The Physical Stack of AI · Sovereign AI compute

Why every country wants its own compute

You can explain why, in the span of about eighteen months, "having our own AI compute" went from a niche industrial-policy idea to a worldwide political consensus — and why none of the declarations on their own actually deliver any chips.

In 2024 you could count the countries with a published sovereign-AI policy on two hands. By mid-2026 more than 90 countries had joined the New Delhi Declaration on AI Impact. The shift wasn't driven by any single event so much as by the realisation, after ChatGPT and then DeepSeek, that the technology behind frontier AI was sitting in a handful of US datacentres and a couple of TSMC fabs — and that the rest of the world had no plan for what to do if Washington decided to turn the valve.

Three things happened in 2025 that turned the topic from policy paper to active procurement. The AI Diffusion Rule was published, then rescinded in May. BIS separately issued GP10 guidance warning that PRC advanced-computing ICs, including Huawei Ascend, remained legally sensitive. And every G7 capital watched the others sign datacentre MOUs they hadn't been invited to. By the time the India AI Impact Summit at Bharat Mandapam closed in February 2026 with the New Delhi Declaration, sovereign AI was no longer aspirational — it was the default frame.

Chapter contains 3 lessons.