypto complexity’ away. Today, RTFKT has catered to very crypto native customers, who understand how NFTs / crypto / wallets work. If Nike decides to bring these types of products mainstream, the onboarding process (e.g. setting up a wallet, fiat on-ramps, purchasing an NFT) must be made as simple as it is ordering a regular Nike sneaker online. Without this, phygital goods will remain a niche product catering to a relatively small crowd.

2. Offering true digital utility. At the moment, the virtual world RTFKT is building is largely a cool gimmick - there isn’t much utility or purpose for any user to use this ‘metaverse’ on a regular basis. In order for the digital wearables to truly offer utility (e.g. social signaling / digital flex), Nike will either have to develop a robust interactive and fun virtual world (very hard) or provide a bridge for users to use these items in other virtual worlds (e.g. Fortnite, Roblox, etc). This may not be as far fetched as it may seem - RTFKT already uses Unreal game engine (developed by Fortnite maker Epic Games). Absent true interoperability / compatibility, the NFTs will remain merely a proof of purchase / proof of authenticity.

3. Avoiding brand dilution. The reason digital sneakers were such a hit with sneakerheads and crypto enthusiasts alike is because they are scarce and signal either being early or being rich (similar to high end luxury goods). As Nike / RTFKT looks to scale the brand and its product offerings, they are risking diluting the brand as each new collection makes the preceding one less scarce and exclusive. This may result in significant brand dilution, and in turn, reduced demand / price of these phygital goods. Maintaining scarcity (both digital and physical) will be key to the long-term success of the RTFKT brand. One way to do so is for Nike to continue on its strategy of exclusive limited editions collabs, such as the recent Luis Vuitton x Air Force 1 x Virgil Abloh drop, which featured IRL experiences, digital galleries, as well as AR functionality enabled by Snapchat Lens. NFT technology also enables Nike to conduct auctions for the super rare drops, allowing the brand to capture value that is currently leaked into the secondary market on eBay and StockX.

Conclusion

Overall, I have been extremely impressed with RTFKT’s execution to date and I believe the Nike acquisition gave it the necessary resources to scale and push the boundary of what’s possible. Similarly, Nike benefits from RTFKT’s crypto-nativeness and track record, as it prepares itself for a world in which the society spends increasing amount of time, energy, and resources on its digital identities and how they present themselves online and in virtual worlds.

I believe RTFKT has all the right tools and team in place to continue growing its ecosystem and I can see a world in which it becomes the biggest NFT brand over the coming years, as it continues to bridge the digital and physical worlds, and imbuing its holders with actual utility (gated commerce/experiences/events + high quality physical apparel). Time will tell how this experiment plays out, but I am certainly excited to see what the future holds for this highly innovative brand.

Disclosure: None of this should be interpreted as financial advice. Do your own research.

Join our growing community by subscribing here:

Most investable assets have three key characteristics:

• Entry price - how much you pay to acquire the asset

• Return of capital - how much capital you earn back over the hold period (e.g. dividends for stocks, interest payments for bonds, rent for real estate)

• Return on capital - proceeds at exit in excess of the entry price you paid

Today most NFT investors focus only on entry and exit price - the vast majority of NFTs are owned for speculative reasons in hopes of selling the asset at a higher price in the future. Some projects also offer return of capital via airdrops (e.g. Yuga Labs giving away MAYC or Otherside land plots to BAYC owners). All told, the end goal remains the same - for most NFT market participants the goal is to end up with more money than what they started with.

But what if we abstract the financialization of NFTs away? What if we removed both the entry price (NFT was free to claim/mint or was earnable) and the exit price (NFT was non-transferable or there was no market for it)? In other words, what can an ownership of an NFT offer to its holder other than financial gain?

In this essay I am going to work through this thought experiment and offer some ideas of non-financial applications of NFTs. I hope to cover the following:

1. Four key types of NFTs (it’s not just JPEGs)

2. Non-financial NFT use cases (Art, Access, Commerce, Identity)

3. Why blockchain? (why do we need web3 for this?)

4. Conclusions (where do we go from here?)

NFTs are not just JPEGs

When most people think of NFTs, they imagine those small profile picture (PFPs) that they see on Twitter (and soon Instagram) - they appear as a simple JPEG (many even wonder why would someone pay thousands of dollars for something that can be ‘right-click-saved’ - topic for another essay).

In reality, NFTs are typically just a link between an on-chain item with some off-chain data (picture, text, music, video, etc). With that in mind, there are some important distinctions between the different types of NFTs available today:

1. Transferrable vs Non-Transferrable NFTs. Most NFTs today are transferrable tokens that one can buy, sell or trade to another person/wallet. These represent the vast majority of NFTs that you might have heard of or seen in the news. Less known, there are also non-transferrable NFTs (or as Vitalik Buterin dubbed them “Soulbound Tokens” or SBTs). These NFTs cannot be sold or transferred to another wallet (though some may be revoked).

2. Static vs Dynamic NFTs. Static NFTs are those that have permanent immutable traits - for example, a Bored Ape or Doodle. Dynamic NFTs on the other hand can be modified - for example an NFT Sneaker from the viral move-to-earn app STEPN, in which users could “level up” / upgrade their Sneakers as they progress throughout the game.

   

These characteristics open up a wide design space for a myriad of different use cases and industries in which these NFTs can be used in new and creative ways.

With that out of the way, let’s explore some of these in more detail.

NFT Use Cases

NFTx Art/Collectible

• Aesthetic/Sentimental Value. I’m listing this as the first use case since it’s the one that is probably easiest to grasp and understand. Some NFTs can valuable to its holder for purely aesthetic/sentimental value - this is very similar to how one would own a piece of art in their apartment that they can’t / don’t want to necessarily sell.

• Intellectual Property. Going one step further, owning an NFT can endow its holder with unique IP rights, enabling them to innovate and create new derivatives (clothes, accessories, games, etc.), providing an avenue for potential monetization .

• Digital Flexing. As our lives become increasingly digital, so will our needs and/or desires to flex. Whether we like it or not, the items we own (physical or digital) are signals for access, wealth, style, etc. Today, most people flex online via social media (e.g. Instagram photos/stories) or (increasingly) with art/PFPs/digital collectibles. In the future, people will be able to flex with verifiable purchases they made in real life. For example, instead of wearing a potentially fake Rolex on Instagram you may display a proof of purchase of a Rolex from an Authorized Dealer - more on this later). The proof itself does not have any financial value, but it may have social/signaling value to the holder. Social media giants are already adding NFT functionality to their platforms:

  • Twitter was first major network to add NFT capability to its platform

  • Instagram has announced it will add NFTs to its app in 2022

NFTs x Access

• Memberships - NFTs can be used as verifiable memberships in various clubs, programs and communities. These scarce tokens can endow users with exclusive access to gated Discord channels / online communities and even IRL experiences. Most current iterations of this use case use transferrable tokens (you can buy/sell them), but it is likely in the future we will have communities which will provide access to its members for free in exchange for doing something (being a contributor, being first/early, working in a certain industry). Some examples of these communities today are:

  • PROOF - A private group of 1000 dedicated NFT collectors and artists

  • GODA - The trusted and curated source for leading contemporary artists looking to explore digital as a new medium

  • Admit One - A 1,000 member private group of Web3 enthusiasts, builders, and thought leaders

• Artist/Creator/Brand engagement - Fans/customers earn NFTs via desired behavior (e.g. being in top 1% of artist’s content consumers or attending X number of games of their favorite sports team), and in return they receive first dibs on front row tickets or exclusive merch / discounts / deals. The NFT itself may not have explicit financial value, but it allows its holder to connect with their favorite brands on a deeper level and be rewarded for being a true fan.

