Sovereign AI is a procurement strategy, not independence

The number that reframes the whole category

Start with the CNAS Sovereign AI Index, which tracks government-backed programmes across compute, models and data. It has now passed 180 programmes worldwide. The headline figure most coverage skips is the concentration: the United States and China between them host roughly 90% of global frontier AI computing capacity. That single line reframes everything below it.

A sovereign programme is not, and cannot be, an exit from that concentration at frontier scale. What it can be is a negotiated position within it. As the CNAS researchers put it directly, sovereignty 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."

Three bands, not two camps

Washington's export controls have sorted the world into three rough tiers rather than a binary split:

• Allied nations with largely unrestricted chip access.
• A much larger group operating under licences and compute caps.
• A small set — chiefly China and Russia — cut off from frontier hardware entirely.

Every government AI programme announced since 2024, whatever sovereignty language it uses, has been a response to discovering which band it sits in and trying to climb. That is why the announcements look so similar: they are bids, not blueprints.

The deals, read honestly

Strip the branding and the pattern is consistent. Saudi Arabia's national champion, HUMAIN, is in practice a Nvidia customer with a large order book and a state balance sheet behind it — up to 600,000 GPUs planned over three years, with xAI as the phase-one anchor tenant and AWS deploying up to 150,000 GPUs in a Riyadh zone. Stargate UAE is a first phase of 200MW inside a much larger campus, developed with OpenAI, Oracle, Nvidia, Cisco and SoftBank. A government wrote the cheque; a lab fills the racks.

India's Common Compute Facility — over 38,000 GPUs offered to startups at roughly ₹65 an hour, with 20,000 more on order — is frequently described domestically as sovereignty. It is more accurately a cost and capability project layered on top of continued hardware dependence, with genuine domestic capability built at the design layer (seven fabricated chips through the Design Linked Incentive scheme) while fabrication and advanced packaging stay elsewhere.

And Europe made the one procurement that genuinely cuts against the grain. In September 2026 the EuroHPC JU signed a €387.8 million (~$451M) contract with Bull for LUMI-AI — choosing AMD's Instinct MI430X over Nvidia's Vera Rubin platform, largely because of a hardware FP64 figure: roughly 288 teraFLOPS against Nvidia's ~33, nearly nine times, for scientific workloads where double-precision is a correctness requirement rather than a preference. Bull's BXI interconnect replaces NVLink. That is a sovereignty argument in its own right: a workflow built on NVLink cannot be migrated off Nvidia without rebuilding the communication fabric from scratch.

What this means for where compute gets bought

Two conclusions follow, and they point in opposite directions for a decentralised network.

The bad news: frontier training is not the market. Every procurement above is anchored on a named hyperscaler or a frontier lab. No sovereign programme is buying preemptible capacity to train a frontier model. If your pitch is "sovereign AI runs on us", you are aiming at the one segment that is structurally closed.

The good news: the tier below frontier is enormous and price-sensitive. India's subsidised GPU-hour pricing exists precisely because the bottleneck for its startups is cost of experimentation, not access to a B300. Sovereign programmes create three things a decentralised network can actually serve: (1) elastic inference and fine-tuning that must respect data residency; (2) cost-capped research capacity for universities and public agencies; and (3) burst headroom that a national facility cannot justify buying outright for a handful of peak weeks a year.

The honest caveat

"Sovereign" and "decentralised" are not the same thing, and it is a mistake to market them as such. A government that wants its compute on its own soil, under its own legal jurisdiction, is not satisfied by a permissionless network whose nodes could be anywhere. The overlap is narrower than the buzzwords suggest: it is verified, jurisdiction-aware, elastic capacity — not "dependence-free compute", which does not exist at frontier scale for anyone outside two countries.

The useful reframing: sovereignty is a spectrum of managed dependency. A decentralised network competes on which layers it can serve without pretending to replace the ones it cannot.

The bottom line

180 programmes, but 90% of frontier compute in two countries and 45% of sovereign silicon from one vendor. The category is real and growing fast; the independence story attached to it mostly is not. Sell into the gap between the two — elastic, verifiable, cost-capped sub-frontier capacity — and stop pretending the frontier is on the menu.