We Wouldn’t Have Paused the Manhattan Project. Why Pause AI?
A two-track answer to the AI race: frontier capability built and tested inside a secure government proving ground, and commercial AI kept capability-limited for the public.
Imagine it is 1942. The United States, worried about the moral weight of the atomic bomb, decides to slow its nuclear research. Meanwhile Nazi Germany keeps going at full speed. Now imagine we go one step further and ship Germany the uranium, politely asking it not to build a bomb.
No serious person would have accepted that policy then. Yet something much like it is being debated today, with artificial intelligence as the weapon and China as the rival.
The race is real
AI is not just another consumer technology. It will shape intelligence analysis, cyber offense and defense, autonomous systems, and the speed at which a nation can make decisions in a crisis. Whoever leads in AI will hold an advantage comparable to what nuclear weapons gave the United States after 1945.
Beijing understands this. Its leadership has said openly that it intends to be the world’s AI leader, and it treats that goal as a matter of national power rather than market share.
Calls in America to slow down, pause, or heavily restrict frontier AI development may be well intentioned. But restraint only works when both sides practice it. A unilateral slowdown does not make the world safer. It simply decides who gets there first.
We are selling the raw material
Advanced AI runs on advanced chips, and the best chips in the world are American designs. On 13 January 2026 the Commerce Department revised its licence review policy for the Nvidia H200, the AMD MI325X and similar chips, moving them to case-by-case approval for export to China provided certain security conditions are met. Roughly ten Chinese firms were subsequently cleared to buy the H200 — Alibaba, Tencent, ByteDance and JD.com among them — each capped at 75,000 units.
The scale is what matters. The Council on Foreign Relations estimated that around a million H200s shipped to China would raise China’s installed AI compute in 2026 by roughly 250 percent compared with relying on domestic chips alone.
There is a wrinkle worth being honest about, because it cuts against the simplest version of this argument. As of August 2026 almost none of those chips had moved. ByteDance and Tencent had taken about 10,000 units each, roughly 13 percent of their ceiling, because Beijing now gates the purchases itself: Chinese firms must justify to their own regulator why a domestic alternative will not do.
Beijing is exercising the restraint we are not. It is doing so to build its own industry, not to slow the race.
That is not reassurance. It is a competitor declining a dependency while it closes the gap, which is exactly what a serious rival does. The license ceiling remains, and policy that can be unused this quarter can be used next quarter.
Where controls do exist, they leak. In March 2026 federal prosecutors charged a co-founder of Super Micro and two others over the diversion of roughly $2.5 billion in Nvidia-powered servers to China, routed through the company’s Taiwan operations and a Southeast Asian intermediary designated as the end buyer. One of the three has since pleaded guilty; Liaw pleaded not guilty and his trial is now set for March 2027. Super Micro itself was not charged and says it was a victim.
And the regime governs objects, not access. Export controls cover physical chips; they do not clearly cover renting the same chips remotely, which is why Chinese labs have used cloud capacity abroad. In January 2026 the House passed the Remote Access Security Act 369 to 22 precisely to close that gap — a vote that tells you the loophole is real and that it was still open when they voted.
That gap is no longer theoretical. Megaspeed, a Singapore company, bought more than $4.8 billion of Nvidia chips over roughly two years and installed close to $2 billion of them in data centers in Malaysia and Indonesia, where the hardware sits lawfully. US investigators allege those machines served Chinese customers remotely, and that a Megaspeed executive used shell companies to reach 240,000 chips barred from export to China. The company denies wrongdoing and says it has cooperated with Commerce inspections. Notice what is not alleged: that anyone carried a chip across a border. Nobody needed to.
The corporate paperwork is its own lesson in how ordinary this looks. Megaspeed was incorporated in 2015 as the subsidiary of a Shenzhen game developer, acquired in 2023 by a British Virgin Islands shell, and renamed. Reporters who visited its Singapore headquarters found the door locked or the office nearly empty. Its Australian subsidiary is registered to a floor shared with forty other businesses, under a director who turns out to be a 29-year-old newspaper vendor four hundred miles away whose father had never heard of the company. One Malaysian address is a dilapidated storefront. None of this is exotic tradecraft. It is a company that looks like a company until somebody goes to the address.
