I hand out a rule to other people and then break it myself, reliably, about once a quarter.
The rule: when a government announces it has built something, go find out what it bought.
In February I did not apply it. An Indian company called Sarvam launched two indigenous large language models trained specifically on Indian languages, and I read the headline and felt exactly what you are supposed to feel: good. A country long described as everyone else's back office had gone and written the software instead of the tickets.
Then, about a week late, I went looking for the technical details of the training run. I found them in a case study published by Nvidia.
That should have been the moment. It took me longer — but it is where this piece begins.
What Sarvam actually built, and it is not nothing
Let me be precise about the achievement, because the rest of this essay is hard on the framing.
The two models were trained from scratch rather than fine-tuned on somebody else's open-source system — a real distinction, and the harder road — using computing resources provided under India's government-backed IndiaAI Mission. The larger one, at 105 billion parameters, claims state-of-the-art performance across 22 Indian languages for its model size and is published under the Apache license, weights and all. Sarvam calls building from scratch a core requirement for a sovereign stack, and about that layer, it is right.
Two dates, because they keep getting merged: the models launched on February 18, 2026, at the India AI Impact Summit in Delhi; the weights went public under Apache 2.0 on March 6.
So far so good — a real model, really trained, genuinely open, in languages the big American labs treat as a rounding error.
Now count the layers underneath it
Here is where things get uncomfortable.
Start with the compute. Nvidia's own case study reports near-linear scaling across 4,096+ Nvidia H100 GPUs. Not Blackwells — H100s, a generation back. And a number Nvidia is happy to publish, because it is Nvidia's.
Where did those chips come from? India answered that itself, in a written statement to the Lok Sabha on March 25, 2026. More than 38,000 GPUs have been onboarded to the national compute facility, said Jitin Prasada, the minister of state for electronics and IT — and then the sentence that concedes the whole argument: GPUs are primarily manufactured in one country. That country is not India. Carnegie is blunter: as of early 2026 those 38,000 GPUs were all supplied by American firms.
Now go one layer below the chip, because this part almost never survives the coverage.
The memory. An accelerator is useless without high-bandwidth memory bonded to it, and HBM is a three-supplier market. In the first quarter of 2026, SK hynix maintained its top position with a 58% share, Samsung and Micron 21% each.
The design software. Every advanced chip on earth is designed using software from three companies — Synopsys, Cadence and Siemens EDA — holding over 85% combined market share. The two largest are American; at 12nm and below, one of them has effectively all of it. Nor can you swap a piece out: the flow itself is the lock-in.
The lithography. The machine that etches the pattern requires an export license, because extreme ultraviolet lithography sits on the Wassenaar list of dual-use technologies. There is one supplier on earth, and it is Dutch.
The packaging. The last step, bonding chip to memory, happens overwhelmingly in Taiwan — where global advanced packaging capacity is currently in severe shortage and Nvidia has already reserved most of TSMC's leading-edge CoWoS lines.
Add it up and you have a model that is Indian in roughly the way a restaurant is French when the chef is French and the kitchen is leased.
Nor is the scale gap close: India has fewer than 80,000 high-end GPUs to America's roughly hundredfold more. What it does have sits oddly quiet — by July 2026 the mission had far more compute than anyone is using — and the layer above was settled long ago, with the facial boarding system at Indian airports running entirely on AWS.
The strongest case against everything I just said
Let me try to break my own argument, because the serious counter is not "actually, India is sovereign."
It runs like this: full-stack autarky was never the goal and would be a foolish one. No country builds its own everything; the United States doesn't. Sovereignty in practice means calibrated interdependence — strategic choice without cutting yourself off from frontier capability. Judged that way, India is succeeding.
Serious people argue it. Stanford's Institute for Human-Centered AI puts it carefully: sovereign AI offerings reconfigure rather than eliminate dependencies on foreign providers, and most still depend on some type of US technology. Its policy work calls the term systematically underspecified and argues that the core policy challenge is calibrating interdependence rather than escaping it. Carnegie agrees full AI sovereignty is beyond the reach of even the most capable middle power actors — which makes holding India to it a rigged test. An academic framing says it best: the foundations of AI, from data pipelines to open-source ecosystems to standards, resist enclosure.
And India is doing the hard part too. Tata is building the country's first mega fab in Dholera, up to ₹91,000 crore of it, in partnership with PSMC of Taiwan. Carnegie India calls that mission's progress on packaging, fabrication and tooling proof that the difficult work of industrial policy is possible in India — while warning that compute purchased is not the same as capability built.
Two caveats on that fab. Its nodes are 28nm to 110nm — power management chips, display drivers, microcontrollers: real, valuable, not the business a 4nm-class accelerator is in. And per that same statement to Parliament, India's design scheme subsidizes access to EDA tools it does not own, while its advanced tape-outs go to 12nm at TSMC.
