Sovereign AI is a lie

Every government wants sovereign AI. The AI Alliance just launched Project Tapestry to build "collaborative foundations for open and sovereign AI." India announced $200 billion in AI infrastructure investment. The EU is pouring €200 billion into its AI Continent Action Plan. France, the UAE, South Korea, and Singapore are all racing to build national AI capabilities. The pitch is compelling: own your AI stack, control your data, secure your future. But look beneath the rhetoric and the dependency chain tells a different story. The chips are American. The cloud is American. The architectures are American. "Sovereign AI" is less a technical reality than a political brand, and understanding the gap between the two matters more than most governments want to admit.

What sovereign AI actually means

McKinsey defines sovereign AI along four dimensions: territorial (where data and compute physically reside), operational (who manages and secures them), technological (who owns the underlying stack and IP), and legal (which jurisdiction governs access and compliance). The Brookings Institution frames it as "a spectrum of strategies to enhance a country's capacity to make independent decisions about critical AI infrastructure," rather than literal self-sufficiency. That spectrum matters. Almost no country is pursuing full autarky. What most governments actually mean when they say "sovereign AI" is some combination of data residency requirements, locally hosted cloud infrastructure, and the ability to fine-tune models on domestic data. That is a far cry from owning the full stack. But the rhetoric rarely makes this distinction. When politicians say "sovereign AI," they imply independence. When the supply chain says otherwise, that gap between narrative and reality becomes a problem.

The dependency chain

Start at the bottom of the stack: chips. Nvidia holds approximately 80% of the AI accelerator market. Every country building "sovereign" AI infrastructure is, almost without exception, doing it on Nvidia GPUs. South Korea's sovereign cloud runs on over 250,000 Nvidia GPUs. India's Reliance Industries is building multi-gigawatt data centers in Gujarat, powered by Nvidia hardware. Even the AI Alliance's Project Tapestry, which aims to enable distributed training of frontier open models, will run on hardware designed in Santa Clara and fabricated in Taiwan by TSMC. Amazon and Google are developing custom silicon, with AWS deploying over 500,000 Trainium chips and Google rolling out its seventh-generation TPU Ironwood. But these alternatives remain concentrated within US hyperscalers. They do not solve the sovereignty problem; they merely shift the dependency from one American company to another. Move up to cloud infrastructure. A recent study of 775 non-US data center projects found that US companies served as operators for 18% of those projects, but accounted for 48% of total data center investment and 56% of AI investment. AWS, Microsoft Azure, and Google Cloud dominate the operational layer of "sovereign" compute worldwide. Amazon's European Sovereign Cloud is governed by a European board and staffed by European employees, but it is still Amazon's infrastructure. The Council on Foreign Relations noted that such quasi-sovereign solutions "may be the closest our allies come to true digital sovereignty." Then there are the models. The Transformer architecture was invented at Google. The most capable foundation models come from US companies (OpenAI, Anthropic, Google, Meta) or, increasingly, Chinese labs (DeepSeek, Alibaba's Qwen). A country can fine-tune an open-weight model on local data, but the base model, the architecture, and the training methodology all originate elsewhere. The dependency is not at one layer. It is at every layer, simultaneously.

The sovereign internet precedent

This is not the first time countries have tried to claim sovereignty over a technology stack they do not control. The "sovereign internet" attempts by Russia and China offer instructive parallels. China's Great Firewall, launched in 1998, is the most successful example of digital sovereignty in practice. It blocks foreign websites, mandates domestic alternatives, and gives Beijing extensive control over information flows. But it required decades of investment, a massive censorship apparatus, and the development of an entire parallel ecosystem (Baidu for Google, WeChat for everything else). Even then, China's tech giants still depend on foreign semiconductor equipment, with ASML's lithography machines remaining irreplaceable. Russia's Sovereign Internet Law, passed in 2019, aimed to create a similar capability. The results have been mixed at best. Russia has developed domestic alternatives like VKontakte and Yandex, and has implemented deep packet inspection across its internet infrastructure. But the economic cost has been significant, with internet shutdowns alone reportedly costing the Russian economy $250 million per month. The Henry Jackson Society noted that Russia's sovereign internet may ironically reduce the Kremlin's control, as frustrated citizens revert to cash and offline activity to escape surveillance. The lesson is consistent: partial sovereignty is achievable, but at enormous cost, and it never eliminates the underlying dependencies. China can censor the internet, but it cannot make its own EUV lithography machines. Russia can filter traffic, but it cannot build competitive cloud infrastructure. Sovereignty over a technology stack you did not invent is, at best, a managed dependency.

