I keep one rule when I audit a claim: a number without a timestamp is a rumor.
When a single figure โ America's lead over China in frontier AI, stated as 2.7% โ reaches me without a methodology, a task suite, a weighting scheme, or an evaluation date, I file it the way I file an anonymous wallet that claims to hold a treasury: unverified. Not false. Just unverified. That distinction, which most coverage of the current AI governance fight collapses into a slogan, is the entire point. The truth is buried in the timestamp.
For the past week I have been reconstructing the policy narrative now circling Washington and the BRICS summit, and the structural problem I keep hitting is not political. It is evidentiary. In my day job I trace fund flows across chains โ I once reconstructed 50,000 transactions across the final 72 hours of the TerraUSD depeg โ precisely because a blockchain is the rare system that records its own history without asking anyone's permission. The AI safety debate has no such ledger. It has press statements, anonymous social accounts, and a scalar that nobody can reproduce.
That is where I want to begin, because the policy fight is downstream of a data fight, and almost nobody is treating it that way.
Context: what is actually on the record
Strip the framing and the reported landscape is narrow. The Trump administration's posture is that any pause in frontier development hands the race to China, and that argument has been made the load-bearing wall of its policy. The hedge against that posture is testimony, not statute. Speaking publicly, Anthropic's Dario Amodei argued that the longer-term approach would be a cooperative attempt to set speed limits on AI progress. He also placed the center of gravity on hardware: to him, the concrete lever is preventing the sale of advanced chips to China and dismantling the smuggling networks that move them.
On the legislative side, the message is thinner than it sounds. Senator Johnson told CNN that Congress had already set the guardrails and that the responsibility now sits with the labs that build the models. Read carefully, that is two claims fused into one: there is a perimeter, and the perimeter is enforced by the people inside it. When you ask the obvious follow-up โ what federal statute compels a lab to comply โ there is no answer in the record. There is no mandatory instrument. The guardrail is advisory. That is not a small semantic gap; it is the difference between a fence and a suggestion.
Then came the more interesting admission. David Sacks, formerly a White House AI advisor, described the frontier as a duopoly โ Anthropic and OpenAI โ and, more usefully, said the quiet part aloud: that when AI executives call for slowdown, they also carry commercial motives, because a single destructive cyberattack traced to a model would expose them to liability claims. That sentence reframes safety spending. It is not compliance cost. It is balance-sheet risk management. And it quietly confirms that the guardrail is being enforced by market actors with a direct financial interest in where the barrier sits.
The sourcing itself is a signal. One of the keystone claims โ that China has committed to providing open-source AI models, development assistance, and training โ rests, in the reported account, partly on social-media accounts rather than an official communiquรฉ. I treat that the way I treat a wallet with no transaction history: not disqualified, but weighted down. A distribution program at that scale leaves fingerprints โ procurement records, compute allocations, model releases, support terms. Absent those, the commitment is a declaration. Declarations are cheap, and I have watched more than one protocol announce a partnership that never produced a single on-chain interaction.
So the reported record is this: open-source AI pitched to the Global South, an alarm number suggesting the US lead has compressed to 2.7%, and a low-cost accusation that Chinese labs have copied American models โ offered without the technical exhibit any auditor would demand. Now let me take it apart the way I would take apart a token.
Core: reconstructing the claims
Start with the 2.7%.
I spent eight weeks of my undergraduate life manually tracing over 500 swaps through Uniswap V1 on Ethereum mainnet, hunting a rounding error in the constant product formula that distorted small-cap pricing. The lesson that stuck was not the bug. It was the discipline: before you trust a number, you ask what produced it, over what sample, under what conditions, and you try to reproduce it. A capability gap expressed as a single percentage fails all four questions at once. Which benchmark? Which task distribution? Was it reasoning, code, multilingual, or agentic tool-use? Weighted how? Evaluated when? A model index is not a thermometer you can hold up to an industry. It is a composite, and composites move differently depending on their internal weights. A 2.7% aggregate gap could mask a dead heat in one capability and a substantial lead in another. Reporting the aggregate as if it were a temperature is the analytical equivalent of quoting 24-hour volume without checking who did the trading.