• Loyalty programs/perks - Generative NFTs can be used as a “digital punch card” which would provide its holders with discounts on coffee/drinks, guaranteed reservation spots, etc (Tascha wrote an awesome piece on this). There has been already some experimentation in this space:

  • Coachella selling limited edition NFTs for life-long pass to its concerts

  • E1even nightclub launched NFT passes that provide its holders with exclusive access to its IRL events at its venues

  • Superlocal is an early stage Foursquare x Web3 startup that will enable small businesses to create and mint NFT-based loyalty programs.

NFT x Commerce

• Token-gated commerce - this is similar to the “access” use case described above - stores/artists/brands can restrict sale of certain items to holders of specific NFTs. Shopify already added an API that enables this functionality.

• Digital Twins - this concept has been brilliantly explained by Nic Carter in his deep dive on NFT use cases in the fashion space. In essence, NFTs can be used to redeem a previously purchased physical good. This means that goods can be both physical and digital goods (hence the “digital twin”). Some examples Nic mentions in his article are:

  • Nike/RTFKT - NFT holders can ‘forge’ a real-life sneaker

  • Tiffany - CryptoPunk NFT holders can purchase a necklace for 30 ETH resembling their own Punk. This is the first example of a known luxury brand leveraging existing 3rd party IP (composability!)

• Anti-Counterfeit/Authenticity - another fascinating use case Nic mentions in his essay is one of anti-counterfeit/proof of provenance for luxury items. Brands can leverage chip-embedded links connecting the physical good to the NFT proving its authenticity. A number of brands are tinkering in this area - Breitling and Vacheron Constantin invested in a ‘luxury good passport’, LVMH, Mercedes Benz, Cartier and Prada joined a consortium to develop a global standard for digital luxury goods.

• Innovative Marketing - Given that NFTs can be ‘airdropped’ into users’ wallets, there is an interesting design space for new marketing campaigns of brands targeting specific demographics (e.g. members of particular communities, or power users of certain apps). For example, Sephora could airdrop NFTs redeemables for in-store discounts to holders of World of Women NFT. To date, airdrops have been largely used by crypto-native companies, but it is only a matter of time until “web2” brands start looking at how they can enhance their marketing efforts using web3 primitives.

NFTs x Identity

Identity is probably the biggest, most fundamental application of NFTs (and probably Web3), yet it is one that is least built out today. As our lives go increasingly digital (and perhaps eventually onto the Metaverse?), digital identity will become more and more important. Today, digital identity is mostly associated with our PFPs or avatars on social media platforms and games. In the future, our digital identities will extend far beyond what you look like on Twitter:

• Credentials - non-transferrable NFTs can be used as verifiable credentials such as university degree, age, accredited investor status, vaccination records and others. In the future, we will likely have something akin to a private “digital backpack” which will house safely and securely a number of important on-chain credentials which we will be able to reveal selectively to whatever service we deem appropriate. A bunch of startups have emerged in this space:

  • Quadrata - a digital passport for KYC/AML/proof of human attestation

  • Footprint - a secure identity verification platform

• Reputation - somewhat adjacent to credentials is reputation - this can be anything from on-chain proof of involvement/contribution to various communities (Reddit is already experimenting with this) to on-chain credit scores (which would unlock cross-border credit history/proof of credit worthiness)

  

Why blockchain?

If you made it this far you may be wondering, why would we need blockchain technology to unlock this functionality? Can’t this be done with existing web2 solutions? In reality, most of these things can indeed be done without blockchain/crypto. However, there are three properties that web3 primitives unlock:

• Interoperability. This is by far the biggest unlock that comes from using blockchain - it introduces standardization and interoperability across the entire ecosystem. This benefits both brands and consumers alike. Brands benefit from composability, enabling them to easily deploy new products/campaigns on top of existing applications as well as seamlessly collaborate with other brands. For consumers the benefit is true ownership of their data/assets and the ability to take these assets with them from one application to another without worrying about compatibility/integration.

• Transparency/Auditability. Blockchain ultimately is a public ledger - it makes it very easy to verify ownership of a particular digital asset and trace it back to the original issuer/owner directly on the blockchain. Importantly, the ledger needs not be fully decentralized (though the libertarian in me would love for this to become the status quo) - web2 incumbents may opt in to use centralized, permissioned ledgers to adopt web3 primitives - time will tell where the various technological solutions/paths will ultimately converge.

• Liquidity. I know I started this piece deliberately omitting the ability to buy/sell digital assets - the primary use case of this property today is speculation and profiteering. However, when overlayed with genuine utility, NFT liquidity allows the asset to end up in the hands of a holder that values it most (this is true only for transferrable NFTs). I may write another piece on this topic and talk about use cases where liquidity may be important - such as ticketing, textbooks, etc.). In the end, I expect NFT liquidity will be used the same way we use Craigslist or Facebook Marketplace - for people to simply offload something they no longer need to somebody else who does.

Closing thoughts

Some of these applications may appear futuristic - and in many cases it is because they are. Brands and artists will need the develop the necessary knowledge and understanding of digital ownership, blockchains, and web3 primitives. Beyond that, they will need the tooling and infrastructure in place to easily deploy their ideas without having to hire an army of blockchain engineers and developers. This will take time. Some will develop this expertise in-house, while others will look to crypto-native startups and advisory firms for help. On the consumer side, the industry needs to abstract any and all complexity associated with interacting with blockchains - we need to get to a point where receiving, sending or redeeming an NFT is as easy as sending a Venmo. Digital wallet wars are still in its infancy phase, and there will be a lot of solutions available in the market offering simple UI/UX for the everyday person to easily access and interact with digital assets and their own digital identity. All of these challenges are critical, but not insurmountable. They create a significant business opportunity for builders and founders to bring this fascinating technology to the masses and I am confident that they will.

We are still very early.

Disclaimer: None of this is financial advice. Do your own research. The author may hold long positions in some of the digital assets mentioned in this essay.

Join our growing community by subscribing here:

Whenever bitcoin comes up in conversation (which happens increasingly often these days), I often hear skeptics push back with statements like

• "Bitcoin has no intrinsic value"

• "Why would I ever invest in something that doesn't produce cash flow? I'd rather own stocks"

• "Bitcoin is too volatile to ever become a global currency or a store of wealth"

While only time will tell whether the Bitcoin experiment will endure over the coming decades, I began to ponder about the roots of such widespread skepticism around the world's largest cryptocurrency. Do people doubt the underlying blockchain technology or is there something else at play here?

Taking a step back, I realized that most people don't really think about what money is - or what underpins it - to most of us money is just something we all grew up using to pay for stuff. Especially for those living in the West, money has been a relatively basic part of their daily life - they haven’t experienced large scale currency devaluations or had to worry about high inflation (until recently…).

This post will explore the concept of "money" and provide a few mental models to think about our current monetary system and some of the alternatives for storing wealth. In particular, I will cover the following:

1. Money 101 (What is money?)

2. The First Hard Money (Why gold?)

3. Does money supply even matter?

4. Can US Dollar remain the reserve currency?

Money 101

In simplest terms, money is just a medium of exchange that provides us with an efficient way to transact with one another. As an alternative to the barter system (e.g. exchanging 12 apples for 8 bananas), money provides us with a common unit to pay for goods and services.

It is generally accepted that for something to be a reliable medium of exchange, it must possess salability - the ease with which it can be sold in the market with the least erosion in its price. Money's relative salability can be assessed in terms of:

• Salability across scale - how easy is it to divide into smaller units (for smaller transactions) or grouped into larger units (for larger transactions)?

• Salability across space - how easy is it to transport around / carry it along as the holder travels?

• Salability across time - how well can it hold its value into the future?

While the importance of the first two properties (scale and space) are fairly obvious, salability across time is worth diving into a bit deeper.

Generally speaking, for a good to function as a robust store of value over time it should possess two properties:

1. It must be very hard for producers to significantly increase its existing supply in response to its price appreciation (the harder to increase the supply, the "harder" the money is said to be).