So the fair version of my complaint is narrower than my headline. Not "India isn't sovereign." Rather: the word is doing work the stack cannot support, at the one layer that is cheapest to reach.
If you think that is a developing-economy problem, look at the bloc that owns the one machine everybody needs
The European Union is richer than India, started earlier, and holds the one genuine chokepoint the West controls. It is stuck in the same place — and unusually honest about it.
Start with the Commission's own text. In its June 3, 2026 legislative proposal — the Chips Act 2.0, COM(2026) 504 final — it wrote that the EU produces less than 10% of global semiconductors and is almost entirely dependent on the United States and Asia for "the most advanced and leading-edge chips below 5 nanometres," and for AI chips. Its summary adds that this is not only about factories: the EU remains dependent on third countries in key areas such as advanced chip manufacturing or semiconductor design. All of it sits inside a broader European technological sovereignty package covering semiconductors, AI, cloud and open source.
Europe legislated the admission. That is more candor than most governments manage.
Its auditors were harsher. The European Court of Auditors found it very unlikely that the EU will meet its target of a 20% share of the global microchip market by 2030 — projecting a rise from 9.8% in 2022 to just 11.7%, against top manufacturers who budgeted €405 billion in three years, a sum that "dwarves the financial firepower of the Chips Act." Its member Annemie Turtelboom said Europe urgently needs a reality check.
Then the machine. JUPITER, at Jülich, is Europe's first exascale supercomputer and the flagship of its computing sovereignty. Per the center's own technical page, the driving chip of the system is the NVIDIA Hopper GPU. How many? Roughly 24,000 NVIDIA GH200 Grace Hopper superchips — with Barcelona's MareNostrum 5 on 4,480 Nvidia Hoppers and Lugano's Alps on 10,752.
And the European processor in it? Rhea1, out of the European Processor Initiative, powered on this year. Scheduled for 2021, it will reach the market five years late, and its own vendor now puts availability at the end of 2026. More than 2,600 go into parts of a JUPITER cluster already running.
Twenty-four thousand American parts. Twenty-six hundred European ones, half a decade behind. If that shape looks familiar, it should: it is Sarvam's 4,096 H100s with more zeros on it.
Europe's 2026 answer is not a factory. It is a tender. On July 30, 2026 — seventeen days before I wrote this — EuroHPC opened a call to establish up to seven AI Gigafactories, each with "a very large number of state-of-the-art AI processors." Bids close November 12, 2026; selection follows in early 2027; operations are expected to begin within 18 months of that. Public money acts as anchor customer to unlock an expected €20 billion-plus of private investment, and the bulk of the public share comes from the next Multiannual Financial Framework, 2028 to 2035. The Commission's project page records 77 proposals from 16 Member States across 60 sites.
Read the specification again: state-of-the-art AI processors. Not European ones. There are none to ask for.
Europe has its own Sarvam, too. OpenEuroLLM is an EU-funded family of models covering all EU official languages, €20.65 million from the Digital Europe Programme, trained on EuroHPC — which is to say, on Nvidia. Same sentence as India's, different language.
At the cloud layer Brussels has legislated four assurance levels of sovereignty for public bodies to choose between, having concluded that over-reliance on non-EU cloud providers poses a significant risk to Europe's digital autonomy. It is not wrong: three US providers control more than 70 percent of the European cloud market, EU providers about 15. You legislate assurance levels when you cannot legislate the chips.
Which brings me to the smartest thing said about any of this, said by a European. Bruegel argues Europe should stop chasing self-sufficiency — leading-edge production in the EU is unrealistic in the near future — and pursue sovereignty through indispensability: control the chokepoints others cannot route around. Then it names the one Europe holds, in words I would be accused of exaggerating had I written them. This Dutch firm, it says, "has a monopoly on extreme ultraviolet (EUV) lithography equipment, required for leading-edge semiconductor production."
That is what sovereignty looks like when you have it. Not a model in your own language. A machine nobody else can build.
Now run the tape forward to a Tuesday in 2029
Here is where I get imaginative — plausible rather than lurid, because the plausible version is worse.
We already know the terms can change with a memo. In May 2025 the Commerce Department announced the rescission of the Biden Administration's AI Diffusion Rule, a framework that would have sorted the planet into tiers of chip access. Its stated objection: the rule would have undermined relations with dozens of countries by "downgrading them to second-tier status" — which concedes that tiers exist and Washington assigns them. The controls were officially paused, not abolished, with a replacement promised for some point in the future.
We also know what a permission slip looks like. In November 2025 the US authorized the UAE's G42 and Saudi Arabia's HUMAIN to buy the equivalent of up to 35,000 Nvidia GB300s each — conditioned on both companies meeting rigorous security and reporting requirements, monitored by Commerce. The trade press caught the grammar exactly right: the U.S. authorizes Nvidia chip exports.
