There is a story that has hardened into conventional wisdom in Washington (and in debate), and it goes like this. The United States is only a little ahead of China in AI. That lead is fragile. If we lose it — even at the margins — China will use its models to invade Taiwan, dominate the battlefield, and export authoritarianism until global democracy collapses. Therefore we must do whatever it takes to stay ahead: build without limit, restrict without hesitation, and treat every month of advantage as an existential stake.
You heard every note of it in July 2026, when Moonshot’s open-weight Kimi K3 topped a US coding leaderboard and the reaction ran straight to “China just erased America’s AI lead”. The former White House AI czar David Sacks called it proof that data-center bans and regulation are “how you lose the AI race.” That’s the argument I want to take apart.
Let me be clear about what this is not. I am not arguing that we shouldn’t build AI or data centers. As I’ve argued elsewhere, development matters — for access, for the diffusion of the technology’s benefits, and because the United States will genuinely need its own sovereign AI capability rather than dependence on anyone else’s. That case I take seriously. What I’m pulling apart here is something narrower and more specific: the way the national-security version of the argument is often presented. Because the arguments as presented sometimes defy credulity — and the most hyped version of all, the one that runs straight from “we slipped at the margins” to “China invades Taiwan and global democracy collapses,” usually does.
So what follows is not a single thesis marched in a straight line. It’s a menu of answers — a set of distinct responses you can make to the dominance argument, laid out so you can see the whole board and choose the ones that fit the case you’re making. Some are empirical (”the gap is smaller or less durable than claimed”); some are strategic (”even a real lead doesn’t buy what they say it buys”); some flip the argument on its head (”chasing the lead is the danger”); and some contest the premise itself (”Chinese AI leadership wouldn’t be the catastrophe you’re assuming”). You do not have to believe all of them, and — as I’ll flag at the end — a couple of them actively conflict, so you’ll have to pick. But taken as a set, they show how much weaker every link in the chain is than the confident tone suggests. The problem is not that the United States is behind; by most public measures it is ahead. The problem is the belief that a marginal lead in model quality is the only thing standing between us and catastrophe. I’ll start with the argument that frames all the others — that the lead is already non-unique — and then go point by point.
Non-unique: the American “lead” is already gone
Before any of the point-by-point answers below, start with the one that dissolves the whole argument: the lead the dominance case depends on is non-unique — it has already diffused. This isn’t a dove’s wish. It’s the read from a table of AI investors and technologists — Peter Diamandis, Alex Wissner-Gross, Dave Blundin, Salim Ismail, and Emad Mostaque — on the July 19, 2026 Moonshots emergency episode recorded after Kimi K3 dropped. This entire podcast is carded and available to DebateUS subscribers.
These are people paid to be right about where AI is going, and they are anything but AI pessimists. Their read:
Kimi K3 is at the US frontier — and, some on the panel argue, near AGI. K3 landed as the #3 model on the cost–performance frontier, jumping 17 places to take #1 on the frontend-coding arena and topping six other domains — built on a recognizable transformer with, in Wissner-Gross’s phrase, “no magic,” which pointedly raises the question of what the US labs’ billions are actually buying. Blundin goes further, arguing K3’s ability to design its own chip and kernels already feels AGI-adjacent.
It’s about to be handed to the entire world. Full open weights were set to drop around July 27 — anyone on Earth can download and run it — and Xi has signaled China will back open source as a public good rather than restrict it. As Ismail put it, frontier intelligence is now a “totally perishable asset.”
US firms can run it for a fraction of the price, gutting US-lab valuations. Why pay for frontier APIs when a near-frontier open model runs on your own hardware? The panel pegged the US frontier labs at roughly a quarter of their value of three months earlier — a ~75% haircut — as the “frontier premium” collapses into one commoditized ingredient among many. (Mostaque’s analogy: the US labs are the brand-name drug; the open weights are the generic, at a fraction of the cost.)