2. It must be resilient to erosion / damage / destruction.

Therefore, when evaluating the relative "hardness" of money, it is helpful to look at its "stock-to-flow" ratio, which consists of the following two components:

• Stock - existing circulating supply of a good less any quantity consumed or destroyed in the period

• Flow - expected new production in the next time period

This ratio should be intuitive - the higher the existing supply of a particular good relative to the expected growth in its supply, the less impact this new supply will have on its price and vice versa.

Let's look at the following abstract example:

As can be seen, Good A has a much lower stock-to-flow ratio, which means that its supply can be dramatically increased year-over-year, which can result with oversupply (and therefore substantial reduction in its value). Meanwhile, Good B’s supply growth is small relative to existing inventory, which will have little to no impact on its price.

The First Hard Money

Although throughout human history money has existed in numerous forms - cattle, seashells, beads, copper, silver, and gold to name a few - the clear winner has been gold, which is still widely held by most central banks and federal reserves.

To illustrate why gold became the dominant store of value, let's compare copper and gold as stores of value. While both have been used as stores of value in the past, there are several critical distinctions between the two:

• Copper:

  • Very abundant on our planet - there are roughly 5.8 trillion pounds of known copper resource on Earth of which only ~12% has been mined throughout history.

  • Has by far the highest recycling rate of any other metal (nearly as much copper is recovered from recycled material as is derived from newly mined ore).

  • Has many industrial applications - virtually all new supply is consumed, not stockpiled. This results in very low circulating supply of copper relative to annual production (e.g. low "stock-to-flow" ratio), which means that whenever copper prices rise due to excess demand, producers can easily flood the market with new supply (ultimately bringing the price back down), making it a poor store of value, with its price highly correlated to the economic cycles.

• Gold:

  • Very scarce - it cannot be synthesized from other materials and can only be obtained from unrefined ore which is extremely rare on this planet.

  • Has very limited industrial applications and is nearly indestructible, which results in virtually all of new gold production being stockpiled, which steadily grows circulating supply.

  • Given that gold exploration has existed for millennia, existing gold stockpiles are massive relative to annual gold production in any given year, resulting in very trivial growth of existing supply (over the past 70 years gold stockpile growth has averaged ~1.5% per year, never exceeding 2%) . This means that even in times of steep price appreciation (due to higher demand for gold as store of value), producers cannot materially influence the price of gold even if they double their production.

To illustrate this point, I include stock-to-flow ratios for some of the most common commodities.

For these reasons, gold has been historically an excellent store of wealth across time. However, in its raw form it was very inconvenient to use - it’s heavy to carry around and hard to divide into smaller units. To enhance gold's salability across space and scale, gold was minted into coins by most civilizations, invigorating a global gold trade. However, given lack of standardization of coins across the world (some coins contained 4 grams of gold, while others contained 9), its salability across space was still suboptimal, limiting the scope of trade between countries.

The Rise and Fall of The Gold Standard

To address this issue, some nations began to switch to paper money backed by gold held in vaults. Each nation's currency was backed by different quantities of physical gold, making conversion as simple as switching between different measuring units (e.g. converting miles to kilometers or pounds to kilograms). For example, the British pound was defined as 7.3 grams of gold while the Deutschmark was backed by 0.36 grams, implying an exchange rate of 20.3 Deutschmarks per 1 British pound.

Britain was the first of many nations to adopt a modern gold standard in 1717, making gold more marketable and incentivizing other nations to join as well. By 1900 ~50 countries were officially on the gold standard. This was the closest the world has ever come to having one global currency, generating record capital accumulation, global trade and rise in global living standards.

This de facto global gold standard was interrupted by the outbreak of World War I in Europe. The tremendous increase in military and defense spending put significant pressure on nations' finances, which were constrained by the amount of gold they owned. As funds began to dry up, governments were left with very few options to choose from to get more money:

1. Increase taxes to increase government revenue - this was politically and logistically impractical to do at the time of war

2. Increase national debt - this would be a very costly source of funding given excess demand for loans and its limited supply, resulting in unsustainably high debt service costs moving forward

3. Inflate the money supply (print money) to pay off expenses - this would require the abandonment of the gold standard

Most nations ended up employing some combination of Options 2 & 3, which resulted in an explosion in their national debt/gdp ratios, as well as a rapid increase in their money supply. Printing money proved to be particularly addicting since it was the most politically palatable - it enabled politicians to increase spending without having to re-allocate funds from other government programs or increasing taxation. Indeed, moving away from the gold standard enabled governments to indulge in spending that they could not previously afford - they were no longer constrained by government revenues they collected through taxation.

Even to this day, long after World War I and II, governments continue to inflate its money supply to meet its ever so expanding spending programs (welfare, healthcare, military, infrastructure, protectionist subsidies for unprofitable domestic industries, etc). Between 1965 and 2015, the global annual money supply growth has averaged ~35%, while the top 10 biggest currencies have grown its money supply at a more modest 7-10% per year.

While 7-10% may not sound like a lot (especially when compared to the global average), such growth rate would still result in the doubling of the money supply every 7-10 years, effectively diluting the incumbent currency holders by 50%! While global reserve currencies (i.e. USD today) may get away with it for longer, consistent increases in money supply will always result in currency devaluation.

Does Money Supply Even Matter?

A helpful mental model to understand the importance of money supply growth is to think of hypothetical company that issues 10% of its market cap in new equity every year. Would you be ok with that as a shareholder? Well, it depends. Let's look at three hypothetical examples:

As can be seen in the tables above, in all three cases, an annual 10% issuance of market cap results in ~50% dilution of shareholder's ownership interest in the company. However, what really impacts the net worth of the shareholder is the relationship between annual growth of the company and the new equity issuance:

• Company A's growth offsets the dilution, resulting in a ~37% increase in shareholder wealth by year 7

• Company B's growth is equal to the dilution, resulting in no change in shareholder wealth

• Company C's growth is less than the dilution, resulting in a ~28% reduction in shareholder wealth by year 7

Similarly, when evaluating the impact of any given country's money supply growth on purchasing power, one must consider the real growth of the economy that may or may not offset the inflation of the money supply. If the newly created money is used productively (e.g. infrastructure, healthcare, education, etc.), it will be highly accretive to that country’s economy and generate real GDP growth. However, if the new money is not used productively (e.g. spent on non-productive financial assets), it will result in the gradual devaluation of that currency (and result in the increase in the value of those assets).

With that in mind, no developed country has achieved average economic growth rates of 7-10% over this 50 year period (for instance, US has averaged only 3.2% since 1948), which results in significant dilution of the purchasing power in each of those countries. Of note, this analysis excludes the impact of COVID-19 pandemic, which resulted in a ~46% increase in money supply in 2020-21 alone (from ~$15 trillion to $22 trillion), sparking wide long-term inflation concerns.

Inflating the money supply (and thereby reducing the currency's purchasing power) to fund government spending that it cannot afford can be viewed as government's expropriation of wealth from its people. To make matters worse, easy monetary policy disproportionately impacts lower income people, who are more likely to not own any inflation-resistant capital (e.g. real estate, stocks). Instead, they largely rely on their nominal wages to fund their day-to-day life, only to see their purchasing power evaporate over the years.

Can Government Effectively Control Money?

To explore this question, let's consider a hypothetical coffee plantation in Brazil responsible for 30% of the world's coffee bean production. Let us assume that it unexpectedly gets struck by a drought, which destroys the entire crop for the year, resulting in a significant global shortage of coffee beans.

