Sovereign AI Runs on Nvidia

Three governments on three continents have declared "sovereign AI" in recent months. All three built it on Nvidia. Bell Canada's "sovereign AI infrastructure" partnership with Cohere runs on Nvidia's DSX AI factory platform ^[1]. In South Korea, Naver expanded its GAK Sejong facility to gigawatt scale under a broad Nvidia alliance with SK Telecom, Samsung, and Hyundai, and labeled the result sovereign AI infrastructure ^[2]. Across 23 European countries, Nvidia and partners are deploying 35 AI supercomputers, and Dell's Arun Narayanan put the framing bluntly ^[3]:

The institutions doing the world's most important research like decoding the human genome, modeling the energy systems of the future and building the sovereign AI infrastructure that nations depend on deserve infrastructure that matches the ambition of their work. — Arun Narayanan

The sovereignty these projects assert lives at the data and governance layer: whose data, whose rules, whose jurisdiction. The compute layer underneath is Nvidia's. Nvidia holds 86% of AI revenue in data centers and roughly 90% of the discrete GPU market ^[4]. Asserting sovereignty on that base is like declaring energy independence by signing a long-term lease on someone else's oil field. What makes the lock-in feel like independence is Nvidia's DSX platform. DSX lets partners deploy AI infrastructure without buying chips upfront, through revenue-sharing and credit-support agreements. Firmus Technologies is using this model to place 170,000 Nvidia chips in a 360-megawatt AI factory in Indonesia, with an eight-year revenue-sharing deal and no initial capital outlay ^[5]. As Firmus's Tim Rosenfield put it:

This is actually a really material way to level the playing field a little bit to give the next a chance to compete with the big guys. — Tim Rosenfield

That framing is the seduction. DSX removes the capital barrier that would otherwise push governments toward domestic alternatives. A government can announce a sovereign AI initiative, sign a DSX deal, and feel it has chosen independence, while the chips, the CUDA software, and the factory architecture all remain Nvidia's. India's sovereign AI push shows what happens when the dependency bites. US export controls forced Anthropic to cut off access to its models for Indian users, exposing the risk of relying on foreign-owned AI infrastructure ^[6]. India is now building domestic GPU clusters and indigenous foundation models ^[7]. Sarvam AI, one of India's domestic model builders, put the lesson plainly:

For AI users, it is clear that you should not confuse access with ownership, or adoption itself as an advantage. — Sarvam AI

The EU reached a similar conclusion through a different door: its Technological Sovereignty Package, unveiled June 3, restricts foreign cloud providers from public tenders in defense, healthcare, banking, and energy ^[8]. But even the EU's response targets cloud and data governance, not the chip layer where Nvidia's dominance sits. The UK is the exception that proves the rule. It chose AMD, not Nvidia, as the primary chip partner for its £1.1 billion sovereign compute strategy, with AMD committing £2 billion over five years ^[9]. That required an active government decision against the default, against the platform every other Western sovereign AI project has adopted. Divergence was possible. It was not the natural path. China is the only country that fully escaped, and it did so wholesale. Zhipu AI trained its GLM-5.2 model entirely on 100,000 Huawei Ascend processors, producing a frontier-class model at less than one-tenth of Anthropic's cost ^[10]. ZTE debuted a full-stack AI factory and 6G prototypes, positioning itself as the Chinese alternative to Nvidia's infrastructure buildout ^[11]. China did not swap one supplier for another. It built a parallel universe: domestic chips, models, 6G, and AI factories. That is the only path that actually breaks the dependency, and it is not incremental. It is total. The implication for every country now shopping for sovereign AI is uncomfortable. You can label the project sovereign, keep data within your borders, write your own governance rules. But if the compute layer is Nvidia's, sovereignty is a story you tell yourself about someone else's hardware. The UK paid the political cost of choosing differently. China paid the economic cost of building everything from scratch. Everyone else is signing DSX agreements and calling it independence.