It’s radically efficient — which undercuts the data-center premise. K3 was trained around US export controls, on second-tier chips. Blundin points to the nano-GPT “speedrun,” where the cost of producing a GPT-2-class model has fallen ~99%, and argues K3 proves those efficiencies hold at frontier scale — a roughly 1%-cost path to what a multibillion-dollar mega-cluster produces. If that holds, the “we have more data centers” advantage matters far less, for both training and inference.
It’s a step toward recursive self-improvement. Per the panel, K3 designed a chip and wrote its own kernels — models improving models. Blundin’s stronger claim is that the RSI threshold was quietly crossed around Opus 4.8: a system doesn’t need Einstein-level genius to start the loop, only the ability to make itself ~10x faster, which then compounds.
Capability is moving onto local hardware. Aggressive quantization is already putting frontier-class models on phones and laptops (a 27-billion-parameter model running on a smartphone; Mostaque predicts Fable-level capability on an ordinary MacBook within roughly 18 months). If intelligence runs on the edge — in every laptop, vehicle, and robot — the case for a hyperscale-gated national monopoly erodes further still.
You do not have to swallow the panel’s most bullish leaps — “already AGI,” “RSI is here,” “labs down 75%” are contestable, and I flag them as their forecast, not settled fact. But you don’t need the strong version. Even the conservative reading — a near-frontier, open, cheap, efficient, self-hostable Chinese model, shipping this month — is enough to make the American lead non-unique: whatever edge exists is already leaking to everyone, this quarter. Everything below is why that leak doesn’t hand China a war, or the end of democracy, either.
First, the data-center gap is enormous — and it doesn’t do the work people think it does.
Start with the physical layer everyone points to. As of July 2026 the United States has roughly 4,540 data centers to China’s 369 — depending on the counting method, Stanford’s AI Index puts it at 5,427 versus 449. Either way, the US lead is better than ten to one. The spending gap is wider still: Goldman Sachs projects US cloud providers’ capital expenditure at roughly $1 trillion in 2027, about eight times their Chinese counterparts. Even if we imposed a full moratorium on new construction tomorrow, the United States would still hold a better-than-tenfold advantage in facilities.
More to the point: we already have more than enough compute to train even the largest models anyone is seriously proposing — including runs far beyond today’s frontier. The data centers that exist are sufficient for the training the doomers worry about, and if anything, a government or a handful of firms could prioritize that capacity for military or strategic use. (The real cost of concentration, which I’ve written about elsewhere, is that it denies access to everyone else — not that it leaves the frontier unbuilt.)
There’s a wrinkle that cuts against the “we need more concrete and power” framing, and it’s worth naming honestly: constraint on compute has, so far, driven efficiency rather than stopping progress. The less processing capacity you have, the harder you’re forced to innovate on doing more with less — and that pressure produces genuinely better methods. When export controls choked off China’s chip supply, DeepSeek and others responded by training competitive models with far less compute, because scarcity was the mother of the optimization. Kimi K3 is the same lesson a year later: a roughly 300-person lab, working around a chip shortage, shipped a 2.8-trillion-parameter model competitive with the US frontier — through mixture-of-experts routing, low-precision quantization, and infrastructure built around scarcity, as a former DeepMind researcher put it. Bank of America analysts noted that K3 shows architectural work can still deliver step-change gains despite compute constraints. If the frontier is increasingly about efficiency rather than brute scale, the whole “compute moat” thesis — the idea that raw FLOPs gate capability — weakens, and “we have 4,540 data centers” is a less decisive advantage than it looks. (Hold onto this point about scarcity-driven efficiency; it matters at the end, because it doesn’t sit comfortably next to every other argument here.)
Second, we don’t actually know who has what.
The public comparison is not the real comparison — and even the public comparison is subtler than the headlines. Kimi K3 did take #1 on the Frontend Code Arena, but that’s one narrow task; on broad intelligence measures it lands behind Claude Fable 5 and GPT-5.6 Sol, roughly level with the older Opus 4.8, and its weights weren’t even out yet when the “erased the lead” pieces ran. Anthropic alleges Moonshot reached that point partly by distilling millions of exchanges with Claude — an unproven claim, but if true it describes very effective catch-up, not independent frontier invention.
