The Sovereignty Paradox Spreads

The case for "sovereign AI" was always about owning the compute. Governments from Brussels to New Delhi declared they would not be dependent on American chips and American platforms. The argument had a clear target: Nvidia's near-monopoly on the GPUs that train AI models. Two years on, that target is still standing. But the dependency surface has widened into domains the independence programs have not addressed ^[1][2][3][4]. Nvidia is no longer just the company that makes the GPU you cannot train a frontier model without. It is extending its platform into four layers that the sovereignty conversation has not caught up to. First, CPUs. Nvidia launched the Vera Rubin platform with its own custom processors, targeting a $200 billion CPU market where it expects $20 billion in standalone revenue this year. Jensen Huang says every frontier model company — OpenAI, Anthropic, SpaceXAI among them — is adopting the platform from launch ^[1]. The lock-in is moving from the accelerator slot into the central processor. Second, telecommunications. Nvidia invested $1 billion in Nokia to co-develop AI-RAN — software that runs 5G-Advanced and 6G base stations on Nvidia hardware. Nokia is porting its radio software onto Nvidia's CUDA platform, turning cellular infrastructure into another place Nvidia's compute stack lives ^[2]. This is the same 6G network layer that sovereignty planners count as national infrastructure. Third, physical AI and robotics. Nvidia's Isaac GR00T platform — paired with Unitree's humanoid robots on Jetson Thor compute powered by Blackwell GPUs — is being adopted by Stanford, ETH Zurich, and UCSD as a research platform ^[3]. Doosan Robotics is integrating its robot operating system with Nvidia's Isaac stack and Omniverse digital-twin tools, targeting industrial humanoid robots by 2027–2028 ^[5]. LG Electronics brings 31 manufacturing facilities and its ThinQ appliance data to a partnership spanning robotics, automotive, and data-center cooling ^[6]. Nvidia is positioning its platform as the default for executing models in physical systems, not just training them ^[3][5][6]. Fourth, the PC. Nvidia co-developed the RTX Spark chip with MediaTek, debuting in devices from Microsoft, Dell, HP, and Lenovo ^[4]. The platform is reaching down from the server rack into the laptop. While Nvidia extends the dependency surface outward, Palantir has done something more quietly radical: it has commercialized sovereignty itself as a purchasable product. Palantir released a nine-point manifesto urging governments and militaries to maintain control over their data and model weights, and expanded its partnership with Nvidia to build secure custom AI models inside government infrastructure — including air-gapped systems and NATO networks ^[7]. Palantir's Alex Karp frames his company as the trusted implementation layer between frontier labs and governments, dismissing OpenAI's deployment business as a farce ^[8]. The pitch redefines sovereignty from owning infrastructure to buying trusted implementation — from a country you build to a product you purchase. And it runs on Nvidia's stack. India is the sharpest case study in how this plays out. India's sovereignty impulse was triggered by a specific shock: US export controls forced Anthropic to disable its models for foreign nationals, hitting Indian IT firms like TCS and Infosys ^[9]. The government responded with the IndiaAI Mission and GPU clusters. But the hardware being deployed — Yotta's 20,736 B300 cards — is Nvidia's ^[10]. The infrastructure is financed by the very foreign firms whose dominance it was meant to reduce: Amazon committed $48 billion by 2030, Microsoft $17.5 billion, Google $15 billion for a hub in Andhra Pradesh with the Adani Group ^[11][12][13]. India offers tax holidays until 2047 for foreign cloud providers ^[12]. Domestic conglomerates like Adani and Reliance provide land and energy; US firms own the compute layer ^[13]. As Jeet Adani put it:

It's part of our $15B investment in India's infrastructure and will house gigawatt-scale compute, a new international subsea cable gateway and clean energy infrastructure. — Sundar Pichai

. India's advantage is the physical plant. The intelligence layer is not Indian. Component-level progress — C2i Semiconductors' power-delivery chip, a Jabil manufacturing plant in Maharashtra — is real but sits below the GPU layer where the lock-in actually holds ^[14][15]. The counter-evidence, examined closely, confirms the pattern rather than breaking it. Amazon is exploring selling its Trainium chips to third parties for European "sovereign AI" demand — but CEO Andy Jassy simultaneously confirmed Amazon ordered a million Nvidia GPUs for AWS by end of 2027, calling Nvidia a partner "as long as I can foresee" ^[16]. Trainium replaces one American vendor with another. Google is scaling its TPU business with billions in infrastructure financing — but Jensen Huang dismisses the cost advantage, and Google's model mirrors Nvidia's revenue-sharing playbook ^[17]. OpenAI and Broadcom unveiled the Jalapeño inference ASIC, claiming 50% cost reduction — the strongest challenge. As Broadcom CEO Hock Tan put it:

any company serious about leading in AI should not depend on third-party GPUs for such a critical part of its stack. — Hock E. Tan

^[18]. But Jalapeño covers inference only; OpenAI still trains on Nvidia. It is a private cost-reduction strategy backed by $122 billion in capital, not a sovereignty initiative. And it is manufactured through TSMC, Broadcom, and Celestia — still within the US-allied semiconductor stack. The one geography where Nvidia has been genuinely displaced is China, where export controls collapsed its market share from roughly 95% to a projected 8%, while Huawei's Ascend 950 chips rise toward 50% ^[19]. That displacement required export-control coercion, not sovereign ambition. Outside export-controlled China, Nvidia still dominates — 74% of the AI inference chip market and $75.2 billion in data-center revenue ^[20]. Competitors like Broadcom (AI revenue up 143% year-over-year) and Marvell are growing rapidly but from a much smaller base ^[20]. The independence narrative is not failing. It is being absorbed, commercialized, and expanded past — in parallel. The platform grows new dependency layers while governments concentrate on the one layer — data-center GPUs — that is no longer the whole story. And the concept of sovereignty itself is being quietly redefined: from something a nation builds to something a nation buys, from infrastructure it owns to a manifesto it purchases, and from the compute it controls to the compute someone else controls on its behalf.