Five Governments Bought AI Sovereignty Last Week. None of Them Bought a Frontier Cluster.

India's Vice President stood up in New Delhi last week to launch the country's sovereign AI stack. The model at the center of it has 30 billion parameters.

Not three hundred billion. Not a trillion. Thirty.

In the same seven days, Anthropic committed about $45 billion over six years to lease 460 megawatts from a single campus in West Virginia, and AWS said it would deploy 2 million more Nvidia GPUs across 2027 and 2028. Both of those are also, in the loose way the word gets used, sovereignty plays.

One of these two definitions is going to get copied about forty times over the next two years. It will not be the expensive one.

## What the cheap version actually looks like

Gnani AI, a Bengaluru voice AI company, built the Indian stack around an open weight model called Evon v3.3 and an agentic layer called Plexus. The design constraints are revealing: on premise deployment, Indic language optimisation, and data that does not leave the country. There is no frontier scale claim anywhere in it and no attempt to beat the labs on benchmarks.

That is a bet that sovereignty is mostly a question of jurisdiction and language, not parameter count. And last week five other governments placed some version of the same bet.

Spain submitted the EU's only majority state owned AI gigafactory proposal, with the public vehicle SETT holding 47.99% alongside Telefonica, ACS and Santander at 15.67% each. Madrid is putting in 720 million euros of public money to mobilise roughly 5 billion in total, plus 300 million for EuroHPC. Mexico's Congress introduced a bill creating a National Center for Artificial Intelligence as a coordinating body. Beijing's Yizhuang zone published China's first dedicated AI4Chip policy package, aiming to cultivate three to five ecosystem leaders in using AI to design chips rather than run them.

MTN and a Dubai investor formed a joint venture to build roughly 150 megawatts of AI ready capacity in South Africa and Nigeria. Microsoft agreed to put HUMAIN's Arabic ALLAM models inside Foundry and M365 Copilot. OpenAI opened an eight week accelerator in Bangkok for ten Thai startups and began commercial operations in Brazil, where it says users already send about 215 million ChatGPT messages a day.

Look at that list and one thing is missing. Nobody in it is trying to train a frontier model.

## The comparison to be careful with

It is tempting to line up 45 billion dollars against 720 million euros and declare the cheap version a bargain. That is not a fair comparison and it should not be made. A six year compute lease is an operating commitment against a specific revenue plan; a state capital contribution is a different instrument with different risk. The two numbers are not substitutes.

What is comparable is the shape of the ambition. Anthropic is buying capacity to keep training. Spain, India, Mexico and the rest are buying the ability to run, host and govern models that already exist. Those are different products, and the second one is available at a price a mid size economy can actually clear a budget committee for.

## The part that should worry the frontier labs

Nvidia forecast 70 percent sales growth last week, well above the roughly 45 percent analysts expected. Nothing in the sovereign turn threatens that yet, because inference capacity is still bought from the same vendor.

But it does threaten a story the labs have been telling for two years: that national AI capability requires a national partnership with a frontier lab. India just demonstrated a route that runs through an open weight model and a domestic vendor instead. If that route works, the labs are selling into a market that has discovered a substitute.

## What to do with this

**If you are watching policy:** track how many national AI announcements over the next two quarters name an open weight base model rather than a lab partner. That ratio is the single clearest signal of whether the cheap version is winning. It was one for one last week.

**If you sell into government:** the buying question has quietly changed from "can you match the frontier" to "can you run on premise, in our language, under our law." Those are procurement requirements, not research ones, and most frontier products answer them badly.

**The question to ask any sovereign AI pitch:** what happens to this stack if the base model's license changes? Every one of these bets except the compute leases is downstream of somebody else's weights.

You can track the underlying deals as they land in our AI deal tracker.