A Manhattan Project for AI? Washington isn't taking over the race — it's taking a board seat

WHAT WAS ACTUALLY ANNOUNCED. On October 4, Donald Trump announced the Super Intelligence Force on Truth Social. It is a federal task force led by Director of National Intelligence Jay Clayton. It has three vice chairs: Andrew Ferguson, chair of the FTC (the competition and consumer protection agency); Emil Michael, the Pentagon's chief technology officer; and Scott Kupor, head of the OPM (the federal personnel office). According to NBC News and The Hill, Condoleezza Rice and David Sacks will advise from outside government. Punchbowl News reports that Ferguson may be the one actually running it. The group has 120 days to deliver a report on risks and opportunities. It held its first meeting on October 15 and set three priorities: cybersecurity, biosecurity, and whether the most powerful models can be kept under control.

The task force grows out of two earlier moves. The first is the White House Accord on Super Intelligence, signed in late September by Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Jensen Huang and Elon Musk. It asks for four layers of control: internal monitoring during training, a dedicated safety team, independent external audits, and board-level oversight. It is not a law. Trump called it “morally binding.” The second move is an executive order requiring official communications to call AI “super intelligence.” Several legal analyses say the order changes the name, not which technologies are regulated. To be clear, none of this is nationalization. The government is not buying labs, running them, or taking their models.

WHERE THE IDEA COMES FROM. The phrase “Manhattan Project for AI” has an author. In June 2024, former OpenAI researcher Leopold Aschenbrenner published “Situational Awareness.” In its chapter “The Project,” he argued that by 2027-28 the US national security apparatus would fold the frontier labs into a government effort, willingly or not. That meant large congressional appropriations for compute and power, and teams working in classified facilities. A few months later, in November 2024, the US-China Economic and Security Review Commission (USCC), which advises Congress, made “a Manhattan Project-like program” to reach AGI (artificial general intelligence) the top recommendation of its annual report.

The most serious rebuttal came in March 2025. Dan Hendrycks, Eric Schmidt and Alexandr Wang published “Superintelligence Strategy.” They warned that a government race for a monopoly on superintelligence would push China to sabotage it before it paid off. They called this MAIM (Mutual Assured AI Malfunction): a form of deterrence, similar to the nuclear kind, in which each power can disable the other's project. Their alternative had three legs: deterrence, nonproliferation and competitiveness. Washington has borrowed Aschenbrenner's label, but not his plan. The Genesis Mission, launched in November 2025, was billed as comparable to the Manhattan Project in urgency and ambition. In practice, it is a Department of Energy program that applies AI to science using public supercomputers. It is a science project, not a takeover of the labs.

WHY THE ANALOGY BREAKS DOWN. The Manhattan Project and Apollo were government purchases with a fixed goal: a bomb and a moon landing. The government was the only customer and the only owner of the result. They were also cheap compared with today's AI spending. The Congressional Research Service (CRS) puts the Manhattan Project at about $22 billion in 2008 dollars, spread over five years. In its peak year, it amounted to 0.4% of GDP. Apollo cost about $288 billion in inflation-adjusted dollars over fourteen years, according to the Planetary Society. In 2026 alone, Microsoft, Alphabet, Amazon and Meta expect to spend roughly $700 billion on infrastructure, according to their own guidance to investors. By our rough estimate, that is more than two percent of US GDP in a single year. It is private money, earned back by selling products to hundreds of millions of users.

There is a second, deeper difference. The bomb had no civilian customers. AI is a general-purpose technology, like electricity, and it already runs inside hospitals, banks and schools. Talent is spread across companies, universities and countries, and a large part of the frontier is open: GLM, Qwen and DeepSeek publish their models. A Manhattan Project requires secrecy and control over a scarce resource. AI in 2026 is neither scarce nor secret. That's why the analogy works as a slogan, but not as a plan.

THE QUIET NATIONALIZATION. If there is no expropriation, where does the government's power lie? In pressure. The clearest example is the fight between Anthropic and the Pentagon. In February 2026, the Defense Department labeled Anthropic a “supply chain risk,” a designation normally reserved for vendors from adversary nations. According to press reports, this happened after the two sides failed to agree on the use of AI for mass surveillance and autonomous weapons. Also according to press reports, then-Defense Secretary Pete Hegseth floated the Defense Production Act, which lets the government prioritize industrial orders. The courts are split. In August, a federal judge in California sided with Anthropic and found retaliation in violation of the First Amendment. In September, the DC Circuit ruled 2-1 for the Pentagon, though under a different rule. Emil Michael, now a vice chair of the Super Intelligence Force, maintains that the designation still stands.

The labs' positions are ambiguous. In March, Sam Altman wrote that it had seemed to him “for a long time” that building AGI might be better as a government project. But he added that it doesn't seem “super likely” on the current trajectory, and he called for a “close partnership” with government. Jensen Huang framed the accord around one principle: the companies carry the primary responsibility for safety. And Anthropic signed the White House accord while suing the Pentagon. Our reading: the industry accepts coordination and oversight as long as it doesn't mean giving up ownership. The government, for its part, prefers discretionary tools to a law that would bind it.

OUR TAKE AND WHAT IT MEANS. The government isn't taking over the race; it's taking a board seat. That is wiser than nationalization. A state monopoly would concentrate power even further, slow the technology's spread and, as Hendrycks and Schmidt warn, invite escalation with China. But the current model has a serious flaw: it rests on voluntary pledges and on pressure that can be arbitrary. We have said it in earlier analyses: the danger is regulating panic, or political loyalty, rather than actual capability. That is the objection raised by Democrats such as Rep. Lori Trahan, who calls the task force a “delay tactic” with no laws or guardrails.

Watch for three signals over the next 120 days. First, whether the report proposes concrete legal duties, such as notifying the government of breaches, hacks and jailbreaks (techniques for getting around a model's safety filters); Crowell and other law firms say this is under consideration. Second, whether the accord's external audits rely on independent, non-saturated evaluations, like those run by the UK's AI Security Institute. Third, whether the supply-chain-risk label used against Anthropic becomes a routine tool for disciplining inconvenient labs. In the short term, the danger is not a runaway superintelligence. It is opaque, personality-driven governance over a technology that will soon affect millions of jobs.

Over the long term, we remain optimistic. The good version of public involvement already exists on a small scale: the Genesis Mission puts public supercomputers to work for science. That is the road to new drugs, cured diseases and abundance. The government adds more value by funding shared scientific infrastructure and demanding verifiable transparency than by trying to own the race. The Manhattan Project built a weapon for a single owner. AI can serve everyone, if governance is built on evidence rather than slogans.