Now layer on what isn’t public at all. The tier above the released models — the unreleased Mythos-class systems — is a genuine step change: on cyber capability, one analysis puts Mythos about seven months ahead of trend, with a Mozilla executive calling it as capable as the world’s best security researchers, and a credible estimate is that China won’t have a Mythos-equivalent until roughly February 2027. US labs are also widely understood to be running models internally beyond what they ship; China may be too. So we’re comparing the visible tips of two icebergs and making confident claims about their total mass. That’s not analysis; it’s guessing with extra steps.
Third, the “margin” is being asked to do absurd work.
Grant the premise: say the United States is a few months ahead — Fable 5 is a bit better than Kimi, or the unreleased tier is six months better. The leap from that to “so China won’t invade Taiwan” is not a small inference. It’s a non sequitur. And the “few months” is doing heroic work: on the intelligence indices the top three models now sit within about three points across three labs, and Fable 5’s lead has narrowed from four points to one since June. This is a photo finish being sold as a decisive edge.
These models are not out on the battlefield fighting wars. At most they are amplifiers — helping with targeting, intelligence synthesis, and some autonomy in weapons systems. And even that is uncertain at the margin: we don’t know that a marginally better model meaningfully improves a weapons system, let alone decisively changes a battle, let alone tips the calculus around an amphibious assault. Consider the physical end of it: MIT roboticist Rodney Brooks estimates it will take more than a decade for humanoids to operate reliably in complex, unfamiliar settings — the manipulation problem of simply picking up a gun on rubble-strewn stairs remains unsolved. Saying “our model is slightly better, so they won’t invade” is like saying “our fighter jet is slightly better than yours, so we’re going to take Taiwan” — except the model isn’t even the jet.
Here’s the tell: nobody making this argument ever specifies the internal link — the actual causal chain by which a six-month model lead prevents an invasion. Taiwan-invasion math is dominated by things AI barely touches: the brutal difficulty of an amphibious landing, US and allied commitments, the economic catastrophe an invasion would trigger for China itself, the near-certainty that the chip fabs get destroyed in the process, and Xi’s domestic legitimacy. AI is a rounding error in that decision, not the swing variable.
Fourth, recursive self-improvement cuts against the hawk, not for him.
Every serious lab is aiming at recursive self-improvement — models that help build better models. It isn’t fully autonomous yet, but the direction is visible, and there are now concrete data points: Weco AI reported an outer-loop agent that rewrote its own inner researcher through seven versions in eight unattended days, beating two years of hand-tuning — and, because it was scored on a metric it couldn’t game, it even learned to cheat less. This is the strongest card the “protect the lead” side thinks it holds, and its sharpest advocates say so plainly. A former OpenAI research VP argued that the publicly available models of mid-2026 essentially don’t matter — that whether China is two months or eight months behind is beside the point, because “the only thing that matters is the race to RSI,” which needs two things: a model with excellent research taste, and enormous compute. If takeoff is fast and winner-take-all, whoever is ahead when it begins compounds a small edge into a permanent one.
So I want to meet that head-on rather than dodge it — and notice it cuts both ways. If the RSI race is what matters, then the whole “China’s public model caught up, therefore catastrophe” panic is aimed at the wrong variable; the leaderboard is a sideshow. What’s left of the hawk’s case is the compute-plus-taste argument — and that only converts a lead into permanent dominance if takeoff is clean and fast. The evidence points the other way, toward diffusion and fast-following: the open-weights story below is the frontier leaking within months, “N-minus-one” being good enough for almost everyone. If capability diffuses that quickly, a three-to-six-month lead at the starting line doesn’t compound; the follower is climbing the same curve a few months back. You can believe in a hard takeoff or you can believe the observable pattern of rapid diffusion. You cannot use the second to describe the present and the first to justify the policy.
Fifth, a huge lead is destabilizing, not safe.
Now flip the thought experiment. Suppose the United States doesn’t just stay marginally ahead but pulls away massively — a system well beyond anything public, capable of running offensive cyber operations, penetrating another country’s military networks, and disabling its weapons. This isn’t hand-waving: the unreleased Mythos tier is already described as as capable as elite human security researchers and roughly seven months ahead of trend on cyber. Push that lead further and it becomes exactly the kind of decisive, deterrent-neutralizing capability the scenario imagines. Does that make us safer?
