The AI Sovereignty Race Left Software Behind

In June, while Russian missiles were still striking its cities, Ukraine's Ministry of Economy signed a deal with Kyivstar and VEON to build a domestic AI data center. The project is modest — 3 to 5 megawatts, tens of millions of dollars — and the military is its primary customer ^[1]. VEON's CEO was explicit about why a country at war would spend scarce resources on server racks.

This reinforces the need for countries to build local AI infrastructure to stay competitive globally, and we are focused on advancing this in Ukraine. — Oleksandr Komarov

The same month, across the Atlantic, local opposition in American states and cities had blocked nearly $100 billion in proposed AI data center projects ^[2]. New York imposed a statewide moratorium. Virginia localities ended by-right development, forcing each project through a political gauntlet. Every jurisdiction was improvising its own framework ^[3]. The juxtaposition is not merely ironic. It captures something that has become visible only in the past year: the question of who controls AI has migrated from software to physical infrastructure, and the countries whose governments can plan, approve, and fund that infrastructure at a national level — whether through authoritarian decree or democratic national planning — are pulling ahead of the market-led model. The US, where local veto power and fragmented jurisdiction scatter what other countries centralize, is the outlier. The trigger for much of this was American export controls. When the US blocked NVIDIA's Blackwell chips from reaching China, Beijing responded by mandating that all state-funded data center projects use exclusively domestic silicon, ordering projects less than 30 percent complete to rip out foreign hardware ^[4]. NVIDIA's share of Chinese data-center compute dropped from 95 percent in 2022 to zero. China's Zhipu AI then launched a 1-gigawatt data center powered entirely by Chinese-made chips, training next-generation models independent of NVIDIA hardware ^[5]. The decoupling was total: both nations had weaponized hardware access, and China had built its way out. India learned the same lesson through a different door. When the US forced Anthropic to disable its models for all foreign nationals, the disruption hit Tata Consultancy Services, Infosys, and a generation of Indian startups that had built on top of American AI ^[6]. Sarvam AI, the Bangalore-based firm that would go on to build models outperforming Google's Gemini on Indian-language tasks, crystallized the conceptual pivot.

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

India's response has been the most comprehensive of any democracy. The Ministry of Power now projects 26.3 gigawatts of additional AI data center load by 2031-32, nearly double its estimate from March ^[7]. The government has secured $70 billion in AI infrastructure commitments from Amazon, Microsoft, Google, and domestic firms, all in the execution phase ^[8]. It is building nuclear power capacity toward 100 gigawatts by 2047, modernizing transmission under a new National Electricity Policy, and manufacturing AI data center components domestically through a Jabil plant in Maharashtra backed by the Telecom Production-Linked Incentive Scheme ^[9][10]. The state is supporting roughly 20 foundational models and allocating 88,000 GPUs to startups. The ambition runs from silicon to software, and the government is planning it as a single integrated stack. Other countries have arrived at the same conclusion through radically different motivations. Saudi Arabia, through its state-backed firm Humain, partnered with xAI and NVIDIA to build a 500-megawatt data center targeting 400,000 chips by 2030 — a direct conversion of oil capital into AI compute capital, explicitly framed as economic diversification beyond petroleum ^[11]. Japan, fearing what its own officials call becoming an "AI colony" dependent on foreign models and hardware, launched FRONTia: a national physical AI infrastructure project with 27,500 NVIDIA Rubin GPUs and a 140-megawatt AI factory managed by a 44-company consortium ^[12]. Jensen Huang, on stage at the launch, made the stakes plain.

Japan cannot outsource its national intelligence. — Jensen Huang

South Korea's government announced a $575 billion joint initiative with Samsung and SK Hynix to build four fabrication plants and a packaging facility — a state-orchestrated bet on physical AI hardware capacity ^[13]. Canada's Bell Canada partnered with Cohere and domestic server-builder Hypertec to build sovereign AI infrastructure explicitly as an alternative to US hyperscalers, with the Treasury Board warning that "Canada cannot compete in the global AI economy without the infrastructure, talent and partnerships to support it" ^[14]. Even Italy and Oman signed a bilateral deal for a 150-megawatt green AI data center — two mid-power nations pooling resources to build physical infrastructure neither could justify alone ^[15]. The motivations could hardly be more different: export-control pressure for India and China, oil-to-compute diversification for Saudi Arabia, wartime necessity for Ukraine, colonial fear for Japan, economic competitiveness for South Korea and Canada. But the response is the same everywhere: treat the physical layer beneath AI — energy generation, chip fabrication, data center construction — as sovereign territory that cannot be outsourced. The United States is running a different playbook. The Trump administration designated data centers as critical national security infrastructure, and the Defense Department is seeking private developers to build AI facilities on military base land, including 207 acres at Fort Hood ^[16]. But the American approach uses national-security designation as a regulatory claim rather than an ownership claim — the government clears the way for private capital rather than directing it. And private capital keeps hitting a wall that state-directed capital does not: local opposition. The nearly $100 billion in blocked projects is not a temporary friction. It is what happens when every county board, state legislature, and municipal zoning commission holds a veto over infrastructure that other countries are building through national plans. The binding constraint, moreover, is not money. Cushman & Wakefield, in a report rejecting the hypothesis of an India AI data center bubble, identified the real bottleneck.

Grid power availability, or the ability to colocate private generation in the absence of available grid capacity, remains the chief concern in the data center industry, as power is necessary for compute. — Cushman & Wakefield

Power, not capital, is the scarce resource — and securing power at the scale AI demands requires the kind of coordinated energy-infrastructure planning that India's National Electricity Policy 2026 enables and that America's fragmented jurisdiction makes extraordinarily difficult. Two counter-currents are worth noting, because both represent bets against the physical-buildout consensus. The European Union is pursuing sovereign AI through cloud contracts rather than physical construction — a 180-million-euro, six-year deal with European providers under its Cloud Sovereignty Framework ^[17]. And Perplexity CEO Aravind Srinivas argued in January that on-device AI is the real threat to centralized data centers.

The biggest threat to a data center is if the intelligence can be packed locally on a chip that’s running on the device and then there’s no need to inference all of it on like one centralized data center. — Aravind Srinivas

Both represent a version of the thesis that software can solve the physical problem. But the physical infrastructure buildout has only accelerated since Srinivas made that argument, and the EU's own cloud contractors acknowledge that technological sovereignty requires actual infrastructure and platforms, not just legal frameworks ^[17]. The market, for now, is voting for steel and silicon. The implication is uncomfortable. A race to own the physical infrastructure beneath AI may favor the countries whose governments can plan, approve, and fund at a national level — whether through the authoritarian decree of Beijing and Riyadh or the democratic national planning of New Delhi, Tokyo, and Seoul. The fragmentation of American jurisdiction is not a bug to be fixed; it is the system working as designed. But it is also what stands in the way. "America is going to win it," declared from a podium, is a different thing from owning the steel, the silicon, and the grid beneath the models. The question the next year will answer is whether a market-led model, constrained by local democracy, can match the speed of governments that have decided the physical stack is simply too important to leave to the market.