Somewhere in an enterprise architecture review this quarter, a CIO is being asked a question that did not exist three years ago: which country’s AI stack should this workload run on. Not which cloud region, which vendor, or which pricing tier. Which country’s stack, governed by whose export licence, subject to whose data law, and reachable within whose timeline.
That question is becoming routine because more than 180 government-backed AI programmes are now underway worldwide, according to the Center for a New American Security’s Sovereign AI Index, spanning compute, models and data infrastructure. Saudi Arabia’s HUMAIN has signed on for 200,000 Nvidia GPUs. The UAE’s G42 runs the Stargate data centre programme with Microsoft as a $1.5 billion anchor investor. France has backed Mistral with state capital through Bpifrance. India has built a Common Compute Facility of more than 38,000 GPUs, offered to startups at roughly ₹65 an hour, with 20,000 more on order. Each government describes this in the language of independence: owning the intelligence layer the way earlier generations of policymakers spoke of owning steel, oil or semiconductors.
The evidence tells a less flattering story, and it is the one CEOs and CTOs actually need to plan around.
What the numbers actually show
The CNAS index finds that the United States and China between them still host roughly 90 percent of the world’s frontier AI computing capacity. Nvidia supplies the chips for 45 percent of all tracked sovereign infrastructure projects. More than three in five of these “sovereign” initiatives involve an American corporate partner sitting somewhere inside the stack, whether that is Nvidia’s silicon, Microsoft’s cloud, or Cerebras’s systems. India’s own Common Compute Facility runs largely on Nvidia hardware procured through a partnership with the domestic cloud provider Yotta. Saudi Arabia’s national champion is, in practice, a Nvidia customer with a 200,000-unit order book and a state balance sheet behind it.
None of this makes these programmes worthless. It makes them something different from what their branding suggests. As the CNAS researchers put it, sovereign AI for most economies means “managing rather than eliminating dependencies, choosing which layers of the stack matter most and accepting reliance on foreign providers for the rest.” That is a precise and useful reframing, and it is the one that should be shaping enterprise strategy, because it describes a negotiated dependency, not a decoupling. Governments are not building alternatives to the American and Chinese AI stacks. They are buying seats closer to the table, positioning themselves within tiers of access rather than outside them.
Those tiers are becoming explicit rather than implicit. Washington’s escalating export controls, including an April ban on Nvidia’s H20 chip that forced a $5.5 billion inventory writedown at the company, have sorted the world into three rough bands: allied nations with largely unrestricted chip access, a much larger group operating under licences and compute caps, and a small set of countries, chiefly China and Russia, cut off entirely. Every government AI programme announced since 2024, whatever its sovereignty rhetoric, has been a response to finding out which of those three bands it sits in and trying to improve its position within it.
Why India’s version of this is instructive, not exceptional
India’s IndiaAI Mission, backed by roughly ₹10,372 crore ($1.2 billion) and expanded through a “Mission 2.0” that adds compute, chip design incentives and semiconductor manufacturing support, is frequently described domestically as a sovereignty project. It is more accurately described as a cost and capability project layered on top of continued hardware dependence. The subsidised GPU-hour pricing lowers the cost of experimentation for startups and researchers who could not otherwise afford frontier compute. The parallel push into chip design, through the Design Linked Incentive scheme that has already produced seven fabricated chips and more than 140 IP cores, builds genuine domestic capability at the design layer even as fabrication and advanced packaging remain elsewhere. And the linguistic and cultural investment behind efforts like BharatGen and Bhashini, aimed at models that work across 22 official languages, targets a layer of the stack that foreign frontier labs have limited commercial incentive to prioritise.
That is a coherent strategy. It is not hardware sovereignty, and treating it as such misreads what India is actually optimising for: cost advantage and linguistic differentiation rather than infrastructure independence. For India’s enterprises and the country’s fast-growing Global Capability Centres, the practical upshot is that competitive advantage will accrue to firms that get good at building AI-enabled products and services on cheap, foreign-underpinned compute, not to firms betting on India controlling its own chip supply chain within this decade. Executives running India operations who assume “local” now means “insulated” are planning against a scenario the data does not support.
What this actually changes in a boardroom
The practical consequence for any multinational is that compute access has become a variable that behaves more like currency exposure than like a procurement line item, and it deserves the same discipline. Export licence approvals for advanced chips are now running 60 to 120 days in many jurisdictions, timelines that can quietly stall a product launch that assumed cloud-like elasticity. The European Union’s AI Act introduces its own complication, with compliance obligations that scale according to training compute measured against a 10^25 FLOP threshold, meaning the jurisdiction in which a model is trained, not merely where it is deployed, now carries regulatory weight. A global enterprise architecture built around a single AI vendor and a single deployment pattern is no longer a simplification; it is a concentration of geopolitical risk that most technology strategies have not yet priced in.
There is a second-order effect worth sitting with, because it cuts against the popular framing of sovereign AI as a democratising force. Fragmenting the global AI stack into tiers, licences and jurisdiction-specific compliance regimes raises the fixed cost of operating globally. That cost is far easier for a handful of hyperscale incumbents and the largest multinationals to absorb than it is for mid-sized challengers or the startups these sovereign programmes are ostensibly designed to help. A policy movement framed as decentralising power over AI may, in practice, be reinforcing the advantage of whoever can already afford to run a compliant AI stack in a dozen jurisdictions at once.
None of this argues against engaging with sovereign AI programmes. Government invitations to anchor a local data centre, co-develop a language model, or participate in a national compute facility are genuine commercial opportunities and, increasingly, a precondition for market access. But they are best understood as instruments of political relationship-building and industrial policy, not as evidence of imminent technological independence, and CEOs who conflate the two will misjudge both the durability of any single country’s AI advantage and the real supply chain risk sitting underneath their own AI roadmap.
The countries racing to build sovereign AI are not, for the most part, racing each other toward independence. They are racing to secure a better position inside a dependency structure that a small number of American and Chinese firms still largely control. The more useful question for a chief executive in 2026 is not which nation will win the AI race. It is whether the enterprise can survive being deployed across a dozen of these dependency structures simultaneously, each with its own licence timeline, compute ceiling and compliance clock.