Scenario 1

1. Global shortage results in a sharp increase in the global price of coffee

2. Hundreds of thousands of individual coffee purchasers (coffee shops, restaurants, grocery stores) adjust their purchasing orders down based on their detailed analysis of expected customer demand in response to the price increase, resulting in reduced global demand for coffee

3. Thousands of coffee producers ramp up coffee production based on their detailed analysis of unmet customer demand, resulting in an increase in global supply of coffee

4. A combination of reduced demand and increased supply results in the stabilization of the global coffee prices, resulting in a new efficient market price at close to pre-drought levels

Scenario 2

1. To prevent significant market volatility, each country's government decides to set the new coffee price based on their top down analysis of the market supply and demand dynamics

2. New coffee price no longer reflects the collective buying and selling decisions of coffee market participants, distorting economic reality of the market

3. Individual coffee purchasers end up buying too much or too little coffee, while individual coffee producers end up producing too much or too little coffee

4. Since the market is unable to clear at a new efficient price (given government's centrally planned price), coffee market endures significant economic deadweight loss for both coffee producers and consumers

In this hyperbolic example, Scenario 2 would result in a complete disaster - no government is able to accurately forecast the correct efficient price (or production level) of a good on behalf of thousands and thousands of market participants. In a functioning economy, the price is the output based on collective actions of market participants that act in their own self interest. To make matters worse, markets are inherently highly dynamic and are constantly impacted by the constantly changing buyer and seller preferences, making it that much harder for any entity (let alone a government) to accurately calculate it.

So just as it is impossible for a government to centrally plan the coffee market, it is equally impossible to centrally plan the market for capital - whether it be its price (interest rate) or liquidity (money supply). As such, a centrally planned loose monetary policy results in suboptimal capital allocations throughout the economy. For example, an artificially low interest rate incentivizes reduced saving and increased spending, which results in reduced capital accumulation. At the same time, it encourages increased borrowing and enables unprofitable projects getting funded.

Government's control of its money supply has permanently increased the role the government plays in our day to day lives. During the international gold standard, foreign exchange rates were relatively fixed and the purchasing power of the gold-backed money was stable. People’s economic well-being was largely a product of their own economic input. In the era of the fiat money, our wealth now depends on the centrally planned fiscal and monetary policy (which directly affects our purchasing power) and relative currency fluctuations (which are affected by trade policies set by the government).

Can US Dollar Remain the Reserve Currency?

So what will happen to the US Dollar in the future? Is fiat money here to stay?

There are several factors that keep government money reigns supreme in our society:

• Banks can only open accounts in government-sanctioned money

• Legal tender laws make it illegal in many countries to use other forms of money for payment

• Taxes have to be paid in fiat

Taxation is by far the biggest competitive advantage that government money has over other alternatives - it’s what gives it its stability. Government can flex taxes (to raise its revenues) to buy and stabilize its own currency. Additionally, governments (and banks) needs to be able to track money flows between market participants to prevent tax evasion, money laundering, terrorist financing, etc. In short, I do not expect fiat money to go away any time soon.

With that said, I do not expect fiat currencies to function as robust stores of wealth over the long-term - even the US Dollar. Why? Simple math.

At the time of writing, the United States has ~$30 trillion of debt, a lot of which has been raised in a historically low-interest rate environment (1.4% blended cost of debt). As we are heading into a high-inflation environment (latest CPI has clocked in at 8.6%), the Fed will be forced to start raising rates to stifle inflation. For every 100bps increase in rates, government’s debt service will increase by ~$300 billion. At the same time, the government is already running a $1.8 trillion budget deficit!

Rising debt service costs will only put additional pressure on the budget deficit. I am not even mentioning all of the underfunded state pensions and social security - all of which will likely have to get bailed out by the government in the future.

So what does this all mean?

The US government has very few options left to meet its future obligations - it WILL have to keep printing money to avoid bankruptcy (raising taxes or cutting spending will be hard both politically and economically). This means that the US Dollar will continue to get devalued over time until a new reserve currency with more sound monetary policy will come to dominate global trade.

To be clear - I do not expect fiat currencies to disappear - nor do I predict the US Dollar to hyperinflate in the next few years - but at different times throughout history we’ve had many reserve currencies (e.g. Dutch Gilder, British Pound, etc.) and they all had the same fate - they were all devalued by overindebted governments to fund its liabilities, giving way to a new reserve currency to emerge.

This is exactly why gold has been such a strong store of wealth over millennia - it has been the only money with a sound ‘monetary policy’ - that is, until Bitcoin was invented, which shares most of the same properties (indestructible, immutable inflationary profile, scarce finite resource).

For Bitcoin to displace gold the world will have to buy into its similarities to and advantages over gold at scale - only time will tell if this experiment succeeds. But one thing is certain - both gold and its digital equivalent will never hyperinflate, nor will its supply be ever controlled by any sovereign state.

We are all in for quite a ride over the coming decades. Stay safe.

Disclaimer: None of this is financial advice. Do your own research. The author may hold long positions in gold and Bitcoin at the time of writing.

Join our growing community by subscribing here:

Oh boy. A lot has changed since I first wrote about STEPN in February 2022:

• Monthly Active Users grew to 2+ million (!) from just 60,000 users

• Daily Active Users grew to 500,000+ (!) from just 15,000 users

• GMT became a top 50 coin by market cap, flipping Axie Infinity

• GST almost 3x’d going from $2.25 to $6.00+

• NFT Sneaker floor almost 2x’d from 6.5 SOL to 12+ SOL

Most importantly, STEPN was able to deliver on its mission and to successfully onboard hundreds of thousands of people onto Web3 while incentivizing them to live a healthier and more active lifestyle.

However, as the game grew in popularity I have been drowning in private messages from people asking me about gameplay mechanics and optimal (read - profit maximizing) strategies. With that, I decided to write a quick guide on how to think about the various ways you can play STEPN. This essay will dive into both the qualitative and quantitative analysis, and I will even provide my own model for readers to play around with their own assumptions to see impact on expected returns.

In particular, this essay will cover the following:

1. What Sneaker(s) should you get? (type, rarity, quantity)

2. What strategy should you pursue? (GST, GMT, Breeding, Mystery Boxes)

3. Nuances (a few caveats / details to consider)

For those that are short on time, please see the below TL;DR recommendations:

• Sneaker type: Jogger (flexible, good value)

• Sneaker rarity: Uncommon (highest yield, quickest payback)

• Sneaker quantity: 3 or 9 (25 or 50 minutes of daily earn time, respectively)

• Game strategy: optimize for GST minting (max efficiency)

Let’s dive in!

Note #1: This article assumes you have basic familiarity with how STEPN works and will primarily focus on profit maximizing strategies. If you are relatively new to the app, I suggest reading my original deep dive first.

Note #2: The game mechanics are subject to change - the app is still in its Beta and management team is constantly monitoring metrics and making tweaks as appropriate to ensure the game’s long term sustainability.

Note #3: This is NOT financial advice - just my opinions. Do your own research. Cryptocurrencies can be very volatile and result in permanent loss of capital.

What Sneaker(s) should you get?

There are four key decisions each new STEPN player has to make:

1. Sneaker type (Walker, Jogger, Runner, Trainer)

2. Sneaker rarity (Common, Uncommon, Rare)

3. Sneaker quantity

4. Sneaker points

Let me quickly explain the tradeoffs of each so you can decide what’s right for you.

1. Sneaker type (Walker, Jogger, Runner, Trainer)

Each shoe type has a specific speed band within which the player must stay in order to earn GST rewards. Below I rank the shoes from most to least flexible.

Conclusion. Most people should get either a Jogger or a Trainer (if comfortable with higher upfront cost) as they are the most flexible. Walkers and Runners are cheaper but are only good for 1 use case, which may require you to get a separate Sneaker to complement.

2. Sneaker rarity (Common, Uncommon, Rare, Epic)

So should you get a Common or Uncommon or your main Sneaker? Or should you go all out and get a Rare? As always, the answer is… it depends.

As a recap, higher Rarity affects earning potential in three ways:

1. Higher base attributes at Level 0 (which is important for boost from Gems)

2. More points to distribute to attributes per each level up

3. Bonus energy units (more earn time)

From a purely ROI perspective, Uncommon sneakers generate the highest yield (and therefore quicker payback period) and generate higher earnings in absolute terms (compared to Commons). On the other hand, Rare sneakers imply the lowest yields (and longest payback period) at current prices, but can earn considerably more income per day when paired with Level 2 or Level 3 Gems (which provide a 70-220% boost to base attributes). Of course, your risk appetite / available liquidity will also be a major consideration when deciding how much capital to have at risk.

