Geopolitics

For years the AI labs talked about geopolitics the way diplomats talk about the weather: a thing that happens, mostly elsewhere, to other people. That posture is dead. By early 2026 the founders of the frontier labs are no longer commentators on great-power competition. They are instruments of it, drafted whether they like it or not, and the most revealing fact about the field is how differently each of them is handling the conscription. The cleanest split runs straight through one question that used to be theoretical: when the United States government asks for everything, what do you refuse?

The Pentagon decided the labs work for it now

The event that crystallized everything was Anthropic's public break with the Pentagon. In a CBS News interview (Feb 2026), Dario Amodei laid out a position that is both maximally patriotic and maximally defiant. Anthropic, he insisted, "has been the most lean forward of all the AI companies" with the US military: first to put models on the classified cloud, first to build custom national-security models, "deployed across the intelligence community and the military." The motivation is unsentimental. "I believe we have to defend our country from autocratic adversaries like China and like Russia."

And yet Anthropic drew two red lines: domestic mass surveillance and fully autonomous weapons. Amodei's worry about the first is precise and unusually honest about the legal vacuum. Buying bulk private data on Americans and analyzing it en masse "actually isn't illegal... it was just never useful before the era of AI." The technology, he said, is "getting ahead of the law." On weapons that fire "without any human involvement," his objection is partly engineering: "the AI systems of today are nowhere near reliable enough."

The government's response was an ultimatum. A three-day window, the threat of a "supply chain risk" designation "normally used against foreign adversaries," and finally a Trump post accusing Anthropic of "selfishness... putting American lives at risk." Amodei's counter is the most interesting tell in the whole dataset: a private company "can choose to sell or not sell whatever we want," but "in the long run... it is Congress's job" to set these rules. He is simultaneously claiming the right to refuse and admitting he shouldn't be the one deciding. That tension, a CEO acting as a constitutional backstop he doesn't think should exist, is the live wire under the entire field.

China is the organizing fiction, and not everyone buys it

The hawkish case is most developed not by Dario but by Daniela Amodei, in Anthropic's submission on the Commerce Department's AI diffusion rule (Anthropic, 2025). The argument is pure realpolitik: "maintaining America's compute advantage through export controls is essential for national security." The numbers are the weapon. Compute doubling every two years means that "by 2027, countries using older chips could face AI training costs that are ten times higher." DeepSeek is Exhibit A for controls, not against them: Chinese labs "openly acknowledge that chip restrictions are their primary constraint," needing "2-4x more power" for similar results. And the historical analogy is the offshoring of solar panels and batteries, with the US share of semiconductor production having "fallen from 40% in 1990 to just 12% today."

Then Demis Hassabis walks in and quietly torches the premise. Asked about "the AI race and the competition with China" (FT, Jan 2026), he says the framing barely survives contact with the actual country: "from what I can see in China there is no AI race... there's no sort of race to reach AGI. There is a lot more focus on applications and finding efficiencies." This is not a minor disagreement. The export-control regime, the Pentagon's urgency, the whole "autocratic adversaries" frame, all of it rests on the assumption of a symmetric sprint toward superintelligence. DeepMind's CEO is on record saying he doesn't see the other runner.

Soft power is the other arms race, and it runs through India

While Washington and Anthropic fought, the labs were quietly running a parallel campaign that looks less like a Cold War and more like a scramble for influence. India is the prize. Hassabis announced DeepMind's "National Partnerships for AI" (DeepMind, Feb 2026), explicitly building on collaborations with "the US and UK governments," now extended to Indian government bodies as the country convened "the fourth global AI summit." The figures are designed to embed: AlphaFold already has "over 180,000 researchers" in India, plus a $30M AI for Science challenge and a deal to turn "two million static textbooks" into Gemini-powered tutors. This is statecraft dressed as philanthropy, and it knows it.

Sam Altman is playing the same board with more swagger. In India (Indian Express, Feb 2026) he urged the country to "play at all of those levels" of the AI stack, from chips to applications, citing vertical integration as a national advantage and name-dropping a conversation with the prime minister. The pitch is abundance: he is "not a jobs doomer," every technological revolution "found new jobs on the other side," and the real constraint is compute. His provocation, that if everyone wants even a thousand GPUs "working for them all the time," you'd need "8 trillion GPUs" and "no way to deliver" them, reframes geopolitics as an infrastructure race. (He also flatly called data centers in space "ridiculous" at current launch costs, a rare puncture of his own genre's hype.)

The contrast with Mira Murati's older framing (Wired, 2020) shows how far the conversation has migrated. Back then the worry was data dignity, facial-recognition bias, and AI as a democratized "platform" like the steam engine. Distribution, not dominance. Six years on, distribution has become a tool of dominance, and the steam engine has a flag on it.

Whose hand is on the chip, and whose donation is in the pocket

The hardware layer is where geopolitics stops being rhetorical. Buried in The Information's January coverage is a fact stated almost in passing: the US government now owns a stake in Intel, making the company "the national champion" and complicating every story about it as "an independent foundry" competing with TSMC (The Information, Jan 2026). The same broadcast was teased around "OpenAI's Greg Brockman Goes MAGA," its "big read" examining his political donations as a signal of "Silicon Valley and politics at large." The labs aren't just being regulated by the state. Their executives are funding it, and the state is buying their suppliers.

Jack Clark supplies the dark undertow on what state adoption actually looks like. National-security uses of AI "will happen quickly," he told Tyler Cowen (Feb 2026), because governments "desperately want growth, and they desperately want efficiency." But his real fear is political backlash: a "political movement... which tries to freeze a load of human jobs in bureaucratic amber," driven "by the chaotic winds of political forces" rather than reason. Geopolitics, in his telling, doesn't end at the border. It comes home as domestic populism over who gets automated.

The kicker

Strip away the export-control spreadsheets and the India summits and you're left with one uncomfortable structure: a handful of private firms now hold capabilities that states want and laws haven't caught up to, and the firms know it. Amodei says Congress should decide. Hassabis says there isn't even a race. Daniela Amodei says the race is everything. They can't all be right, and the only thing they agree on is that none of them asked to be holding the gun.