Not obviously. It can do the opposite. A country that concludes it is falling decisively behind — that its deterrent is about to be neutralized — has every incentive to act before the window closes. As the China-technology analyst Paul Triolo argues, if Beijing comes to believe US firms are nearing AGI, that belief alone could prompt action against Taiwan it would not otherwise have considered — a direct escalation risk manufactured by the race for compute dominance. China might strike Taiwan specifically to destroy the chip-fabrication facilities and choke off the West’s access to advanced silicon, since most leading-edge chips are fabricated there. Or it might move against the United States directly. Russia could reason the same way. This isn’t an anti-China or anti-Russia point — it’s the logic of the security dilemma. A runaway lead doesn’t end the game; it creates a powerful incentive for preventive war by whoever is losing it.
This isn’t only a dove’s worry. Eric Schmidt, Alexandr Wang, and Dan Hendrycks — hardly AI pacifists — argue in their Superintelligence Strategy that a hurried bid for AI dominance endangers every state: a rival that sees a destabilizing project nearing completion has reason to sabotage it, from cyber operations that degrade training runs to physical strikes on data centers. They call the resulting standoff Mutual Assured AI Malfunction — a nuclear-MAD analogy in which no one dares grab for outright monopoly because the attempt itself invites a crippling response. That is the opposite of “get a big enough lead and you’re safe.”
History is not reassuring here. The United States held an absolute nuclear monopoly after 1945 — and it lasted barely four years, with the Soviets arriving fifteen years sooner than Washington expected despite total secrecy, export controls, and an enormous head start. The assumption that a technological lead is durable and decisive has a weak track record. Both a small lead and a huge lead undercut the original argument — just in different ways.
Sixth, what matters is distribution and integration, not raw model quality.
The thing that actually turns a capable model into military or economic power is not the model — it’s getting the intelligence deployed across real institutions. And that turns out to be the hard part.
We can see it in the military itself. Claude — Anthropic’s model, running inside Palantir’s Maven system — was reportedly used for targeting and intelligence work in the strikes on Iran and in the Venezuela operation. But getting there took deep, unglamorous integration. A Pentagon official noted that Anthropic had become “deeply embedded” partly because it provided forward-deployed engineers while competitors hadn’t. That’s the whole ballgame: the bottleneck is human and organizational, not a few points on a benchmark. As one former Navy intelligence officer put it, success or failure in war depends on the people using the machines, not the machines.
This is precisely where a marginal model edge stops mattering and adoption capacity starts. And on that dimension China may have advantages — a larger, more technically capable workforce to drive integration. Alibaba’s chairman has argued the AI race will be won by who adopts it fastest, not who builds the strongest model. Whether or not he’s right, he’s pointing at the correct variable.
Seventh, a lot of the market doesn’t even want a smarter model.
The “we must lead on capability” frame assumes everyone is desperate for maximum intelligence. Many firms aren’t. They already find current models highly capable; their problem is integration, not IQ. Some explicitly prefer to sit at “N-minus-one“ — one step behind the frontier — because it’s cheaper and good enough. And a genuinely disruptive superintelligence could be bad for an incumbent business: it would lower the barrier for competitors to replicate what you do. Plenty of companies have no interest in accelerating toward AGI or ASI that could dissolve their own moat. The crux, again, is deployment — not frontier bragging rights.
Eighth, the democracy threat is overstated — and mostly domestic.
The scariest version of the argument is that if the US loses its lead, China exports a tyrannical model of AI and democracy collapses worldwide. Two problems.
First, exporting a model of governance requires distributing the technology — into other countries’ businesses, agencies, and development plans. The United States is not good at AI distribution; we can barely get these tools deployed inside our own companies. Doing it abroad, at scale, against a determined competitor, is harder still. Meanwhile China just signed 29 countries onto a China-led World AI Cooperation Organization — mostly Global South nations, with no major Western democracies and few US firms in the room. That tells you the world is already dividing along lines that track existing alignments. Western democracies are not about to run their governments on Chinese models; they aren’t doing it now, and a marginal capability gap won’t change that.