Conclusion. For most people, an Uncommon or Common Sneaker is the way to go.

3. Sneaker quantity

As a reminder, the number of NFT Sneakers you own determines the amount of “Energy” you have, which in turn determines the amount of your daily move2earn time (1 Energy = 5 minutes). Crucially, your Energy levels do not increase with every additional Sneaker you own - instead, there is a scale with specific “break points” at which your Energy will increase:

Therefore, unless you are buying/minting Sneakers for potential capital appreciation (i.e. selling them at a higher price in the future), for most users it makes sense to own 1, 3, or 9 Sneakers since any additional Sneakers in your portfolio will not earn any additional income until the next “break point” is met.

Similarly to the Sneaker Rarity, the primary constraint here will be a user’s liquidity and risk appetite since each Sneaker costs at a minimum ~$1,200. With that said, no matter what type of Sneaker, portfolios of 3 Sneakers or higher will always result in higher ROI and quicker payback period (despite higher upfront investment).

Finally, one more thing to consider here is your schedule. Before going with a 9 or 15 Sneaker portfolio, ask yourself if you can commit to daily walking/running of 45-60 minutes - if the answer is no, it might make sense to go with less Sneakers since you wouldn’t be maximizing your ROI.

Conclusion. Get 9 or 3 Sneakers for best results (if comfortable with upfront cost).

4. Sneaker Attributes

Last thing to cover is point allocation at each level up. The optimal allocation will depend on the game strategy of your choice (we will cover this more in the next section). But as a rule of thumb:

• For earning GST → maximize Efficiency

• For Mystery Boxes → maximize Luck

• For earning GMT (not available yet) → maximize Comfort

Please note that you will also need to balance this with some allocation to “Resilience” to reduce your Sneaker’s repair costs (especially relevant if you are exercising for 45+ minutes a day). You can play around with points allocation in this helpful calculator which can estimate your revenue, repair costs and profit.

Summary

Before we get into the more detailed math and some more nuanced strategies, let’s quickly recap the recommendations for an average user (again, these can vary based on your circumstances, DYOR):

• Primary Sneaker Type - Jogger

• Primary Sneaker Rarity - Uncommon, if too expensive get a Common

• Shoe quantity - 9 or 3

Now let’s dive into how you can actually make money playing STEPN.

What Game Strategy Should You Pursue?

What makes STEPN fun is that there are multiple ways to play the game and earn rewards. At present, the primary strategies users can pursue in the game are:

1. Mint GST token (best risk/reward)

2. Earn Mystery Boxes (higher risk, higher reward?)

3. Breed Sneakers (lowest risk, lowest reward)

4. Mint GMT (not yet available)

But which strategy is the best one?

Mint GST (Best Risk/Reward)

This strategy is the most well understood and at present offers the best risk/reward. To truly maximize your daily yield you will need to own multiple Sneakers (9 or 3 preferred), have one primary high earning Sneaker (preferably an Uncommon), and exercise daily. The primary risk players assume is the price of GST token itself, which in turn will affect the price of the Sneakers purchased (if GST drops, so will the Sneaker price). As such, upfront capital investment should be balanced with your risk appetite and expected payback period at various GST prices.

Below I provide an illustrative returns analysis for 3 and 9 Sneaker portfolios with a Level 28 Jogger as a primary Sneaker. Given that prices/game mechanics can change any time, please refer to this model see impact on ROI under different assumptions.

A few quick observations from the above maths:

• If you are going with a Common jogger as your primary Sneaker, you are better off buying a Level 28 Sneaker (vs gradually leveling up from level 0)

• If you are going with Uncommon jogger, it is more economical to level up from zero (though it will take ~18 days to reach max level and your earnings will not be maximized until then)

• Rare Sneakers currently yield the lowest (given high upfront cost) but will earn by far the greatest daily income in absolute terms

• Upgrading Sneaker to Level 28 from Level 0 will take ~18 days, which means that your Sneaker will not be earning at its max potential until then

Earn Mystery Boxes (Higher Risk, Higher Reward?)

Users can earn Mystery Boxes containing Gems from their daily workouts which can be used to boost Sneaker attributes or sold on the in-app marketplace. The frequency and rarity of the Mystery Box depends on duration of the exercise (and therefore # of Energy spent) and on the Sneaker’s Luck attributes. While exact probabilities are not known, the STEPN community was able to crowd-source unofficial data to get a better idea of what is required to earn various Mystery Boxes:

A few quick observations:

• To earn a Common Mystery Box you will typically need to spend at least 5 Energy (25 mins) per run and have more than 10+ Luck attribute

• To earn an Uncommon Mystery Box you will typically need to spend 10 Energy (50 mins) per run and likely have 20+ Luck attribute

• To earn a Rare Mystery Box you will likely need to spend 15 Energy (75 mins!) per run and have 30+ Luck attribute

This make Mystery Box earning a fairly advanced strategy that requires significant upfront investment (for most people this would mean 9+ Sneakers with a dedicated Sneaker with maxed out Luck). But what does this mean economically?

Firstly, not all gems are created equal. Higher Level Gems are significantly more valuable than Level 1 Gems, with Level 4 Gems priced in the tens of thousands:

However, earning an Level 2 or Level 3 Mystery Box DOES NOT guarantee that you’ll receive a gem of that same level (i.e. a Level 3 Mystery Box might contain multiple Level 1 or 2 Gems instead of a Level 3 gem. Luckily, the STEPN community was able to crowdsource the expected profit from each type of Mystery Box based on thousands of data points:

This means that consistently harvesting Uncommon or Rare Mystery Boxes can still be quite profitable, adding thousands of dollars in potential weekly income in addition to regular GST rewards (which would be lower for Sneakers with lower Efficiency and higher Luck). Given that each box takes about 1 week to unlock, a user can expect to earn up to 4 boxes per week, generating $600 to $2,400 on average, with attractive upside from receiving an Uncommon or Rare Gem.

Conclusion. If pursuing a Mystery Box strategy, go with at least 9 shoes and 1 Uncommon (resulting in 10 energy) with a 30+ Luck attribute.

Breed Sneakers (Lowest Risk / Lowest Reward)

To play STEPN each new user needs to own at least one Sneaker. However, given that there is a finite number of Sneakers in circulation, existing users are economically incentivized to breed (create) new Sneakers to onboard new users.

This is arguably the lowest risk (most quantifiable) and therefore lowest reward strategy in the entire game. However, it is an essential component of the game as it directly impacts new user growth, so it is worth exploring in more detail.

As a reminder, each Sneaker can be used for breeding a maximum of 7 times, with each consecutive mint requiring increasing quantities of GST and GMT:

As long as the price of a Sneaker is greater than the cost to mint, users are incentivized to mint new Sneakers and earn a risk-free profit:

A few quick notes on the above math:

• Mint prices can change from time to time - management is constantly monitoring supply/demand dynamics and tweaking economics of shoe minting as appropriate

• At current price it is profitable to mint Common Sneakers and unprofitable (on average) to mint Uncommon Sneakers. This is because when you open an Uncommon shoebox there is a 25% chance you will get a Common Sneaker, resulting in a steep price discount of Uncommon Shoeboxes relative to the Uncommon Sneakers.

• Minting process can be time consuming. Each Sneaker must be upgraded to Level 5 first which will take a total of 30+ hours. Following each mint there is also a “cool down” period during which the Sneakers are not available for secondary mint.

Conclusion. This is a strategy that can be pursued in parallel to GST minting once your primary Sneaker is maxed out (since you cannot level up two Sneakers at once). Otherwise, for conservative players, this strategy can provide an attractive 6-7% “risk free” return on capital every few days.