So: 2029. Pick a country — India, Spain, Brazil, it hardly matters. Three years into its national model, four ministries and a hospital network running on it. Then Washington ties accelerator export licenses to something adjacent but not technical: a data-localization law, a content-moderation posture, a vote in a multilateral body. Nothing is seized, nothing switched off. The next tranche of chips simply moves from "approved" to "under review," and depreciation on the last tranche does the rest. Eighteen months later the sovereign model runs on hardware nobody will service.
And through all of it the minister is still correct. The model is theirs. That is what makes the scenario work.
There is a number for how much of this got sold last year. On Nvidia's February 25, 2026 earnings call, chief financial officer Colette Kress said the company's sovereign AI business had more than tripled year over year to over $30 billion, driven primarily by customers in Canada, France, the Netherlands, Singapore and the UK — against total fiscal 2026 revenue of $215.9 billion, up 65%. National sovereignty is a product line, and better than one dollar in eight of the whole company.
The seller's framing is the tell. Every country, Kress said on that call, "will build and operate some parts of its AI infrastructure, just like with electricity and Internet today." Some parts. She is describing her market accurately, which is more than most of her customers' press releases manage.
What the people who study this for a living are saying
None of this is a fringe reading, and it does not sort neatly by ideology.
The historian Chris Miller, who maps these chokepoints for a living, told an audience at Carnegie Mellon in February 2026 that even the United States — holder of the largest share of the chip industry — is fundamentally reliant on Europe, Japan, Korea, Taiwan, and China for critical links in the chain. In a 2024 interview at the American Enterprise Institute he compressed it further: of advanced chips, around 90 percent of them are made in one island — extraordinary, he said, for earthquake risk and more importantly for geopolitical risk.
From a different direction, the AI Now Institute wrote in March 2024 that India was far from having the capability to manufacture the GPUs that have become the bedrock of AI development, and called the country's foray into AI "more reactive than strategic." Two years on, the manufacturing half has not changed.
In Delhi, the Observer Research Foundation said it plainly in July 2026: virtually every sovereign AI model trained on Indian financial data depends on hardware whose export is subject to a foreign government's licensing decisions, not India's own.
And the market skeptics deserve more airtime. A panel convened by Discourse Magazine in October 2023 noted that even on the industry's own projections, the whole subsidy edifice buys a couple of percentage points uptick in global market share — while everybody else answers with subsidies of their own. Find that dispositive or not; do not skip it.
The tidiest formulation I have read arrived this month, from an Indian columnist arguing that sovereignty in AI is not something a country has or lacks but a property of each layer of a very tall stack — and that the layer earning the headlines, a national model, is the cheapest to reach and the least decisive.
Cheapest to reach. Least decisive. That is the whole essay in five words.
What does this mean for you?
When you read "sovereign," ask which layer. Model, compute, cloud, chip, memory, design tools, lithography, packaging. A claim about one is not a claim about the others, and the press release will never tell you which it means.
Treat "trained in our country" as geography, not control. It tells you where the electricity was spent, not who can decline to sell you the next machine.
If you build on a national model, read the license and the hardware in one sitting. Sarvam's weights really are Apache 2.0 and really are yours; that part is durable. The cluster is a rental, and so is yours.
Watch the export-control calendar, not the launch calendar. Launches are announced months ahead by people who want your attention. License policy changes on a Tuesday, in a notice nobody live-blogs — and it decides what your country can buy in 2029.
Ask the question at home. Whatever country you are in has an AI strategy. Find the paragraph about compute. If it uses the verb "procure" rather than "produce," you know which layer your government is on.
My vote, having finally read the whole stack
Building the model was worth doing, and I want that said plainly, because the argument above is easy to misread as contempt. It isn't. A country with 22 official languages that waits for San Francisco to care about Maithili will wait a long time. Sarvam built the thing, trained it the hard way, and published the weights. That is a public good, and it is India's.
What I object to is the word doing the marketing. Sovereign implies the capacity to keep operating when somebody else would rather you didn't. Measured that way, a national model is a flag over a rented building — the lease written in Washington, the fittings from Hsinchu and Icheon, the machine that cuts the key sitting in Veldhoven.
The honest alternative is not defeatism; it is Bruegel's. Find the layer where you can make yourself impossible to route around, invest there like you mean it, and buy the rest openly with the political risk priced in. India has candidates: advanced packaging, memory, or the open RISC-V instruction set its supercomputing mission already designs processors around — chosen, tellingly, because the closed ones are licensed. Ten-year projects with no launch event in year one, which is exactly why they are the real test.
The next time any country announces it has built its own AI, the useful question is not whether the model is good. It is: which layer, and who can turn it off?
Ask it about India. Ask it about Europe. Then ask it about home.
Sovereignty you can lose to a policy memo was a subscription all along. If this landed, forward it to the next person who says "sovereign AI" without blinking — and subscribe for the next one.