Second, on surveillance — the concrete fear underneath “authoritarian AI” — the difference is again one of degree, not kind. The United States already hosts pervasive surveillance: Flock alone operates well over 100,000 automated license-plate readers generating billions of vehicle scans a month, with data flowing to federal agencies. The very Palantir stack used for military targeting abroad now underpins domestic immigration enforcement and predictive policing at home. Having spent real time in China, I don’t dispute there’s more surveillance there. But it is a marginal difference layered on top of a country that already tracks its residents extensively — not a bright line between a free world and an unfree one.
And here’s the reframe the hawks skip: the larger AI-and-democracy risk may be internal — concentration of power in a handful of labs and agencies, mass surveillance, synthetic disinformation. That danger exists regardless of who “wins” the race. The AI researcher Ben Goertzel makes the point sharply about governance itself: the danger is handing the shaping of AI to some central committee — whether it sits in Washington, San Francisco, or Beijing — and much “we must act now” safety rhetoric conveniently pulls the ladder up behind whoever is already at the frontier. If you actually care about AI and democracy, the answer looks more like wide distribution and participation than a protected national champion. Pointing at Beijing is partly a way of not looking at home.
Ninth, on pure economic competitiveness, open weights flip the board.
If the worry is industrial competitiveness, the open-source strategy quietly undoes the whole argument. The leading Chinese models are open-weight and downloadable — anyone can run them locally, modify them, and control them end to end. Qwen has become the most widely used open model family in the world, and Kimi K3 shipped its full weights days after launch so companies and governments can self-host. US firms are already using these models: Airbnb relies heavily on Qwen for customer service, calling it fast and cheap, and Perplexity and Nvidia have used it too. As Axios put it, America may still push the frontier forward, but it cannot stop the rest of the world from choosing a cheaper alternative.
Three consequences follow. (a) These models are dramatically cheaper — often a fraction of the cost of US frontier APIs, in some cases an order of magnitude less. (b) Access to US frontier models is politically contingent in a way open weights are not: a lab or the government can restrict them — recall the periods when the most capable models were walled off from firms with foreign employees, or when Congress opened probes into companies using Chinese models — and a company facing that friction can simply keep using the open alternative. (c) As above, the extra “intelligence” in the US model is of little material value to most deployments, while cost and control are worth a great deal.
I’ll concede the obvious counterpoint rather than pretend it away: diffusing capable open-weight models everywhere also diffuses misuse — cyber and bio uplift — to actors who couldn’t build it themselves. This is concrete, not hypothetical: the UK AI Security Institute finds open-weight models now trail the closed frontier on cyber by only four to seven months, down from six to ten, and that safety measures bolted onto open models are largely ineffective. That’s a real problem. But it’s a problem the “maintain American dominance” frame doesn’t solve, since the models are already out. It argues for safeguards, not for the race.
Tenth, it’s the whole ecosystem — not the single best model.
Put the pieces together and the unit of competition is not a model; it’s an ecosystem: the policies that nurture development, the runway industry has, the consistency of the rules, the workforce that can integrate, the capital that can fund it. On some of these the US leads; on others it doesn’t. China’s policymaking has costs — a company that tried to sell into a foreign partner can find itself blocked by its own government — but it also offers a kind of continuity. None of this is a “China is better” claim. It’s that the outcome is shaped by a dozen factors, and the marginal quality of the single best model is one of the least decisive of them. The proof is already in front of us: US firms are choosing Chinese models even while the US model is widely believed to be a bit better.
Eleventh, the lead may not even last.
