Everyone Is Trying to Own AI Now

The convert nobody expected is Washington. On October 5 the Pentagon banned Anthropic — the company refused to remove its safety guardrails and got switched off in return — and launched GenAI.mil, a single military AI platform serving more than three million personnel ^[1]. Beside it sits the Energy Department's plan to purchase and own up to ten nuclear reactors outright ^[2]. Carl Coe, the department's chief of staff, put the doctrine in Washington's own words.

But this is a national emergency. — Carl Coe

The state that taught the world access is revocable has begun acting like a tenant on notice. The same doctrine now appears in nearly identical form across every class of actor. In May, China certified nine domestic AI chips — from Huawei, Alibaba, Biren and four others — as a standalone category under its national security framework, a procurement list steering state agencies away from imported silicon ^[3]. In December, India published a white paper declaring compute and models a "shared national resource" to be kept out of the hands of "a few global firms and urban centres" ^[4]. In July, Japan opened a national AI factory to avoid becoming an "AI colony" ^[5]. Britain signed a £31 billion tech deal and announced an Nvidia/OpenAI campus it calls sovereign ^[6], even as the Alan Turing Institute's director warned that allied access could be restricted or switched off at any time ^[7]. Microsoft, one of the biggest buyers of rented compute, unveiled its Maia 300 chip with the stated aim of cutting its Nvidia reliance ^[8]. The commercial version of the same logic is growing: analysts expect roughly 75% growth for the specialized AI clouds CoreWeave and Cerebras ^[9]. And at least a dozen nations have now committed hundreds of billions to sovereign AI, with the UAE planning to seat its national AI system as a cabinet advisor ^[10]. What is striking is that the actors do not all name the same teacher. India's IT minister ties the indigenous-model push to the precedent of Washington cutting off foreign access to Anthropic's models ^[4]. Britain's institute director cites the same cutoff ^[7]. China's answer extends a long state drive to swap imported technology for domestic into AI itself ^[3]. The Pentagon cites a company's refusal to strip its guardrails ^[1]. Same doctrine, no single trigger. Jensen Huang saw this coming in June 2025.

Shielding Chinese chipmakers from US competition only strengthens them abroad and weakens America’s position. — Jensen Huang

The race he described is now the whole market's logic. Except almost none of it has actually been built. Japan's "sovereign" factory runs on 27,500 Nvidia GPUs bought from the very vendor it was meant to stop depending on ^[5]. Huawei's flagship chip, projected to bring in roughly $12 billion this year, was assembled around nearly three million TSMC dies smuggled through a shell company ^[11][12]. Microsoft's Maia 300 is a 2027 inference chip, and the company still partners with Nvidia and AMD on training ^[8]. The reactors are still a plan ^[2]. Satellite analysis suggests around 40% of U.S. data-center projects may miss their 2026 dates by more than three months, though OpenAI, Oracle and SB Energy all deny it ^[13]. At the bottom sits Kenya: a $1 billion Microsoft-G42 data center, the imported version of sovereignty, stalled before construction over the purchase commitments it could not meet. President William Ruto put the arithmetic plainly.

A project of this size would strain Kenya's energy grid, saying the country would need to switch off half the country to power it. — William Ruto

Washington's own entry reads the same way. The ban, the platform launch, the reactor purchase plan, even proposals to nationalize OpenAI and Anthropic if they falter ^[14] — the most decisive-sounding moves in the race, and every one of them still an announcement. For now, the ownership race has produced a planet of buyers holding paper: white papers, procurement lists, purchase plans, and chips still built on someone else's dies.