What I Heard in Delhi About Sovereign AI
Observations and thoughts on sovereign AI after the AI Impact Summit
This week I was in New Delhi for the AI Impact Summit, where AI sovereignty was the topic of many panels and conversations.1
What stood out was both the level of interest and the range of views on what “AI sovereignty” should mean in practice.
One camp argued the term is vague and needs more definition to be operational — probably the closest thing I heard to an objection in principle. Another argued that the basic intuition of sovereignty makes sense but is only feasible for wealthier countries that can afford to build significant parts of the AI stack domestically. A third view, closer to my own, is that AI sovereignty is difficult and maybe impossible as a blanket ambition, but achievable for targeted objectives — through concrete strategies that reduce but don’t eliminate interdependence and disruption risks. A fourth held that sovereignty is largely achievable through open models, open software, and shared computing resources.
That’s not a comprehensive accounting, but it captures the breadth. What’s notable is the view I didn’t hear: that sovereignty isn’t a worthwhile goal at all. Not long ago, I would have expected more pushback — the argument that interdependence is the point, and that the task is managing it well rather than unwinding it. That argument hasn’t disappeared, but it’s no longer functioning as pushback against sovereignty. It’s now a point along the sovereignty continuum, representing one version of what AI sovereignty could look like.
Another thing that stood out is that, though the subtext — and often the text — of sovereignty comments was about achieving some distance from the United States, I heard very little acknowledgment that the Trump administration has expressed support for sovereign AI. Part of that may be unfamiliarity with the U.S. position. But part of it reflects a deeper disconnect. Many countries are pursuing sovereignty as a hedge against dependence on the American AI stack, while Washington is framing sovereignty as something the American stack can deliver.
That mismatch is what I’ve called the U.S. AI sovereignty gap, and it’s getting harder to ignore. According to my research, government-backed sovereign AI projects outside the U.S. and China have grown from roughly 50 across about 30 countries in 2024 to roughly 130 across more than 50 countries today. Reducing reliance on foreign tech — including the United States’ — is a recurring theme of these projects.
I lay out that argument in detail in a new Lawfare essay, “The Sovereignty Gap in U.S. AI Statecraft.”
Two conversations, one summit
Since the essay’s publication and during the summit itself, Trump administration officials provided a clearer picture of how they are trying to meet sovereign AI demand without conceding the underlying premise that many countries are hedging against.
At the summit, White House OSTP Director Michael Kratsios delivered remarks that appeared, in part, to be an effort to bridge the sovereignty gap. The day before, he made a similar case in the Financial Times, in more development-oriented terms.
The core message across both is that the best AI stack is made in America, the United States is eager to deliver it to partners, and the American AI Exports Program is the vehicle — backed by financing tools, a proposed Tech Corps, and the rejection of centralized global AI governance.
It’s important to note that the administration’s sovereignty argument this week is more serious than simply “buy American.” It has two parts that deserve engagement on their merits.
The first is an ownership claim. The administration claims that it isn’t just offering access to AI systems. Rather, it’s presenting a model in which infrastructure can be built and operated in-country under the partner’s authority. In Kratsios’s framing, that means sovereign infrastructure, sovereign data, sovereign models, and sovereign policies within a country’s borders, under its government’s control — “[t]hey build it; it’s yours.” That’s a stronger proposition than a deployment license, even if the practical meaning of “yours” will depend on the commercial structure and the supply chain that supports it.
How seriously should partners take the “it’s yours” claim? They should take it seriously as a promise about where infrastructure sits and who operates it, but not as a promise of autonomy. In the same remarks, Kratsios ties the pitch to “secure and robust supply chains” and “world-leading American security.” And the administration’s broader strategy is explicit that these export packages are supposed to meet U.S.-approved security requirements and standards. So it seems that this is ownership-with-conditions — and partners should assume the U.S. retains leverage at the points that matter most for long-term capability: upgrades, replacements, frontier-model access, and compliance with evolving security requirements.
The second is a speed argument. Kratsios argues that countries face a choice between two paths to sovereignty, and one of them is a trap. Waiting to build the full AI stack domestically means falling behind at a moment when the gap between AI haves and have-nots is widening fast. The alternative is to adopt the best available stack now, and build national champions alongside it rather than instead of it.
