The Moat Already Moved

The dominant debate in AI right now is about model weights: whether frontier models should be open or closed, whether Llama or Mistral can close the capability gap with GPT and Claude, or whether decentralizing the model layer will distribute power more broadly. It's a real debate, and it's worth having. But the crypto industry ran a similar "let's decentralize and give power to the little people" experiment a decade earlier, and the results bode evaluation.

What Crypto Taught Us About Where Power Lives

Blockchain was supposed to be the great financial equalizer. In meaningful ways, it delivered. Traditional remittance corridors still charge 6-7% on a $200 transfer on average globally, and in some routes the costs are staggering: according to a 2023 Migration Policy Institute report drawing on World Bank data, sending $200 from Tanzania to Uganda incurred fees of at least 39%. Bruh, what. Blockchain-based remittances have meaningfully cut costs in corridors like Nigeria and Vietnam. The protocol worked.

But the protocol wasn't where power ended up living.

Almost nobody interacts with Bitcoin or Ethereum at the protocol level. They use Coinbase, Binance, a wallet app, a bridge UI. The protocol decentralized. The interface consolidated. And the exchange layer consolidated even more aggressively: according to data from CryptoCompare, Binance's share of global crypto trading volume on centralized exchanges rose from 48.7% in Q1 2022 to 66.7% by Q4, driven in part by FTX's collapse, which wiped out its closest competitor and handed that competitor's market share to the dominant player. By 2025, Binance captures roughly 40% of global spot trading volume, with no close second. One CEO's tweets were enough to trigger a market-wide bank run. Not very decentralized..

The lesson from crypto isn't that decentralization failed. It's that the moat migrated. The protocol layer opened up, and power moved to whichever layer stayed closed: the on-ramps, the exchanges, the sequencers. Decentralizing one layer doesn't decentralize the system. It relocates the chokepoint.

The Master Switch is a book that documents a pattern across 20th century comms tech: telephone, radio, film, the internet. Each began open. Each consolidated. It seems to be the path of greatest efficiency. Because we know this pattern, we can be more intentional about where we allow the inevitable bottlenecks to form.

Is AI Repeating This?

The open source movement in AI is decentralizing the model layer. This is genuinely useful and worth supporting. But there's reason to ask whether the model layer is where power will ultimately concentrate, or whether the chokepoint is somewhere else.

The most obvious candidate is data. You can download Llama. You cannot download the hundreds of billions of human interactions that shaped GPT-4's RLHF training. RLHF = Reinforcement Learning from Human Feedback. You cannot replicate whatever enterprise data agreements give frontier models specialized knowledge of legal, medical, and financial domains. Open source can close the capability gap in model architecture over time. Whether it can close the data gap is a different question.

Hardware is easier to regulate than data. You can count servers, you can't count secrets. Researchers at GovAI have made a related observation about compute governance: compute is at least trackable and regulable. In a 2024 Lawfare piece, Lennart Heim, Markus Anderljung, and Haydn Belfield argue that governing compute is one of the more tractable entry points for AI policy precisely because it's visible. Data accumulation, by contrast, is harder to see and harder to govern. An NBER working paper on competitive moats in generative AI notes it's too early to know whether proprietary human feedback (the billions of user interactions that shape how models improve over time) will create durable advantages that open source simply can't replicate. But it flags this as one of the central open questions in AI competition.

The companies releasing "open" AI models are sharing the recipe but hiding the ingredients. A 2025 paper from MIT and Hugging Face researchers adds another layer: even as open-weight models proliferate, transparency about training data is collapsing. In 2022, the majority of downloaded models disclosed something meaningful about their training data. By 2025, fewer than 40% did. If you've ever wondered what the AI you use every day was trained on, the answer is increasingly: nobody's saying. The ecosystem is becoming more open in one dimension, model weights, while becoming more opaque in another: what data was used, and where it came from.

The Foundation Question

The foundation models we use today were trained on the collected output of human civilization: journalism, novels, code, photography, music theory, academic research. Much of it was used without explicit consent. Many of the people who created it didn't know it was happening and weren't compensated.

In 18th century England, land that farming communities had shared and maintained for generations was legally privatized through enclosure. Productivity increased. The gains flowed to landowners while the people whose labor had sustained the commons were excluded from the value their work had helped create. The enclosure wasn't illegal; it was accomplished through legislation that powerful interests shaped.

Whether AI training on humanity's collective output without consent or compensation constitutes a contemporary version of this is an ongoing argument. Kate Crawford traces the case carefully in Atlas of AI. Shoshana Zuboff's surveillance capitalism framework points in the same direction. Human experience is rendered as raw material for a production process that users never agreed to join. The ongoing litigation — including New York Times v. OpenAI — suggests courts are still working out the legal contours.

The "publicly available" defense does deserve engagement. But publicly available doesn't automatically mean available for unlimited commercial exploitation at civilizational scale. A musician playing on a street corner is performing publicly. That doesn't mean a record label can build a catalog from the performance without asking.

What Would It Look Like to Get This Right

Crypto was built to operate outside institutional frameworks — that was the point. AI isn't. It's being built inside large corporations that are already regulated entities, in countries with functioning institutions, and policymakers are engaged from the start rather than arriving a decade later.

If data is where power is likely to concentrate, then data is where governance needs to focus.

The good news is some proposals already circulating are worth taking seriously.

• Data commons: collaboratively governed repositories where creators contribute data under shared standards and researchers can access it for public-interest uses, rather than ceding all value to whoever scraped fastest. The GovLab and others have been developing blueprints for this since 2024.
• Consent and compensation frameworks at the point of data collection, rather than buried in terms of service. Longpre et al.'s 2024 "Consent in Crisis" paper documents how web sources are rapidly restricting AI crawling precisely because no such framework exists; creators are opting out because opting in was never designed to mean anything.
• The EU's Data Union Strategy, part of the AI Continent Action Plan published in April 2025, proposes a common internal market for data with shared standards and interoperability across sectors. The premise is that the infrastructure for AI development shouldn't be owned entirely by whoever had the capital to build it first.

None of these proposals is sufficient on its own, and none has been implemented at meaningful scale. But they point at the right layer. The open source debate is doing real work on model weights. The harder and more consequential question is what happens to the data that trained those models, who controls it, and whether the people whose knowledge powered the algorithms get a say.