Mint GMT (not yet available)

The final strategy to earn a return in STEPN is to mint GMT. This feature is not yet available in the game, but is expected to be added in H1/2022. We don’t know the exact game mechanics for this other than the Sneaker’s GMT earning potential will depend on the Comfort attribute and on the total number of users minting GMT (i.e. the more users earn GMT at a given time, the less GMT earning will be available per user and vice versa).

Given lack of any specifics on how this will work exactly, there are separate approaches to playing this strategy:

1. Do nothing - when details come out consider ‘resetting’ your primary Sneaker into a GMT earner (by maxing out Comfort attribute)

2. Level up a new Sneaker on the side without attributing any points until GMT earn mechanics are clarified (then optimize appropriately or sell Sneaker)

3. Sweep Comfort Gem floors in anticipation of price appreciation ahead of GMT earning functionality

Nuances

STEPN is an incredibly complex game with many moving parts and variables. Make no mistake - this is by design - obfuscating ROI with complexity makes the game:

1. More fun to play and explore, and

2. Harder to exploit / rent-seek

3. Provides management with more levers to control / protect the economy

With that said, there are a few nuances worth mentioning / flagging as you begin your journey within the STEPNverse.

Gems

We briefly touched on these earlier on but these are worth mentioning once more - higher level Gems (i.e. Uncommon, Rare and Epic) can substantially increase the earning power of your Uncommon or Rare Sneaker. For reference, a Level 3 Efficiency Gem applied to a Level 0 Rare Sneaker with 25 base Efficiency (middle of the range) would result in Efficiency of a 100 (!) - this is what a fully maxed Common Sneaker would have at Level 30!

Conclusion - don’t sleep on Gems, especially if you have a higher rarity Sneaker.

Multi-Purpose Sneakers

In the preceding section we talked about different strategies you can employ in the game and the different Sneakers you may need to pursue them. In reality, many users choose to have a “multi-purpose” Sneaker - one that is “pretty good” at multiple things (versus optimizing for just one). This is particularly practical to do with Uncommon and Rare Sneakers given the higher available points to distribute across each attribute (for example you can have Luck at ~30-50 points to increase chances of earning Mystery Boxes and then the rest allocated to Efficiency for GST earning).

Conclusion - consider owning a more universal Sneaker that be used for multiple strategies instead of a Sneaker optimized for just one strategy

Renting

The high upfront investment required to play the game is no doubt a limiting factor for widespread adoption - not everyone can fork out $1,200 for a virtual NFT Sneaker. To reduce onboarding friction and provide an additional revenue stream for existing users, management plans to introduce “renting” functionality, which would enable new users to “lease” Sneakers from existing users in exchange for a revenue share. While this will likely result in tremendous new user growth (due to lower barriers to playing the game), this could have unexpected consequences on the prices of GST and NFT Sneakers themselves:

• Increasing utilization rate of otherwise “idle” Sneakers should reduce demand for buying new Sneakers resulting in 1) lower price of Sneakers 2) lower minting demand → less GST burn/deflation

• Renters introduce the risk of being parasitic/extractive to the ecosystem as they convert their daily GST rewards into USDC, providing additional sell pressure

Conclusion - we don’t know yet how renting will work, but it is something that may introduce fluctuations in the price of Sneakers and GST in the future

Liquidity

One of the nice things about blockchain-enabled games is that you have full control over your digital assets which you can sell to other users on the open market. While this is certainly the case for STEPN as well (it’s in-app marketplace is fantastic), it is worth noting that liquidity varies greatly across different items sold on the platform. Unsurprisingly, the more expensive the Sneaker or Gem is, the less buyer demand there will be (less users can afford a high price tag), resulting in diminished liquidity.

Conclusion - keep in mind you might not be able to easily dispose of your higher priced items due to a smaller buyer universe (relevant for fully maxed Sneakers, ultra rare Sneakers/Gems, etc).

Conclusion

Since its launch in late December 2021, STEPN took the crypto world by storm - and rightfully so. It is one of the most innovative, daring and impactful web3 projects of the year and it has pushed the boundaries of what’s possible in the play-to-earn.

The game is being developed real-time, and if one thing is for certain is that the founders will continue to innovate, iterate and improve the gameplay as the app continues its meteoric rise.

I am excited to continue following STEPN on its journey of onboarding millions more people onto web3 and plan to keep STEPN daily to get my daily exercise in ;)

Disclaimer: The author has long positions in GST, GMT and NFT Sneakers mentioned in the essay. This is NOT financial advice. DYOR.

Join our growing community by subscribing here:

If you have used crypto, it is very likely that you have come across stablecoins - you know, those tokens that are always priced at $1.00? Perhaps you swapped dollars for USDC on Coinbase to send money to one of your wallets. Or maybe you deposited money into your Gemini account and converted it into GUSD to earn an 8.5% yield on your savings. Or maybe you paid for your online order with UST using the Chai app. Or did you receive a payment in USDT from overseas for that consulting project?

You might ask, why talk about stablecoins? Aren’t they… kinda boring?

No! Stablecoins have by far one of the largest TAMs in crypto (the $130 trillion global money supply itself). Indeed, stablecoins have already become ubiquitous within crypto. By the end of 2021, the combined stablecoin market cap reached an astonishing $180 billion (and still growing!). This should makes sense - as our lives (and economic interactions) become increasingly digital, we need the financial infrastructure to provide a stable medium of exchange to execute frictionless peer-to-peer transactions. Herein lies an opportunity - successful scalable stablecoin projects have the potential to generate massive upside for investors.

One such project is UXD Protocol - a new stablecoin on the Solana network that is backed by a stable of industry veterans such as Multicoin Capital and Alameda Research. UXD aims to eliminate the tradeoffs that exist in incumbent stablecoin solutions today by offering a fully decentralized, capital efficient, and stable asset pegged to USD.

UXD Protocol caught my attention due to its incredibly elegant business model and clean tokenomics design. The more I dug into the project and its roadmap, the more impressed I became. I believe UXD has the potential to become a major player in the stablecoin arena, offering early investors an asymmetric risk/reward profile given the project’s early stage of development and user adoption.

In this essay I will cover the following topics:

1. Stablecoin 101 (what are they and why do we need them?)

2. UXD Business Model (how does UXD work and make money?)

3. Tokenomics & Valuation (is UXD an attractive investment?)

4. Conclusion (what’s the verdict?)

Let’s dive in!

Stablecoin 101

Stablecoins are blockchain-based versions of fiat currencies (such as USD or EUR) that offer price stability while still providing key crypto-native properties such as cheap and fast transaction settlement, 24/7/365 trading, privacy, and security.

Given that stablecoins are typically pegged to a fiat currency (e.g. 1 USD stablecoin = $1.00), they do not offer any price appreciation potential (unlike say Ethereum or Solana). Instead, they are primarily used as a price-stable unit of exchange within the crypto ecosystem, mitigating the price volatility seen in other crypto assets (which can move up or down in price 10%+ in a single day) . Some of its primary use cases include:

• Peer-to-peer money transfers (especially cross border)

• Purchases of goods and services (provides instant transaction settlement)

• Buying/selling cryptocurrencies on trading platforms that do not offer a fiat on/off ramp

• Borrowing / lending money via decentralized finance (DeFi) applications

• Yield farming (providing liquidity on decentralized exchanges)

Stablecoins have seen a massive rise in popularity over the last year, following a cambrian explosion of app development within DeFi and the overall increase in peer-to-peer transaction activity on blockchain networks. As of this writing, the total market cap of all stablecoins is nearly $180 billion.

As can be seen from the chart above, stablecoins come in many different flavors, each employing different mechanisms to maintain the peg to the fiat currency. Let’s explore the three dominant stablecoin types and their respective drawbacks / risks.

Note: If you are already familiar with stablecoins, feel free to skip to the “UXD Business Model” section.