The open-source angle

Open-weight models are the most realistic path to something resembling AI sovereignty, and everyone knows it. Meta's Llama, Google's Gemma, Mistral's models, Alibaba's Qwen, and DeepSeek have made frontier-capable AI freely available for download, fine-tuning, and deployment. The Linux Foundation's survey found that 82% of organizations are already developing customized AI solutions to maintain control over their capabilities. CNBC reported that experts see open-source models and cloud computing as key enablers for countries building sovereign AI environments. Project Tapestry takes this further, proposing a federated training platform where participants can help build a shared open foundation model while retaining the ability to create sovereign derivative models tailored to their own needs. Yann LeCun, appointed as the project's Chief Science Advisor, has long argued that open models are essential for preventing AI concentration. But open-weight is not open-stack. You can download Llama and fine-tune it on your own data, but you still need Nvidia GPUs to run inference at scale. You still need cloud infrastructure to serve it. You still need the electrical grid capacity to power it. The model layer is the one layer where sovereignty is most achievable, and even there, you are building on architectures and training techniques developed by a handful of US and Chinese labs. Open source reduces lock-in. It does not eliminate dependency.

Who benefits from the narrative

If sovereign AI is more aspiration than reality, who benefits from maintaining the fiction? Nvidia, for one. Jensen Huang has been the most effective evangelist for sovereign AI, describing it as a "five-layer cake" spanning energy, chips, computing infrastructure, AI models, and applications. Every layer of that cake requires Nvidia hardware. The company has signed sovereign AI partnerships with South Korea, India, Japan, and dozens of other countries. Sovereign AI is, in practice, Nvidia's most effective sales pitch. US hyperscalers benefit too. AWS, Azure, and Google Cloud are rolling out "sovereign cloud" offerings that meet local compliance requirements while keeping the operational layer firmly under American control. They get to sell the same infrastructure twice, once as a global product and again as a localized, sovereignty-compliant variant. Local consultancies and system integrators benefit from the complexity. Every sovereign AI initiative needs local partners to navigate regulatory requirements, manage data residency, and integrate with government systems. The more complex the sovereignty requirements, the more consulting revenue they generate. And politicians benefit from the narrative itself. "We are building sovereign AI" is a powerful message that signals technological ambition, national security awareness, and forward-thinking governance, regardless of whether the underlying reality matches the rhetoric.

Singapore as a case study

Singapore is perhaps the most honest example of a small nation navigating the sovereign AI contradiction. Six million people, no natural resources, a land area smaller than most major cities, and zero chip fabrication capacity. Singapore cannot build a sovereign AI stack in any meaningful sense. It knows this, and its strategy reflects that honesty. The National AI Strategy 2.0, launched in 2023, does not promise independence. It promises relevance. The strategy emphasizes three shifts: from opportunity to necessity, from local to global, and from narrow applications to broad capability building. In early 2026, the government committed over S$1 billion to the National AI Research and Development Plan, funding public-sector AI research through 2030. On governance, Singapore has been pragmatic. The Model AI Governance Framework is voluntary, industry-friendly, and built around principles of explainability and human-centricity. AI Verify, an open-source governance testing toolkit, lets companies assess their systems against international standards without the punitive compliance costs of the EU AI Act. In January 2026, Singapore published the world's first governance framework for agentic AI, addressing autonomous systems that can reason and act independently. But Singapore remains deeply US-dependent. Its data centers run on American hardware. Its cloud infrastructure is operated by American hyperscalers. Google expanded its AI investments in Singapore in February 2026. The government's own deployment of agentic AI in the public sector runs on Google's air-gapped cloud. Singapore's approach is not sovereign AI. It is strategic dependency management, and it might be the most realistic model for most countries. Rather than pretending to own the full stack, Singapore focuses on what it can control: governance frameworks, talent density, regulatory clarity, and regional connectivity. It accepts the dependency and works to minimize the risks.

What actually matters

The honest framing is not "sovereign AI" versus "dependent AI." It is a question of which layers of the stack a nation must control, which layers it can safely rent, and which layers it should share. Data residency has real value. Keeping sensitive government and citizen data within national borders, subject to domestic law, is a legitimate and achievable goal. Local fine-tuning matters too. A model trained on local languages, cultural context, and domain-specific data serves its users better than a generic global model. But pretending that data residency and fine-tuning equal sovereignty is misleading. The compute layer, the chip layer, and the cloud operations layer remain concentrated in a handful of American and, to a lesser extent, Chinese companies. That concentration is a structural feature of the AI industry, not a temporary condition. The countries that will navigate this best are the ones that drop the sovereignty theater and focus on practical resilience: diversifying suppliers where possible, investing in open-source ecosystems, building governance frameworks that attract investment, and maintaining the talent base needed to adapt as the technology evolves. Sovereign AI, as politicians use the term, is a lie. But strategic AI capability, built on honest assessments of dependency and realistic investments in the layers that matter, is very much achievable. The distinction matters more than most governments are willing to admit.

References