OpenAI is an L1, Distillation is a Hard Fork: How AI Stole Web3’s Stack Wars

by @allquantor

Fat Protocols and Fat Models

If you spent 2016 reading venture capital blogs, and I hope, for your personal health, that you were doing literally anything else, you probably ran into Union Square Ventures’ Fat Protocol Thesis.

The theory went like this:

In the Web2 internet, protocols were "thin" and applications were "fat." Nobody got rich owning HTTP, TCP/IP, or SMTP. Instead, companies like Google, Meta, and Amazon built fat applications on top of those thin protocols and captured roughly 100% of the economic value in the observable universe.

Web3 was supposed to invert this. The base protocols (Ethereum, Bitcoin, Solana) would be "fat" because they held shared state and had native tokens. If anyone built a multi-billion-dollar app on top of Ethereum, the ETH token would go up, making the protocol fat and the application a thin, low-margin wrapper fighting over crumbs.

When ChatGPT launched, Silicon Valley immediately tried to run the exact same mental playbook.

They looked at OpenAI and said: *Aha! OpenAI is an L1.

*The Rosetta Stone

To understand what is happening in tech right now, it helps to map the crypto boom directly onto the AI race. It turns out Silicon Valley is playing the exact same board game, just with different game pieces:

The last two rows are the ones that actually determine who gets rich. In crypto, decentralized exchanges didn't care about website pageviews -- they cared about TVL and order flow. In AI, a platform with billions of casual "write a poem about my cat" prompts is dramatically less valuable than a platform that manages a company's entire codebase, tax filings, and customer support operations.* *

Let’s call it the Fat Model Thesis

The logic was simple, elegant, and extremely convenient for people trying to justify a $100 billion valuation for a chatbot company:

Frontier models are scarce, blindingly expensive to train, and useful across thousands of applications. Therefore, OpenAI or Anthropic gets to sit at the base layer of the new internet, collecting a toll on every single automated thought in the global economy. If you raised $10 million to build a neat writing assistant on top of GPT-4, you weren’t really a software company; you were just a "prompt wrapper" waiting to be executed when Sam Altman decided to add your core feature as a default button in the next product update.

Which worked brilliantly for about eighteen months. Except there is a minor, hilarious structural flaw in the comparison:

An AI model is not a blockchain.

A blockchain has a native token, an immutable ledger, and massive network effects. If you build Uniswap on Ethereum, Ethereum gets stronger. If you build a $5 billion SaaS company on top of GPT-4, OpenAI does not automatically get a cut of your equity, and your users do not need to buy "GPT-tokens" to stake the network.

OpenAI’s economic power didn't come from immutable cryptographic lock-in but from temporary capability scarcity.

And in tech, temporary capability scarcity is just an invitation for someone else to pirate your* math.d.*

Distillation is a Hard Fork

In crypto, if you don't like the governance or direction of a blockchain, you execute a "hard fork." You copy the codebase, copy the ledger history, copy the account balances, and launch a rival chain.

In AI, Meta releasing Llama was the first big hard fork. But then the market discovered something much funnier and far more disruptive: Distillation.

Distillation is the ultimate, non-consensual hard fork

If OpenAI spends $500 million and six months of GPU cluster sweat to train a smarter model, a bored developer in Shanghai or San Francisco can prompt that model a few million times, collect the outputs, use those outputs to train a tiny open-weights model for $50,000, and get 95% of the performance.

You spend half a billion dollars teaching a computer how to reason through complex legal docs, and three months later, a rival lab forks your intelligence onto a laptop chip for a fraction of the cost.

This ruins the "Fat L1 Model" thesis entirely. If you sell raw intelligence by the token, you aren't an invincible tech monopoly. You are ConEd. You are selling electricity or water.

You are in a commoditized, race-to-the-bottom utility business.

Work Under Management

So what do the "L1" AI labs do when they realize their infrastructure is getting commoditized?

They frantically run up the stack to become the application.

This is why every major AI lab has abandoned the polite fiction that they are neutral infrastructure providers. They are vertically integrating because selling raw API tokens is a punishing business, while owning the customer workflow is a glorious business.

In crypto, decentralized exchanges didn't care about website pageviews; they cared about TVL and order flow. In AI, a platform with billions of casual "write a poem about my dog" prompts is economically useless compared to a platform that manages a company's entire codebase, tax filings, or customer service ops.

The winning metric is actually Work Under Management** (WUM).**

If you map the tech stack wars out, it looks like a familiar 5-phase loop:

Who Opens a Bank Account for an Agent?

Which brings us to the final, beautiful twist in the story:

What happens when the software agents start hiring each other?

Suppose an autonomous AI agent working for your company needs to hire a specialized sub-agent to solve a complex math proof, optimize a database, or buy server space. The transaction takes 200 milliseconds and costs $0.003.

How does the agent pay?

An AI agent cannot walk into a Chase branch, present a physical driver's license, and open a checking account. It has no Social Security number. If you hand your agent a corporate Visa card and it attempts to execute 400 micro-transactions per minute, Visa’s fraud algorithm will freeze the card before you can even finish reading this sentence.

Traditional banking was built strictly for slow, carbon-based bipeds moving money between 9:00 AM and 5:00 PM. It is completely useless for a non-human population of autonomous software agents executing millisecond economic decisions.

And this is the ultimate irony of Silicon Valley over the last decade:

For ten years, crypto developers built permissionless, API-native, programmatic, streaming digital money. And for most of that decade, critics (often correctly) asked: "Why do humans need this when Apple Pay works fine?"

It turns out crypto wasn't built for hyper-financialized humans. Crypto was accidentally built as the native financial infrastructure for autonomous AI agents.

Agents can hold capital, sign smart contracts, and negotiate real-time pricing without asking a human manager for permission.

AI didn't kill Web3’s stack wars but inherited them.

Frontier models provided the brain, vertical applications captured the workflow, and permissionless crypto provided the pulse.

The non-humans have finally arrived, and their first order of business is opening a digital wallet.