1. Fiat-backed stablecoins (e.g. USDT, USDC, BUSD). These stablecoins are backed 1:1 with off-chain USD reserves in traditional bank accounts. These reserves are custodied with centralized providers (e.g. Tether in the case of USDT or Binance in the case of BUSD). To date, this has been the biggest and most popular type of stablecoin by market cap. The biggest drawback of this solution is centralization, which results in three primary risks:

   • Trust - the user must trust that the custodian of the off-chain deposits actually has enough reserves to back every on-chain stablecoin, which typically requires an audit done on a regular basis by a 3rd party audit firm. The potential danger of such structure came to light when Tether found itself in hot waters when an allegation surfaced that it used $850m of Tether funds to cover a loss at the parent company.

   • Regulatory - Lack of clear regulatory framework around custody of reserves, as well as potential privacy / reporting concerns

   • Censorship - Risk of asset freezes / blacklisting by the custodian (Tether recently froze $160m worth of assets upon request of law authorities)

2. Crypto-backed stablecoins (e.g. DAI). To address the issue of centralization crypto-backed stablecoins rely on trustless/decentralized crypto assets such as Ethereum. The biggest issue with this design is that it is capital inefficient - given the significant price volatility of the underlying reserves crypto-backed stablecoins require overcollateralization - this means that every $1.00 of stablecoin has to be backed by 1.5-2.0x of collateral capital. Additionally, to grow the reserve base protocols such as MakerDAO (providers of DAI) began accepting non-decentralized assets such as USDC - injecting the very problem of centralization these protocols were meant to address in the first place!

3. Algorithmic stablecoins (e.g. UST, FRAX, FEI). Algorithmic stablecoins were created to address the shortcoming of the fiat-backed and crypto-backed stables. They are completely on-chain (i.e. trustless and decentralized), and are backed by no collateral (i.e. capital efficient). Instead, to maintain its peg to USD these coins rely on some sort of algorithmic rebalancing of supply and demand of the stablecoin. In the case of Terra’s UST (biggest algo stablecoin at the time of writing), the peg is achieved through an algorithmic incentive mechanism. At any time, users on Terra can burn $1 of LUNA to mint 1 UST, or burn 1 UST to redeem $1 worth of LUNA. Here’s an example:

   • If UST price increases to $1.05, traders can burn $1.00 of LUNA (Terra’s native token) in exchange for minting 1 UST, generating a $0.05 risk-free profit. This would reduce LUNA supply and increase UST supply.

   • If UST price drops to $0.95, traders can burn 1 UST in exchange for minting $1.00 of LUNA, once again, generating a risk-free profit of $0.05. This would reduce UST supply and increase LUNA supply.

   While UST’s market cap has ballooned to $14 trillion, much of its popularity has been driven by an incredibly high 19.5% yield it offers its holders via Terra’s DeFi app Anchor. To maintain such a high APY, Terra's parent Terraform Labs recently had to replenish the yield reserve. If the yield reserves do ever run out, the APY on UST will likely be <10%. If this occurs, the demand for UST might significantly diminish, causing investors to migrate funds to other protocols, thereby drying up UST liquidity and introducing de-pegging risk, which has raised concerns around the long term stability of the UST peg.

Stablecoin Trilemma

To summarize, for a stablecoin to be an effective proxy for digital cash, it must possess the following three properties:

• Decentralized & trustless - fiat-backed stables such as USDC and USDT are capital-efficient and relatively stable, but are governed by centralized entities that store its reserves in off-chain traditional bank accounts, introducing a single point of failure risk (trust, censorship, fraud)

• Capital efficient - crypto-backed stables such as DAI require 1.5-2.0x as much capital as the amount of stablecoin it produces making it highly inefficient

• Price Stable - it must maintain a tight peg to the underlying fiat currency even in times of significant market volatility. Algorithmic stables such as UST, FRAX or FEI all rely on multi-token rebalancing mechanisms that introduce a lot of reflexivity and a risk of a potential de-pegging.

Until recently, no existing stablecoin possessed all three of those properties.

Enter UXD.

UXD Business Model

UXD is a new stablecoin on the Solana network that aims to maintain a perfect peg to the USD. The UXD Protocol operates as follows:

1. User deposits $100 of crypto collateral (e.g. ETH, BTC, or SOL) and receives 100 newly minted UXD tokens (1 UXD = $1.00 USD)

2. The protocol immediately shorts the equivalent amount of collateral to achieve a perfectly hedged position. This means that if ETH drops/increases in price by 10%, the short position would fully offset that price change, resulting in no change in the value of the underlying collateral base.

3. UXD relies on arbitrageurs to maintain a perfect peg to the USD. For example, if UXD trades at $1.05, users can deposit $1.00 of collateral to receive UXD. If UXD trades at $0.95, users can redeem each UXD for $1.00 of collateral. This means that whenever the price of UXD deviates from $1.00 traders will be able earn a risk-free profit by arbitraging the token back to the peg.

UXD’s elegant design is intended to address the shortcomings of existing stablecoins:

• It is fully trustless and decentralized and is backed by on-chain deposits

• It is capital efficient - each UXD is backed 1:1 by its reserves

• It is price stable - UXD is insulated from market volatility given its perfectly hedged collateral nor does it depend on reflexive multi-token incentive mechanisms

However, to fully grasp UXD’s business model we need to add a bit more nuance around how the protocol actually works and how it generates revenue.

A Quick Detour - Perpetual Swaps 101

First, we must understand how exactly the protocol is able to “short” its collateral base to achieve a perfectly hedged market neutral exposure. UXD does so via perpetual swaps - a type of financial derivative. Without getting into too much technical detail, this instrument can be described as follows:

“Perpetual swap” contract is a way to gain exposure to an underlying asset without having to “settle” the asset at any point in time.

Each contract has two parties - a buyer (“long” position) and a seller (“short” position). Contract price may be higher or lower than the spot price of the underlying asset, which would require either the buyer or the seller to make a payment to the counterparty to compensate for the difference:

An important takeaway here is that in bullish markets the funding rate is typically positive - the “short” position receives a payment from the “long”, while in bearish markets the funding rate is typically negative - the “short” makes a payment to the “long”. This should be intuitive - the funding rate acts as an incentive to enter the other side of the trade: the more bullish the market → the higher the difference between contract and spot price → the higher the funding payment an investor receives for entering a short position. This is exactly why funding rates typically oscillate between positive and negative, depending on underlying market conditions.

Ok, let’s get back to UXD.

UXD Revenue Model

Ok quick recap.

• Every UXD token is backed by $1 of cryptocurrency collateral.

• The protocol eliminates all market price exposure by “shorting” its collateral reserves via perpetual swap contracts.

• Depending on market conditions, the protocol will either receive payments from its swap position (when funding rate is positive) or make payments (when funding rate is negative).

As demand for UXD stablecoin grows over time, so will the protocol’s deposit base, or “total value locked” (given that each UXD is backed 1:1), as well as its swap positions. Therefore, UXD’s gross revenue model can be distilled into the following formula:

Revenue = Total Value Locked x Funding Rate, where

Total Value Locked = UXD Market Cap

This simple revenue model invites two important questions:

1. How big can UXD deposits get? (i.e. what is the TAM?)

2. What would a long-term funding rate look like? (i.e. what are the unit economics?)

Let’s tackle TAM first.

A Multi-Trillion Dollar Market Opportunity

At its core, stablecoins attempt to provide a digital-native equivalent of cash. Critically, stablecoin use cases extend far beyond that of just crypto - in the future we will be using this digital infrastructure to facilitate frictionless peer-to-peer transactions, as well as to pay for goods and services. This means that the total TAM for stablecoins is the global money supply itself - about $130 trillion. As such, I expect the demand for stablecoins to grow exponentially over the next decade (B2B payments sector represents a $35 trillion opportunity alone).

UXD is committed to providing a trustless, decentralized solution backed by on-chain crypto collateral. Therefore, a bet on UXD Protocol is a bet on both growing UXD adoption as well as on the growth of the overall crypto market cap (currently at $2 trillion vs ~$475 trillion of all global assets).

