Tech & AI

America's Proposed AI Manhattan Project: Beat China, Staffed by China

By Rick Renko, editorPublished Aug. 24, 20265 min readSources checked

MMG editorial graphic depicting an AI research compound between abstract United States and China motifs. — MMG editorial graphic. This is an illustration, not documentary imagery of a government facility.

The AI arms race has a Manhattan Project recommendation, a deterrence doctrine named MAIM, and a personnel problem nobody can slogan away.

The idea is real. In 2024, the bipartisan U.S.-China Economic and Security Review Commission recommended that Congress establish and fund a Manhattan Project-like program to race toward artificial general intelligence. But there is an important distinction: Washington has not ordered a secret desert superintelligence lab into existence.

The desert compound comes from Superintelligence Strategy, a 2025 national-security paper by AI researcher Dan Hendrycks, former Google CEO Eric Schmidt and Scale AI founder Alexandr Wang. The paper describes — and mostly warns about — what a monopoly-seeking American AI project could look like: invoke the Defense Production Act, concentrate scarce chips in a secured U.S. facility, and bring top researchers together to push for superintelligence before China gets there.

Then comes the awkward part. The authors note that a large fraction of the researchers such a project would need would necessarily be Chinese nationals, because Chinese researchers make up a significant share of frontier AI talent in the United States. That is not the same thing as saying those researchers are spies. In fact, the authors' own argument is that simply excluding Chinese nationals could kneecap American competitiveness.

So the plan to out-China China may require Chinese citizens working in America. Somewhere, a slogan writer has developed a stress headache.

The deterrence doctrine that rhymes with “maim”

Cold War I had Mutually Assured Destruction — MAD, a name so on-the-nose it became shorthand for the whole era's insanity. The same Superintelligence Strategy paper proposes a modern cousin: Mutual Assured AI Malfunction. MAIM.

The basic idea is that if one superpower makes an aggressive bid for a destabilizing AI monopoly, rivals deter it by making clear they can sabotage the project first — through cyberattacks on data centers and, in the paper's most extreme scenarios, potentially physical attacks. It is Cold War logic wearing a hoodie, except the acronym somehow got less reassuring.

MAIM is not official U.S. doctrine. But it is not a random forum post either: its authors include Schmidt, Hendrycks and Wang, and the concept is now part of the serious policy conversation around superintelligence.

Meanwhile, Washington is purging Anthropic

If the AI arms race were coherent, at least the theater would be satisfying. It isn't.

In July, the Air Force Research Laboratory told contractors to inventory their use of Anthropic products and remove them from defense-contract systems by Sept. 1 — ahead of a broader Pentagon deadline at the end of September. The move followed the Pentagon's earlier clash with Anthropic over safeguards involving domestic surveillance and autonomous weapons.

So, during the same year Washington is treating frontier AI as a strategic national asset, parts of the defense establishment are also forcing one of America's frontier AI companies out of their systems. Geopolitics, apparently, now includes enterprise software migration.

The chip policy is similarly awkward. The U.S. still blocks China's access to some of Nvidia's newest hardware, but it has also licensed limited sales of powerful H200 chips to selected Chinese buyers. By July, shipments had begun. Whether that is contradiction or calibration depends on which memo you are holding.

China is winning the half of the race you can put on a factory floor

While Washington argues about models and vendors, China is scaling hardware.

Reuters reported that Chinese manufacturers delivered more than 40,000 humanoid robots in the first half of 2026, about 97% of global deliveries. That does not mean 40,000 useful robot workers are replacing people tomorrow — plenty are still research machines, demonstrations or early deployments — but the manufacturing advantage is no longer theoretical.

The intellectual-property numbers point the same direction. Morgan Stanley counted 7,705 Chinese humanoid-robot patents over five years, versus 1,561 in the United States: roughly five to one. A separate LexisNexis analysis adds an important caveat: U.S. patents tend to score higher on average quality. China has more tickets in the raffle. America may still have some of the better tickets.

And the software gap is getting cheaper

America still has the edge at the frontier. Bloomberg's August analysis found the best U.S. systems continuing to lead overall capability benchmarks. But the gap is narrowing, and price is changing the contest.

In a Bloomberg coding test published Aug. 20, Chinese models from Moonshot AI and Z.ai produced passable websites at roughly a 75% discount to Anthropic's Claude. That is a narrow test, not proof that Chinese models are 75% as good at everything. It does show why “who has the smartest model?” is becoming only half the question.

The other half is: who can afford to deploy acceptable intelligence everywhere?

What “superpower” is quietly turning into

For eighty years, superpower status ran on a fairly stable formula: nuclear stockpiles, GDP, blue-water navies and reserve currency. AI and robotics are adding two new terms: who controls the systems that do the thinking, and who can manufacture the machines that do the doing.

The United States still leads the first category. China has built a frighteningly large advantage in the second. Both are trying to close the other's gap.

The strangest part is that the strategy papers are much better at describing how to stop the other side from winning than what a stable victory would actually look like. The deterrence framework is called MAIM. The hypothetical monopoly project may depend heavily on Chinese nationals. One government is simultaneously restricting Chinese access to AI hardware and licensing some of it back in.

That is not proof the United States is losing. It is proof the race is messier than the “America versus China” scoreboard suggests — and the winner may be whichever country can combine frontier intelligence, cheap deployment and physical scale before the other one stops arguing with itself.

Send this to the person who thinks the AI race has a clean scoreboard.

Sources & verification

Article evidence and links checked Aug. 24, 2026. The “desert compound” is identified as a scenario described in Superintelligence Strategy, not as an already-approved government construction plan. Editorial process →

Rick Renko · Editor, Mulholand Material Generalists

MMG is an independent, one-person publication based in Northwest Arkansas. Every article is researched against primary sources where they exist, drafted with the help of AI research and writing tools, and reviewed, edited and approved by a human before it is published. More about how this works · Editorial process · Send a correction