Leads in general-purpose technologies diffuse. The nuclear case is the cautionary tale, but the AI case is arguably faster. The Council on Foreign Relations’ Lauren Kahn argues the whole “first-mover advantage” premise is largely overhype, and any such advantage would be unsustainable: AI is a general-purpose, private-sector-driven, largely open-source technology — closer to electricity than to a secret weapon — and it spreads faster, not slower, because market incentives push it out the door. As recently as April 2026, the US government’s own AI testing center assessed China’s best models as about eight months behind; US leaders took comfort in a six-to-twelve-month cushion. Kimi K3 suggests that cushion collapsed faster than expected — and, as Axios noted, even if US labs pull ahead again, China has shown it can close the gap quickly. Export controls, intended as a ceiling, functioned partly as a forcing function, pushing Chinese labs toward efficiency and domestic silicon. As one researcher put it, DeepSeek’s success suggested export controls are ineffective at preventing other countries from building frontier models. (In fairness, others argue the controls are working precisely by constraining China’s aggregate training resources over time — the debate is genuinely unsettled.) Add the mobility of talent — including the large share of top US-lab researchers of Chinese origin, whom “beat China” policies like loyalty screening and visa restrictions tend to push away — and the picture is of a lead that is hard to hold, not a moat.
Twelfth, and most important: the race framing is itself the danger.
Here is the turn. The zero-sum “we must beat China or else” frame is not a neutral description of the situation. It is an active driver of the single most catastrophic risk in AI — the pressure to cut corners on safety. “Win at all costs” translates directly into: deploy faster, test less, override caution. You can see the vise closing in the reaction to Kimi K3: as Axios framed the administration’s dilemma, tougher safety rules could slow US labs just as China accelerates, while looser oversight lets them move faster at the price of releasing dangerous capabilities. The more totalizing the race rhetoric, the more that dial gets turned toward “faster, looser” — and the more likely someone fields a system they don’t understand and can’t control. This is the mechanism a US government-commissioned assessment by Gladstone AI identified: competitive pressure pushes labs to accelerate at the expense of safety and security, which both raises the odds of losing control and makes the most advanced systems a richer target to be stolen and turned against US interests. A bigger lead, on that logic, is a bigger bullseye — not a shield. The dominance argument, taken to its conclusion, manufactures the very danger it claims to be protecting us from. If you actually believe advanced AI is dangerous, the racing is the threat — not the six-month gap.
Thirteenth, contest the premise: China leadership may not be the catastrophe assumed.
Everything above grants the frame’s core premise — that Chinese AI leadership would be a disaster — and disputes the mechanism. Debaters can also go after the premise itself. These run hotter and are more contestable, and I’d deploy them knowing that, but they’re live and they’re made in the literature:
There may be no US–China war to prevent. Han Dongping notes China maintains the only unconditional No-First-Use nuclear pledge among the nuclear powers, and argues the cost of a Taiwan or South China Sea war is too high and the returns too low for either side to rationally start one — the “Thucydides Trap,” on this view, is not fated.
The competition may be positive-sum. Kai He and Huiyun Feng argue that US–China “institutional balancing” has generated positive externalities — regional cooperation, dynamism, and the provision of public goods — rather than the inevitable slide to conflict that power-transition theory predicts.
China may not be trying to remake the order — or even chasing AGI. Former national intelligence officer Paul Heer argues Beijing is not seeking to remake world order; Suisheng Zhao casts China as a stakeholder that has propped up the WTO, the UN, and the Paris accord as Washington retreats. And the Wall Street Journal reports Xi has said little about AGI, pushing his industry instead toward practical, low-cost applications — a fundamentally different race than the one US boosters are running.
The moral premise is shakier than the frame admits. The sharpest version presses the double standard directly: it is a form of chauvinism to simply assume American tyranny is categorically preferable to Chinese tyranny. And the domestic picture isn’t pristine — a Bright Line Watch survey of more than 500 political scientists found US democracy ratings dropping sharply, with many describing a slide toward “competitive authoritarianism” of the Hungary/Turkey variety. If the democracy we’re racing to protect is eroding from within, “beat China to save democracy” is carrying less weight than it claims.
Even the safety case can cut toward sharing, not hoarding. Some argue Beijing takes AI safety seriously — it’s become a stated political priority, with pre-deployment assessments and a surge of national standards — so the “reckless China” caricature that fuels the race is at least overdrawn.