Today, the combined stablecoin market cap sits at ~$180 billion, or 9% of total crypto market cap of ~$2 trillion. For illustrative purposes, I provide what total stablecoin market cap could look like for a given size of the overall crypto market and a stablecoin market share:

As can be seen from the above example, it is not unimaginable that the total stablecoin supply will be in the trillion of dollars, and while we can’t tell for certain how much UXD will be able to capture, the total opportunity is clearly massive.

Attractive, but Volatile Unit Economics

The other key value driver for UXD is the realized funding rate, which determines how much revenue the protocol can generate on each dollar of its deposit base.

As explained earlier, funding rate can be either positive or negative, depending on market conditions. Although the funding rate on BTC and ETH swaps has typically been positive (which means swap buyers pay swap sellers), it’s also been incredibly volatile. However, as crypto markets have significantly matured over the past few years (both in terms of market cap and liquidity), the funding rate volatility has reduced substantially.

Assuming crypto will continue to experience exponential growth for years to come, the funding rate is likely to remain largely positive somewhere in the 1-5% range (annualized). Applying these rates to a range of potential UXD market cap outcomes illustrates just how profitable the protocol can become in the future:

With that said, it is clear that there will be periods of time with negative funding rates, requiring UXD to make payments to the buyers of the swaps, which could result in prolonged cash outflows.

To mitigate this risk, UXD has set up an “insurance fund” (initially funded with $57 million from the UXP token sale) to create a buffer against sustained periods of negative funding rates. At the moment, 100% of protocol’s revenues are allocated to the insurance fund. In the future, UXP tokenholders will be able to vote on how the gross revenues generated by the protocol will be allocated between the insurance fund and profit distributions.

Now that we covered UXD’s business model and key value drivers, let’s bring it all together by looking at the protocol’s tokenomics & valuation.

Tokenomics & Valuation

UXP is the native governance token of the UXD Protocol. It has a fixed total supply of 10 billion tokens with a multi-year emission schedule. The UXP token serves three primary function:

1. Governance decisions

2. Liquidity provider of last resort (should insurance fund get depleted)

3. Profit distribution to staked UXP tokenholders (subject to voting)

UXD Protocol’s “profit” can be viewed as whatever is left over after funding the insurance fund and covering overhead. While actual profit distributions are subject to a governance vote, this analysis assumes an illustrative 50% profit share.

Overall, there are three primary ways to earn a return via UXD Protocol:

• provide liquidity for UXD tokens and earn an APY

• provide liquidity for UXP token and earn an APY

• stake UXP and earn protocol’s distributable profits (subject to vote)

So is investing in UXP attractive at current prices?

Valuation

Given how new and unique this protocol is, finding relevant comparables is next to impossible. Equally, given a very limited track record, forecasting the protocol’s future growth is just as tricky. With that in mind, and based on some of the TAM analysis presented above, I ended up evaluating three illustrative growth scenarios:

• Base Case = ~$140bn TVL by 2031, LT funding rate of 3.0%, 70% probability

• Downside Case = $6bn TVL by 2031, LT funding rate of 1.0%, 20% probability

• Upside Case = $410bn TVL by 2031, LT funding rate of 5.0%, 20% probability

Summary of the key operating assumptions for each scenario can be found below:

You can find my model here if you’d like to play around with your own numbers.

Using a Discounted Cash Flow (DCF) analysis and a target 25% IRR I derived a blended valuation of ~$0.38 (on a fully diluted basis), implying a ~6.6x upside from current price of $0.058. In the upside case, this analysis implies a ~21x return from current levels. Therefore, I see the risk/reward here skewed asymmetrically to the upside, and as such I view UXP as a very attractive investment at current levels.

Note: At the risk of sounding redundant, given how nascent this business is, this valuation framework is highly sensitive to assumptions on TVL growth as well as long term funding rates

Key Catalysts

On a more qualitative side, this is the perfect time for UXD to launch its product. More specifically, I see the following near to medium term catalysts driving the protocol’s growth and adoption:

1. Robust product roadmap for 2022 - UXD team looking to add UXP staking functionality, integrate cross-chain capability, as well as introduce other fiat currencies such as JPY, EURO, GBP. All of these updates should drive meaningful demand for both UXD and UXP tokens.

2. UXD’s native blockchain, Solana, is rapidly expanding with hundreds of new DeFi, NFT, and Gaming projects launching over the coming months. A Solana-native stablecoin will find an easy product-market-fit.

3. Given the global geopolitical tensions and associated recent asset seizures/freezes by governments around the world has brought back its focus on the notions of “self-sovereignty” and decentralization. Demand for censorship-resistant and trustless financial products will only keep getting bigger.

4. Recent spike in volatility is driving investors into price-stable “risk off” assets which should be bullish for near-term demand for stablecoins, and particularly those that are insulated from market fluctuations.

5. Long-term secular tailwind of growing broader crypto adoption across numerous verticals and industries should continue to drive peer-to-peer transaction growth and, in turn, increased demand for stablecoins.

Key Risks

While I genuinely believe in the protocol’s high growth potential, I would be remiss if I didn’t mention a few potential risks associated with this protocol and its future development:

1. Funding rate risk - funding rates have been quite volatile in the past and it is anyone’s guess where they will settle long term. Given that funding rates are the primary revenue driver for UXD, a low/near-zero funding rate will materially reduce the protocol’s profitability.

2. Liquidity constraints - UXD currently relies on on-chain perpetual swaps on Mango Markets (Solana’s native perp market). New UXD can only be minted if there is sufficient liquidity in the perp market to enter into offsetting short positions. Given relative immaturity of on-chain perp markets on Solana network (measured in the 10s of millions USD vs off-chain markets measured in the 10s of billions USD), lack of liquidity could potentially significantly constrain UXD’s supply growth. To mitigate this, UXD accepts multiple types of collateral (ETH, BTC, SOL to start), increasing total available liquidity. It is also worth noting that liquidity on Mango Markets has been steadily rising and is expected to continue to do so in tandem with continued growth of Solana ecosystem and broader crypto market.

3. Regulatory issues - given that UXD peg relies on hedging via perpetual swaps, UXD Protocol is currently not available in some key geographies such as United States due to regulatory issues. This lack of access could potentially slow down the adoption / supply growth of UXD. Management said they are working with counsel on potential workarounds. Of note, all users can still buy UXD or UXP tokens on the open market - it is only the minting/ redemption of UXD that is unavailable in certain geographies.

4. Insolvency risk - whenever funding rates are negative, the protocol has to pay out money to the “long” counterparty of the swap. If funding rates stay negative for a prolonged period of time there is a risk of insurance fund depletion, which would trigger the protocol to issue new UXP tokens (diluting existing holders) to fund the cash shortfall. Of note, UXD’s current insurance fund could sustain ~12 months of negative 10%+ funding rates at $500 million TVL. UXD published a detailed study on impact of changes in funding rate on protocol solvency.

5. Smart contract risk - this is true for any on-chain protocol, but is particularly true for newer, less mature projects. UXD smart contracts have been audited by two independent smart contract auditors and no major issues were identified.

Conclusion

UXD Protocol is a very ambitious and innovative product designed specifically to solve the so called stablecoin trilemma. I believe its censorship-resistant and capital efficient solution will find a natural product-market-fit within the burgeoning Solana ecosystem, driving significant demand for UXD.

This project is still relatively under the radar (only $21 million of TVL!) and I see asymmetric risk/reward at this stage of its development. I have been and will continue to invest at these levels.

With that said, it is very important to keep in mind a number of key risks involved such as potentially volatile unit economics (funding rate), as well as potential growth constraints (limited on-chain perp markets, not available in some countries such as the US due to regulatory issues).

As always, this is definitely NOT financial advice - this thing could moon or go to zero. Always DYOR and only invest what you can afford to lose.