On 12 November Nvidia announced that Jensen Huang had resigned as president, chief executive and director, effective immediately. The company gave no reason. The resignation followed his admission of an affair and reporting that the woman involved, a former Megaspeed executive, was assessed by US authorities to have been recruited by Chinese intelligence; Nvidia’s board has said it has no reason to believe national security was compromised, and no evidence has been established that he knowingly assisted Beijing. Those investigations will take their course. The structural point does not wait on them: the company that supplies the world’s AI compute lost its chief executive to a counterintelligence problem, and the chips at the center of the story never had to leave the countries where they were legally installed.
Would they use it?
Does anyone doubt that Hitler, given the bomb first, would have used it to dominate Europe and crush the Soviet Union? Authoritarian regimes do not leave decisive weapons on the shelf.
China’s government already uses AI for mass surveillance at home and has been explicit about its ambitions toward Taiwan. It would be naive to assume a Chinese lead in military AI would be used gently.
Full speed, with the guardrails in the right place
The answer is not reckless acceleration. It is a two-track system.
The first track is national security AI. American companies should push frontier capability as fast as possible inside secure, controlled environments that simulate what China could actually launch: cyber intrusion against the power grid, disinformation at scale, attacks on financial systems, and AI-driven offensive operations. These systems get built, tested and hardened to detect and defeat adversary AI in a real conflict.
The second track is commercial AI for the public, which stays regulated and capability-limited. We already understand this model. Americans have a constitutional right to own handguns and rifles; no one walks into a store and buys an RPG. Powerful, useful tools for businesses and citizens, with the most dangerous capabilities reserved for national defense.
Alongside both tracks, chip export policy should be tightened rather than loosened, the remote-access loophole closed, and smuggling networks prosecuted aggressively.
A blueprint: the Secure AI Proving Ground
The first track needs a concrete home: a government-sponsored, privately operated AI lab, run by two of America’s largest AI companies, located on a protected military installation.
The design is simple. One company plays the United States. The other plays the adversary. Their models are pitted against each other in continuous, realistic simulation — one side launching cyberattacks, influence operations and assaults on critical infrastructure, the other detecting, defending and counterattacking. Every exercise exposes a weakness, and every weakness becomes the next thing to fix.
The companies switch roles on a regular rotation. The team defending the country this quarter attacks it the next. Each builds genuine expertise in both halves of the problem, neither side’s blind spots go untested, and no single company becomes the permanent keeper of America’s defensive playbook.
The facility is closed. No Wi-Fi, no phones, no modems, no connection of any kind to the public internet, with the physical security and personnel vetting we apply to our most sensitive defense programs. That isolation defends against the exact threat the lab exists to study. It is also the part most likely to be traded away for convenience, and it is the part that cannot be.
The engine of the whole arrangement is competition. Neither company wants to lose to its rival, and that pressure drives progress faster than any mandate.
Who pays, and who owns it
The companies would be compensated two ways. A government contract covering the cost of running the lab and rewarding performance. And the right to release selected advances to their commercial divisions once reviewed and cleared as safe for public use — the same path that took GPS and the internet from military programs into ordinary life.
The top-tier capabilities, the ones too dangerous for public release, would be owned by the United States government. Ownership should not mean unchecked control. Oversight rests on three independent legs: the two companies, who understand these systems better than anyone and would have the access to confirm they are aimed outward and never at Americans; the House and Senate intelligence committees, with regular classified briefings; and an independent inspector general empowered to audit the program and report violations directly to Congress.
No single party holds all the keys. The government holds the weapon, the builders keep it pointed outward, and Congress and the inspector general keep everyone honest, the companies included.
The public never sees the top tier, just as civilians never buy military-grade weapons. What the public gets is the downstream benefit: better commercial AI, stronger cyber defense, hardened infrastructure, and the knowledge that if an AI-driven attack ever comes, American systems have already fought that battle a thousand times.
Answering the critics
The nuclear analogy fails because Germany was far behind. True in hindsight, and only in hindsight. In 1942 nobody in Washington knew how close Germany was, or that the war would end in 1945. Germany had world-class physicists. The Manhattan Project was not built on certainty that Germany was ahead; it was built on the understanding that being second was unacceptable. We do not need proof that China is winning to justify racing. We need only recognize what losing would cost.
