Sovereign AI Is Turning Into Trade Policy: The Deals Now Being Signed Between Governments

The loudest AI stories of the past month were about money and machines: leases on GPU clusters, debt facilities behind gigawatt data centres, the financialisation of compute itself. That story is real, and we have written about it. But it is not the only thing happening, and it may not be the most consequential one for the next two years.

Look at what actually landed in our deal tracker in the first two weeks of July, and a different pattern shows up. A meaningful share of the week's AI agreements were not signed by labs, or by venture funds, or by hyperscalers. They were signed by governments and by industrial incumbents, and their subject was not a model. It was infrastructure.

## The agreements nobody covered

On 7 July, India and Indonesia concluded intergovernmental agreements to cooperate on artificial intelligence, telecoms, digital public infrastructure and startups, folded into a broader strategic partnership package. The same day, Pakistan and Saudi Arabia reached an agreement in principle to expand their digital corridor with new submarine and terrestrial connectivity, discussed at the AI for Good Global Summit.

Neither carries a disclosed value, which is normal for memoranda of this kind, and neither will move a stock price this quarter. That is exactly why they are easy to miss. But read them next to the Karandaaz Pakistan partnership with D-Tech, which embeds an AI assistant inside digital finance apps with the stated goal of usability and financial inclusion, and the shape becomes legible. Three of the week's agreements involve South and Southeast Asian states, all of them concern plumbing rather than products, and all of them treat AI as a category of national infrastructure to be negotiated bilaterally, alongside cables and payment rails.

This is what AI adoption looks like when it is procured rather than sold.

## Meanwhile, on the factory floor

The industrial deals rhyme. Rockwell Automation and Cisco formed a partnership to co-develop AI-ready, software-defined manufacturing for Indian factories. Siemens and IFS entered an industrial AI partnership built around a shared executable digital twin that links design, production and asset lifecycle data.

Notice what is absent from both. There is no frontier model in the headline. There is no chatbot. The value being created is in the connective tissue: getting a factory's design data, its production data and its maintenance data into one representation that a model can act on. The hard part was never the intelligence. It was the integration, and the companies that own the integration layer are ninety-year-old industrial firms, not labs.

That is a useful corrective to the way most AI coverage, including ours, tends to be weighted. The frontier is where the capability comes from. It is not necessarily where the adoption happens.

## The startup layer is still running the old playbook

To be clear, the classic pattern has not stopped. In the same window, Dragonfly led a Series A into Venice AI, a privacy-first, uncensored assistant that is spending part of the round on its own GPU capacity. Nissay Capital led a $4 million Series A into GenerativeX to scale a forward deployed engineer model, which is a polite way of saying the company sells engineers who build the AI system inside the customer.

Both are interesting. Both are also, in the scheme of things, small, and both are competing in a market where the buyer already has ten options. The state-level agreements are competing in a market where, in many countries, there is no incumbent at all.

## What this implies, and how it could be wrong

If the pattern holds, three things follow over the next eighteen months.

**First, the unit of AI adoption in much of the world becomes a corridor, not a subscription.** A country does not buy seats. It signs a connectivity and cooperation agreement, builds or rents capacity, and then runs domestic services on top. The commercial opportunity is downstream of a diplomatic one, which means the sales cycle is measured in treaties.

**Second, the power constraint gets worse before it gets better.** We argued in Megawatts Over Models that electricity, not algorithms, is the binding limit. Every one of these state agreements adds demand to grids that are already the tightest part of the stack. Bilateral AI cooperation is, in practice, a promise to build data centres somewhere.

**Third, the incumbents quietly win the deployment layer.** If the integration problem is the real problem, then Siemens, Rockwell, Cisco and their peers are better positioned than any lab to capture the margin on industrial AI, because they already own the data and the relationships. The labs supply a component. The incumbents sell the system.

Here is the honest caveat. Intergovernmental memoranda have a long and unglamorous history of going nowhere. An agreement in principle on a digital corridor is not a cable in the seabed, and a strategic partnership package can be signed, photographed and forgotten. The falsification test is simple and worth watching: within twelve months, do these agreements produce financed, sited, contracted capacity, or do they produce a second round of agreements? If it is the latter, this was diplomacy, not infrastructure, and the frontier labs remain the only story that matters.

Our bet is that at least some of it converts, for an unsentimental reason. The states signing these deals are not doing it because they are excited about AI. They are doing it because they have watched compute become a strategic input, priced and allocated abroad, and they would prefer not to rent it forever. That motive tends to survive a change of government.

We track every one of these agreements as they are announced. If you want the pattern rather than the noise, the deal tracker is here, and you can get the analysis by email.