In 2024, sovereign AI was an aspiration mentioned in conference keynotes. By 2026 it is a budget line in most of the G20. The UAE has a national AI company. So does Saudi Arabia. So does France, India, Japan, Qatar, and more than twenty other governments now hold direct partnerships with Nvidia. But the US and China still control 90% of the compute and own every one of the top 50 frontier models. This is the trajectory of what "sovereignty" actually buys when the underlying chips still come from one country.
The Center for a New American Security's own Sovereign AI Index states the concentration plainly: the United States and China control 90% of the computing power needed to develop and deploy frontier AI, and own all 50 of the top-ranked AI foundation models between them. Every other government pursuing "sovereign AI" in 2026 is, in practice, negotiating for a slice of that concentrated stack — not replacing it.
The index tracks more than 130 national infrastructure and model projects worldwide. Roughly 70% involve at least one foreign technology partner, and four-fifths of those partners are American. The United Arab Emirates and Japan alone account for over two-thirds of total disclosed sovereign AI investment globally — meaning even the "sovereign" race itself is concentrated among a handful of capital-rich states, not evenly distributed across the G20.
What sovereignty actually means in 2026 is a portfolio choice, not a binary. TechPolicy.Press frames it precisely: full-stack sovereignty (owning chips, cloud, models, and data end to end) is politically attractive but capital-intensive and structurally exposed at the semiconductor layer no country outside Taiwan and the US genuinely controls. Compute sovereignty — controlling infrastructure for sensitive workloads while still relying on global foundation models for everything else — is the far more common, far more achievable version actually being built.
Sovereignty in AI needs managed interdependence, not isolation — the lesson every serious national programme has converged on by 2026, whether or not the political rhetoric around it admits it.
HUMAIN was launched under Saudi Arabia's Public Investment Fund in May 2025 with the most complete mandate of any sovereign AI programme: data centers, cloud infrastructure, models, and applications, all under one state-backed entity. The centrepiece is the Hexagon facility in Riyadh — a $2.7 billion, 480-megawatt Tier IV data center — alongside a 200-megawatt Qualcomm inference partnership and a joint venture with telecom operator STC.
The UAE's parallel bet is Stargate — a 5-gigawatt compute campus in Abu Dhabi backed by G42, OpenAI, Oracle, and Nvidia, with a November 2025 US authorisation letting G42 purchase up to 35,000 Nvidia Blackwell chips under a novel "Intergovernmental Assurance Agreement" — the structural innovation that let Washington relax export controls for a foreign counterparty whose ownership could be formally audited. This is the model other Gulf and Asian states are now negotiating toward: not independence from US chips, but an audited, treaty-like relationship that earns access to them.
India's approach runs on two tracks simultaneously: national foundation models and raw compute expansion. Sarvam AI, the state-anchored model effort, launched Sarvam-30B and Sarvam-105B at the February 2026 India AI Impact Summit. In parallel, the IndiaAI Mission's compute facility is scaling from 38,000 to 100,000 public GPUs by December 2026 — still entirely built on American chip supply.
The most instructive deal in the entire sovereign AI landscape may be the one India signed with the UAE in February 2026: Emirati state-owned G42, working with Cerebras and the Mohamed Bin Zayed University of AI, agreed to fund and build an 8-exaflop supercomputer inside India — Emirati capital, American chips, deployed on Indian soil. Sovereignty, in this deal, means capturing the economic and strategic value of hosting and operating the infrastructure, not owning every layer of its supply chain.
Not every sovereign AI programme is chasing raw compute scale. France's Mistral positions itself explicitly as a publicly-supported, open-source European alternative to US labs — paired with a genuinely unusual hedge: a $30-50 billion joint 1-gigawatt data center with the UAE, introducing "virtual data embassy" concepts that let French government workloads run on foreign soil under sovereign legal control.
The UAE's Falcon and Jais models, and India's Sarvam and BharatGen, share a different rationale: building foundation models trained specifically for Arabic and Indian-language contexts that no US lab has prioritised at the same depth. This is sovereignty as cultural and linguistic self-determination as much as economic strategy — a genuinely distinct rationale from the compute race playing out in parallel.
This publication has covered the data center buildout (Letter 134), the memory bottleneck behind it (Letter 144), and the trust economy AI abundance creates (Letter 147) — all of which implicitly assumed a small number of US labs would keep supplying the models running on that infrastructure. Sovereign AI complicates that assumption directly: more than twenty governments have now signed direct compute partnerships with Nvidia alone, and the customer base for AI infrastructure is fragmenting from a handful of hyperscalers into a handful of hyperscalers plus dozens of sovereign buyers, each with its own procurement rules, data-residency requirements, and geopolitical constraints.
For infrastructure investors, this is a genuine demand tailwind — sovereign buyers are, on current evidence, willing to pay a premium for auditable, jurisdictionally-controlled compute. For model-layer investors, it is a genuine competitive threat — Mistral, Falcon, and Sarvam are real, funded, improving alternatives to the default assumption that US labs win every market by default.
Most "sovereign" AI today is sovereignty of convenience, not independence. Roughly 70% of tracked national programmes still depend on a foreign — overwhelmingly American — technology partner. Calling this sovereignty is not wrong, but it is a specific, narrower kind: control over deployment, data residency, and procurement, not control over the underlying chip and model supply chain.
The UAE and Japan's disclosed spending dwarfs almost everyone else's. Most other "sovereign AI" national programmes involve a few hundred million dollars, not the tens of billions Gulf states are committing. Treat headline national AI announcements skeptically until capital commitment and actual GPU delivery are disclosed.
The Intergovernmental Assurance Agreement structure — pioneered for UAE-US chip access — is the template to watch. If it scales to more countries, it becomes the mechanism by which Washington keeps effective control over frontier compute even as it appears to "distribute" sovereignty. If it doesn't scale, expect friction as more nations seek chip access without accepting the same audit terms.
The language-and-culture rationale (Mistral, Falcon, Sarvam) is the most durable version of this trajectory. Unlike the pure infrastructure race, a genuinely better Arabic or Hindi-context model is a real, defensible product advantage that raw compute scale alone doesn't erase — this is where sovereign AI programmes are most likely to produce lasting, exportable value rather than a one-time infrastructure bill.
Part of an ongoing journal — observations recorded when something in the world economy is worth saying. No schedule. No noise. Not investment advice.