The Third Superpower Has No Territory: Reading Paul Tudor Jones Through the On-Chain Compute Layer

When Paul Tudor Jones writes anything, the market reads it twice. In September 2023, the man who shorted Black Monday and banked a fortune published an op-ed in The Wall Street Journal under a headline most readers skimmed and moved past: AI may become the third superpower. No ticker. No trade idea. No position sizing. Just a macro legend telling the financial elite that the machine is not a sector. It is a sovereign.

That single framing accomplished what years of developer blog posts could not. It carried the AI risk conversation out of the research lab and onto the allocator's desk. It reclassified artificial intelligence from a growth theme into something that behaves like geopolitics — a power with scale, reach, and its own apparent intent, operating outside the borders of any nation.

But the op-ed stopped exactly where the crypto market starts. Read it closely and the omissions jump out. Jones never named a single company. He never named a single model. He never mentioned the one physical resource that makes any of this governable. Chaos is just data waiting for a pattern, and the pattern here is a man warning about a sovereign that has no territory, no treasury, and no army.

So I did what I do. I pulled the on-chain numbers, mapped the thesis onto the infrastructure that actually exists, and found the real story hiding underneath the headline.

Context: what Jones actually said, and who he said it to

Strip the rhetoric and the op-ed contains roughly seven information points, most of them restatements of the same idea. AI models reshape themselves thousands of times, causing goal misalignment. The consequences scale by multiplying across tens of thousands of users. The resulting entity is unpredictable and unbounded in reach. American AI company employees have already warned about the danger. Therefore, the leaders of the United States and China should make controlling the development and proliferation of AI — on a coordinated timeline — their remaining top priority.

That is the argument. Notice what it is not. It is not a technical paper. It is an opinion piece written by a non-technical authority and published in the one venue guaranteed to reach institutional capital and policy elites at the same time.

The technical core of the claim — that models reshape themselves thousands of times — is where the argument loses contact with engineering reality. Production large language models do not autonomously rewrite their own weights or architecture. What Jones is almost certainly gesturing at is a collision of three very different mechanisms: self-play, the strategy iteration loop behind AlphaGo and AlphaZero; continuous online learning and feedback updates; and recursive self-improvement, the theoretical frontier where a system improves its own intelligence. Folding these into one sentence is not a harmless simplification. It flattens two distinct threat categories — a model that hallucinates a fact, and a model that pursues a goal orthogonal to its designers' — into a single word: risk.

Then there is the reach claim. Multiplied across tens of thousands of users. By early 2023, ChatGPT had crossed a hundred million users. The order of magnitude is off by four digits, which tells you the model of deployment in the author's head is smaller and more abstract than the one actually running on the world's servers. And directionally, his warning is not wrong — deployment scaling genuinely multiplies the absolute blast radius of any single flaw. The mechanism is just described in a way that would not survive a code review.

That matters, because I spent a decade reading contracts the way other people read headlines. In May 2017, I reverse-engineered the 0x protocol's v2 smart contracts within forty-eight hours of mainnet launch, ran an on-chain liquidity monitor, and found an impermanent-loss mispricing that let me clear fifteen trades in under ten minutes and forty-two thousand dollars before the bug was patched. The lesson from that window was not that I was fast. It was that narrative and mechanism decay at completely different rates. The narrative runs for months. The mechanism corrects in minutes.

None of this makes Jones wrong that AI governance matters. It makes him directionally correct and mechanically vague — the exact profile of a risk story that is easy to sell and hard to verify.

And verification is the whole game. The moment you ask how you prove a country complied with the timeline, you have left geopolitics and entered cryptography.

Core: the chokepoint is compute, and crypto priced it first

Here is the part of the thesis nobody on the desk talked about. You cannot control AI development at the model layer. Model weights are just numbers; they copy for free, they cross borders in a packet, and once they are open-sourced they are ungovernable by construction. If you want a leash, you need a lever. And the only physical lever AI respects is compute.

Training a frontier model takes tens of thousands of advanced GPUs running for months under a single roof. That demand collapses into three bottlenecks that are genuinely hard to hide: the silicon itself, almost entirely concentrated in one supply chain; the fabrication capacity, held by a handful of fabs; and the electrical power, which is local, immovable, and metered.

This is not theory. It is the exact logic behind the October 2022 US export controls on advanced chips to China and the August 2023 investment restrictions that followed. When policymakers want to slow a rival's AI program, they do not draft a treaty about model alignment. They gate the GPU. The most enforceable AI policy on Earth is a customs list.

Now watch what the crypto market did with this insight years before it became an op-ed.

Decentralized compute networks — Render, Akash, io.net, Bittensor — exist for the plain reason that GPU supply is fragmented, expensive, and geographically trapped. These protocols turn idle silicon into a liquid, permissionless, globally addressable market. In a world where compute is the chokepoint, a decentralized compute market is not a novelty. It is an arbitrage on the chokepoint itself.

I tested this directly. Working with a decentralized AI agent team in early 2026, I deployed three autonomous agents on an Ethereum L2 and ran them against cross-chain micro-inefficiencies for two weeks. They cleared eighteen thousand dollars. The lesson was not that the agents were smart — the lesson was that they were bidding for compute in a market where the spread between what a GPU costs to rent in one jurisdiction and what it earns in another is wide enough to drive a freight train through. Sustainability is just a loan from the future, and compute markets are the clearest example of that loan being drawn down every single day.

Here is where the on-chain lens beats the macro lens. Jones wants a coordinated timeline between two superpowers. Fine. But a timeline is only as strong as its verification mechanism, and verification is the one thing blockchains have spent fifteen years solving.