Selling chips to China keeps Beijing dependent on American technology. That is short-term thinking, and the last six months undercut it directly: offered the chips, Beijing told its champions to justify why they could not buy domestic instead. Every chip sold buys time for China’s own industry. And the most advanced manufacturing still sits on Taiwan, an island Beijing has vowed to bring under its control. If that changes, the question of who depends on whom flips overnight.
Nobody in Washington is actually pushing a slowdown. Today, largely true. But policy does not stand still, the next presidential election is two years out, and a future administration could impose restrictions that slow American labs while China’s keep running. The time to build a durable framework for secure, full-speed development is while the question is still hypothetical.
Win the race, control what we build
I argued in AI Is Nuclear in Reverse that AI inverted the nuclear sequence — arriving as a consumer product and becoming a weapon system afterward — and that the right posture is to race and harden rather than to race or pause. The Secure AI Proving Ground is what hardening looks like when you give it an address, a budget and a chain of oversight.
The lesson of the 1940s is simple. When a hostile power is racing toward a world-changing capability, the responsible choice is to win the race and control what you build. We should not slow down while our rival speeds up. And we certainly should not hand them the fuel.
Sources
- Bureau of Industry and Security, “Department of Commerce Revises License Review Policy for Semiconductors Exported to China” (13 January 2026) — case-by-case review for the Nvidia H200, AMD MI325X and similar chips, subject to security conditions. bis.gov
- The Next Web, “US clears H200 sales to 10 Chinese firms, but not a single chip has shipped” (14 May 2026) — roughly ten approved buyers including Alibaba, Tencent, ByteDance and JD.com, each capped at 75,000 units. thenextweb.com
- Tech Times, “Nvidia H200 Chips Enter China at 13% of Quota” (20 August 2026) — ByteDance and Tencent took about 10,000 units each, with the remainder gated by China’s own regulator. techtimes.com
- Council on Foreign Relations, “The New AI Chip Export Policy to China: Strategically Incoherent and Unenforceable” — roughly a million H200s would raise China’s 2026 installed AI compute by about 250 percent. cfr.org
- Fortune, “Supermicro cofounder arrested over $2.5 billion in GPU smuggling” (19 March 2026) — charges against Yih-Shyan “Wally” Liaw and two others; servers routed via Taiwan and a Southeast Asian designated end buyer; the Taiwan general manager at large. fortune.com
- CBS News, “Megaspeed’s vacant offices raise questions about ties to Nvidia and China” (13 November 2026) — Singapore headquarters, the Sydney subsidiary and its listed director, the Malaysian addresses, the 2015 incorporation as a Shenzhen game developer’s subsidiary and the 2023 acquisition by a British Virgin Islands shell; US investigators’ allegation regarding 240,000 chips barred from export; Megaspeed’s statement that it takes its compliance obligations seriously and has cooperated with Commerce inspections. cbsnews.com
- Fortune, “2 scandals, 1 destination — Why Nvidia can’t shake China smuggling allegations” (10 November 2026) — more than $4.8 billion in Nvidia chips bought by Megaspeed, with nearly $2 billion installed in Malaysian and Indonesian data centers alleged to serve Chinese customers remotely; the current status of the Super Micro prosecution. Fortune states there is no evidence the two cases are directly connected. fortune.com
- NVIDIA, “NVIDIA Announces CEO Transition” (12 November 2026) — Jensen Huang “has resigned as President, Chief Executive Officer, and a member of the Board, effective immediately”; Colette Kress named interim CEO. The release gives no reason for the resignation. nvidianews.nvidia.com
- The Register, “Jensen Huang resigns as Nvidia chief executive” (12 November 2026) — the board accepted the resignation on 11 November per the Form 8-K; the admitted affair, and the reporting that the woman involved was a former Megaspeed executive assessed to have been recruited by Chinese intelligence. Reported as allegations. theregister.com
- Al Jazeera, “Why did Nvidia CEO Jensen Huang resign, and who is Colette Kress?” (12 November 2026) — Nvidia gave no reason for the departure; the board has said it has no reason to believe national security was compromised. aljazeera.com
- The Register, “Congress votes to kick China off remote GPU services” (13 January 2026) — the House passed the Remote Access Security Act 369–22 to extend export controls to cloud access to controlled chips. theregister.com