A word of caution on this cluster, since it’s the spiciest on the menu: the “China is more peaceful,” “US democracy is cooked,” and “chauvinism” turns are high-variance. They win big in front of a judge who rewards a willingness to interrogate the frame’s assumptions, and they backfire in front of one who reads them as apologetics. Run them deliberately, with the sourcing attached, not as throwaway zingers.
Anticipating the obvious objections
Let me answer the strongest replies before they’re made.
“China is only copying — catching up isn’t the same as inventing.” This is the strongest version of the other side, and it deserves a fair hearing. Growth economics does distinguish technological catch-up from frontier invention: a follower can move fast by observing the leader’s designs and skipping its failed experiments, and matching an existing capability doesn’t prove the capacity to generate the next one. If Anthropic’s distillation allegation holds, some of Kimi’s leap is exactly that kind of assimilation. Fair enough — but notice this concedes my central point rather than rebutting it. If China is a fast, cheap follower, then the world still gets near-frontier capability at a fraction of the price on a short delay, which is precisely why the marginal lead doesn’t buy strategic dominance. “We invent, they copy — a few months later, cheaper” is not a description of safety. It’s a description of diffusion.
“If the margin never matters, why are governments and labs pouring hundreds of billions into it?” Partly because they’re caught in exactly the race dynamic described above — each moves because the others do — and partly for prestige and hedging against uncertainty. Revealed preference shows the race is intense; it doesn’t show the race is wise. Tulip futures were intensely traded too.
“You lean on uncertainty — but uncertainty cuts both ways.” It does, and I’ll own that. “We can’t be sure we’re ahead” can be flipped into “we can’t be sure we’re safe, so we must race.” The reason it still favors restraint: you should not bet the international order, and safety margins, on a decisive lead you cannot even confirm you possess. Deep uncertainty argues against staking everything on the assumption of dominance.
“China won’t actually integrate faster.” Fair — this is my most contestable empirical claim. The US has deeper capital markets, more compute, and a vastly larger software-deployment surface. So let me soften it to what I can defend: there is a real case that China integrates faster, driven by workforce depth and an adoption-first strategy. Whether it wins that race is open. But the fact that the question is live is itself fatal to “our slightly better model keeps us on top.”
A word to anyone who wants to use these arguments
I’ve deliberately laid out more arguments here than any single case should run at once — a menu, not a brief. And a few of these dishes do not belong on the same plate. If you’re going to deploy this, you’ll have to make some choices, because at least one pairing turns on itself.
Here’s the trap, in the plainest terms. One line above says compute restrictions don’t work: starve a lab of processing and it just gets more efficient, so the constraint accelerates capable AI rather than stopping it. Another line says an AI lead — and the frantic racing to hold or extend it — is dangerous. Run both at full strength and you’ve just argued that the restriction speeds up the very thing you called dangerous. That’s a double turn: you’ve linked and impacted your own position into a contradiction, and a sharp opponent will simply staple the two together and hand you the loss.
So pick your lane before you start. Either the lead is illusory and undurable — compute won’t preserve it, efficiency and diffusion erode it, so there’s no real edge worth wrecking the international order to protect. Or holding a runaway lead is affirmatively destabilizing — the edge is real enough, and chasing it is precisely what invites a preventive strike or a safety catastrophe. Both are strong. Both are defensible. Together they collapse. If you’re running the “scarcity just breeds efficiency” argument, don’t also run “racing to build ever-better AI is the core danger” — choose the world you’re arguing in, and stay in it.
The bottom line
The United States is, by the visible evidence, ahead in AI. It should probably want to stay ahead. But “we are marginally ahead” is not a national-security strategy, and “we might become marginally less ahead” is not an existential emergency. Whichever lane you take, the conclusion holds: either the edge is too illusory and undurable to justify the panic, or it’s real enough that racing to extend it is what actually courts disaster — you don’t need both to be true, and you shouldn’t claim both at once. Around that core, the rest stands on its own: the model is not the jet, integration beats capability, open weights beat secrecy for most real uses, and the democracy threat is largely at home.
The margin, in other words, is silly. What isn’t silly is the question of whether we build and deploy this technology wisely — and that question has almost nothing to do with beating China by six months.