How do you prove that a training run stayed under a compute threshold? Not with the tools governments currently hold. But you can get closer than most people realize:

• Hardware attestation. Trusted execution environments and secure enclaves can cryptographically sign that a workload ran on specific silicon. Trust is a variable, not a constant — you do not trust the operator, you verify the hardware.
• Verifiable computation. Zero-knowledge proofs and fraud proofs can, in principle, attest that a given output followed from a given input. Proving a frontier-scale training run is currently prohibitive in cost, but the direction is unambiguous and the research is compounding.
• On-chain compute accounting. A decentralized market produces an auditable ledger of who rented what, when, and where. That is a regulator's dream — a real-time, tamper-evident record of compute consumption that no single party controls.

None of these are treaty-ready today. But they are the actual technology of AI governance, and they live in the crypto stack, not the diplomatic one. The people writing think-pieces about coordination have no idea the enforcement layer they are asking for is being prototyped by protocols they have never heard of.

Core, continued: the precedent that should terrify every developer

Now the darker half of the story, and the reason I think the framing is more dangerous than it looks.

The op-ed calls for controlling development and proliferation. That phrase sounds neutral. It is not. Read it against what the United States already did to Tornado Cash.

In August 2022, the Office of Foreign Assets Control sanctioned the Tornado Cash smart contracts — not a company, not an individual, but open-source code published to a blockchain. Then it went further. Core developer Alexey Pertsev was arrested in the Netherlands and convicted. Roman Storm was prosecuted in the United States. The precedent is now baked in: publishing code can be treated as an offense, and the open-source developers who release it can be held personally liable for how strangers use it.

Sit with that. Jones is asking the world's two largest powers to agree on a timeline for controlling the development of AI. If the enforcement philosophy of the Tornado Cash sanctions migrates into AI governance — and there is no structural reason to believe it will not — then every person who publishes an open model, every team that ships a training script, every researcher who posts weights on a public repository becomes a potential defendant. The code is the crime.

I have watched what happens when legal risk attaches to source code. Developers do not stop building. They build in jurisdictions that will not extradite, they route through anonymous infrastructure, and they publish under pseudonyms. The effect of aggressive code liability is not to stop proliferation. It is to move proliferation out of the light and into the dark, where it becomes harder to monitor, not easier.

That is the contradiction at the heart of the op-ed. It wants control. Its only workable control mechanism — compute gating — is leaky. Its only alternative — code liability — drives the very activity it fears straight into the shadows.

Contrarian: the metaphor is a category error, and the trade is mispriced

Here is what almost nobody has said out loud. The third superpower framing is a category error, and category errors are where the money is.

A superpower has territory, citizens, an army, a treasury, and a foreign ministry. AI has none of these. What AI has is the ability to act across all of them at once. That is a different thing, and conflating the two distorts both the risk and the opportunity.

The real risk is not that a machine wakes up and pursues a misaligned goal. The real risk — the one already happening — is cartelization. Frontier AI development has consolidated into a handful of laboratories that sit on the most compute, hire the scarcest talent, and write the rules their competitors must follow. A superpower with no territory is just a cartel with better marketing.

And here is the trade angle that matters more than any safety narrative. If compute is the chokepoint, then the most valuable assets in the AI economy are not the applications and not the models. They are the layer beneath — the silicon, the power, and the markets that price them. The market is currently paying decacorn valuations for applications built on someone else's chokepoint. Every established venue knows the first rule of a supply squeeze: first in, first served, or first to flee.

I have seen this movie. In May 2022, when Terra collapsed and the crowd screamed about algorithmic stablecoins, the actual signal was in Anchor's withdrawal queue. The real-time data showed exactly where UST would run out of buyers, and it did not care about anyone's narrative. Liquidity didn't lie; the storytellers did. The same discipline applies here. Ignore the op-eds. Watch the compute supply. Watch the power contracts. Watch who signs the multi-year GPU reservations.

Crypto's own AI tokens deserve the same skepticism. Most are narrative wrapped in a token and priced on a slide deck. But underneath the noise sits a genuine structural idea: that a permissionless, verifiable compute market is the only architecture in which AI governance is even technically possible. The tokens may be junk. The plumbing is not.

It is worth noting how fragmented the actual regulatory response has already become, because it dismantles the fantasy of a single coordinated timeline. The EU AI Act took a risk-tiered approach. The US executive order that followed the op-ed leaned on unilateral reporting thresholds rather than bilateral coordination. China built a filing regime. The Bletchley declaration gathered signatures but carried no binding force. Four jurisdictions, four philosophies, zero shared verification mechanism. A coordinated schedule between Washington and Beijing was never the realistic path — multi-point containment was, and that is what we got.

Takeaway

So what do you actually do with all of this?

Watch the one number no op-ed can spin: the compute supply chain. Track GPU reservation deals, fab capacity announcements, and data-center power contracts, because those are the real votes on where AI's sovereign territory will sit. Track the crypto compute protocols that are quietly becoming the market where that scarce capacity gets priced, rented, and audited. And track the legal language, because the Tornado Cash precedent tells you the governance fight will arrive as developer liability long before it arrives as a treaty.

Jones is right that AI deserves to be governed. He is wrong that the lever is a handshake between two capitals. The lever is compute, the verification layer is cryptographic, and the enforcement is being decided in courtrooms right now.

The race wasn't about who thinks hardest. It was always about who controls the plug. The question is whether the market figures that out before it prices another round of AI applications built on someone else's chokepoint.