Which brings me to the thing I cannot stop noticing: the open-weight claim and the volume claim share a failure mode.
Open weights mean the parameters are downloadable and runnable on your own hardware, as opposed to a closed API you rent by the call. China's reported commitment to hand weights to the Global South is a real strategic choice โ it is the same route DeepSeek and Qwen walked through 2024 and 2025, and it is genuinely different from the Western default. But a downloadable model is not an adopted model. Distribution is not deployment, and deployment is not dependence. I have watched this exact conflation destroy credibility before. In 2021 I pulled 10,000 transactions from a Bored Ape floor and found that five interconnected wallets generated roughly 30% of reported volume through self-washing. The number on the dashboard was real in the sense that the trades existed. It was false in every sense that mattered. Wash trading is the ghost in the machine โ and the ghost does not only live on chain.
Apply it here. A model's download count can be inflated by mirrors, by automated pulls, by researchers who never deploy anything, by vendors who re-upload under their own banner. A download is not a user. If the West's read on China's open-source diplomacy is that downloads are climbing and therefore the ecosystem is shifting, the West is reading a vanity metric. The auditable questions are harder and duller: fine-tune activity, inference volume, enterprise adoption, support contracts, the number of organizations that put the weights into a production path and kept them there. Nobody in the current debate is supplying those, and without them the open-source flood is a narrative, not a measurement.
There is an even more basic gap in the reporting: the license. Open weights come with terms โ permissive, or restricted to non-commercial use, or subject to field-of-use limits. The entire geopolitical force of an open-source offer depends on which terms attach. A model that cannot be commercialized is a training tool, not an economic lever, and it competes differently. The story treats open as a category when it is a spectrum, and the spectrum is where the leverage lives.
Now take the compute lever, because that is where the argument is strongest and also where it is most naive.
Amodei is right that chips are the binding constraint in the short run, and the smuggling admission โ that a network exists to move advanced silicon โ is exactly the kind of gap I hunt for in liquidity. Nominal supply is not tradeable supply. I have built depth-chart analyses where the order book looked deep until you realized half the bids belonged to one actor ready to pull them on a tick; liquidity evaporates when logic fails. An export control regime that does not account for the informal channel is pricing nominal restriction, not effective restriction. If the reported smuggling network is real, the control is a screen with a known hole, and everyone who reads on-chain data knows that screens with known holes leak.
But the deeper problem is the equation itself โ the assumption that compute capped equals capability capped. It does not hold, and it never has. Algorithmic efficiency moves the floor. Mixture-of-experts routing, quantization, distillation, sparsity โ each of these lowers the silicon required to reach a given performance level, and they compound. I spent 2024 building a model that correlated ETF inflows with on-chain exchange reserves across 180 days, and the most useful thing it taught me was humility about single-variable stories. The market did not behave as reserve accumulation predicted in isolation; it behaved as reserve accumulation interacted with everything else. Capability is the same. Treating export controls as the master variable ignores the one force that has repeatedly defeated hardware restrictions: software that needs less hardware tomorrow than it did yesterday.
The constraint the debate keeps ignoring is not silicon but power. Training and inference are electrical before they are computational, and the grid is the one input no algorithm compresses. I have watched infrastructure narratives run into physics before โ the swap that looks profitable until you price the gas, the farm that looks sustainable until you price the emissions. Compute expansion is no different. Data-center energy supply, transmission capacity, and cooling are the true upper bounds, and they move on a timeline measured in years, not quarters. A race that ignores its power budget is not pricing its own pace.
There is a cleaner way to say it. The race framing treats progress as a speedometer โ a single, observable, negotiable rate you can dial up or down. Amodei's speed-limit language depends on that metaphor. But frontier progress is not a knob. It is a sum of many independent curves, several of which move in jumps. Pattern recognition precedes prediction, and the pattern here is that every attempt to treat a complex adaptive system as a dial has failed โ algorithmic stablecoins tried it, and the dial did not turn when the stress arrived.
Which brings me to the part of the record that worries me most, and it is not the competition. It is the responsibility transfer.
The reported arrangement is: no mandatory federal compliance, and full responsibility assigned to the labs. That is not a guardrail. That is a liability with no ceiling and no verification mechanism, handed to a handful of firms that compete with each other on the very dimension they are being asked to police. In the Terra post-mortem I ran, what killed the system was not one bad actor โ it was that the stability mechanism had never been stress-tested against the assumption that it might actually need to work. It was designed to look stable in calm conditions. It held until it was asked to hold, and then it did not. Safety promises made under competitive pressure share that architecture. They function as long as nothing tests them.
And Sacks's own framing concedes the problem: if safety spending is risk management, then it competes for capital against growth, and in a race the growth line wins every quarter it is permitted to. The only thing that reverses that ordering is an external auditor โ a statute, a court, an insurer, a standard with teeth. The record offers none. What it offers instead is a fragmentation risk nobody is pricing: if Washington vacates the field, state legislatures and foreign regulators fill it, and the labs end up navigating a patchwork that costs more in compliance friction than a single federal standard ever would. The absence of a rule is not the absence of rules. It is the presence of unpredictable ones.
The duopoly claim deserves its own line, because it cuts against the way the race is narrated. If two firms hold the frontier, then the US-versus-China story is partly a story about two American companies and their pricing power. Concentration is comforting in a sprint and dangerous in a regime where safety is self-policed โ because a duopoly has both the strongest incentive to defend the barrier it helped set and the least external oversight. I have learned to distrust any market where the loudest voices for orderly competition are the ones already ahead. That is not fraud. It is structure. But structure with no auditor is precisely where systemic risk accumulates quietly.
Contrarian: the correlation we keep assuming
Here is the move I want to make, and it will be unpopular on both sides of the cable.
Everyone in this debate assumes a correlation that has never been demonstrated: that whoever leads is somehow less safe, or that the party responsible for safety is necessarily the party holding the frontier. Neither is established. The race frame and the safety frame are being argued as if they sat on the same axis. They do not. They are two different measurements that the reporting keeps stacking into one bar chart until the chart looks like it says something.
The blind spot is this. The conversation is spending enormous energy on who is ahead and almost none on whether the models do what their builders claim. Both nations are racing past a verification gap they have jointly created. There is no independent evaluation regime, no reproducible safety standard, no third-party audit with subpoena power, no insurer pricing the tail. The US says regulation would cede the race. China proposes a governance framework that is, on the record, a framework โ not a technical standard. Neither side has produced the one artifact that would matter: a published, reproducible demonstration that a frontier model behaves within stated bounds under adversarial conditions.
There is a corollary the open-source enthusiasm tends to skip. Open weights are not open governance. Handing someone the parameters does not hand them the ability to audit training data, evaluation methodology, or the safety fine-tuning that was or was not applied. A downloadable model is a black box you are permitted to run, not a system you are permitted to inspect. If the strategic goal is trust, open weights deliver distribution without delivering verification โ and distribution without verification is exactly the surface on which vanity metrics thrive.
There is a tail risk both sides are underpricing by treating the absence of regulation as stability. The current regime holds only because nothing has tested it. A single destructive incident โ a cyberattack traced to a model, a misuse that produces mass harm โ would not produce a measured response. It would produce a retroactive one, written in the dark, under pressure, with the labs as the obvious defendants. Systems built on unverified trust do not degrade gracefully; they fail at the exact moment they are finally asked to prove something. The calm is a position, not the same thing as solid ground.
Correlation is not causation, and an index showing a 2.7% lead is not evidence that the leader is safer, faster to deploy, or better positioned. It is one number. Volatility is the tax on unverified trust, and the system is levying that tax on everyone โ it simply has not sent the bill yet.
Takeaway: what to watch next
Ignore the readouts from the next round of summits. They will produce statements, and statements are not evidence. Watch instead for the first artifact that can survive an audit: an independently reproducible evaluation, a model card with benchmark results someone outside the lab can verify, a safety claim with a named auditor behind it, a data point with a timestamp and a method attached.
That is the only signal that will tell you whether any of this is improving. Everything else, on both sides, is an unverified balance sheet. History is written in blocks, not promises โ and the AI race, so far, has produced almost no blocks worth reading. In the noise, the signal remains silent, until someone finally publishes the ledger.