The National Champions Initiative is pitched as a way to support that parallel track — so that, in Kratsios's framing, no country has to choose between capability now and an indigenous AI industry later. Saudi Arabia and the UAE are perhaps the clearest examples of this approach in practice — both have moved to import U.S. AI infrastructure at scale while building domestic AI industries and champions (Humain in Saudi Arabia, G42 in the UAE) in parallel. These are deals and structures that I examined in a recent CSET issue brief and they’re live experiments for this theory of parallel AI development.
For governments with the resources to sustain a parallel track, that tradeoff may be entirely rational — adopt now, hedge later. KSA and the UAE can afford to do both simultaneously. For countries facing the financing and capacity constraints Kratsios himself describes, the “hedge later” part is probably not for them. Instead, the financing and implementation toolkit Kratsios describes — DFC, EXIM, USTDA, MCC, a World Bank fund, a Tech Corps under the Peace Corps — seems to be designed for exactly these governments that are less likely to build up national champions.2
It's worth noting that Kratsios's ownership vision sits in some tension with the business models of the U.S. hyperscalers and frontier labs whose participation the AAEP depends on. Those companies primarily monetize through managed services and API access, not infrastructure transfers. Whether the sovereignty offer Kratsios described is one that the U.S. AI industry is actually prepared to deliver — and whether every partner wants that model rather than managed cloud regions or API access to frontier models — are open questions as the program takes shape.
What happens next?
One question the Kratsios package doesn’t answer is what happens after deployment — when the infrastructure is “yours” but the supply chain that keeps it current and functional is not.
The deployed infrastructure depends on continued access to U.S. chips for upgrades and replacement, U.S. frontier models for leading-edge capability, and a broader supply chain that Washington controls through export licensing. Ownership of the hardware doesn’t resolve the question of what happens when those upstream dependencies are subject to someone else’s discretion. Sovereignty exists — until something changes.
That’s what the sovereignty conversation in Delhi was actually about. Not whether countries can acquire and deploy AI systems inside their borders — most governments already assume they can — but how to reduce exposure to external discretion over time. What happens during a sanctions dispute, an export-control shift, a diplomatic crisis, or a change in vendor policy?
There’s also a scope question. The AAEP is a selective program for a curated set of government-backed deals, but most U.S. AI exports will continue to flow through ordinary commercial channels. The sovereignty concerns I heard in Delhi weren’t primarily about curated deals backed by the government. They were about the broader technology relationship — standard commercial terms, standard jurisdictional exposure, no continuity assurances. If the administration’s sovereignty framing is confined to the AAEP, it doesn’t reach the part of the relationship that most countries are actually worried about.
Closing takeaways
The most striking part of the summit was how quickly the ground has moved. Even among those who aren’t fully comfortable with the concept of sovereign AI, there was broad acceptance that managing dependency and disruption risk is now a legitimate state objective. I’d seen it in policy statements and in the data, but hearing it discussed with such depth and conviction was different.
There’s also something the AAEP can’t fix on its own, no matter how well it’s designed or implemented. The program doesn’t operate in a vacuum. It operates inside a broader relationship between the United States and its partners — and right now, that relationship is under strain. In several conversations in Delhi, the concern wasn’t really about the specifics of U.S. AI exports. It was about whether the United States is a trustworthy and predictable partner.
When that confidence erodes, how good the financing terms are or how capable the Tech Corps is matters less than it otherwise would. The sovereign risk calculation still points toward hedging. That impulse has structural drivers that would exist under any administration. But the current trust environment is pushing sovereignty strategies toward more defensive and autarkic forms than they might otherwise take.
The administration is engaging with the sovereignty shift rather than dismissing it, and the ownership and speed arguments Kratsios made are stronger than the sovereignty debate sometimes acknowledges. His basic diagnosis — that most countries cannot afford to wait for full self-sufficiency and that the gap is widening in ways that will be hard to reverse — is correct, and it's a point that sovereignty advocates need to engage with more seriously than many currently do.
But if the sovereignty-compatible label is going to stick, the offer will eventually need to address what happens after deployment: continuity, jurisdictional exposure, portability, and the political durability of the commitment itself — which is ultimately, at least in part, a question of trust and healthy relationships between the U.S. and its AI partners. And it will need a broader U.S. posture that gives partners reason to believe the commitment extends beyond one program and into the entire AI export ecosystem.
Conflict of Interest Disclosure: The views I share in this essay are my own. Although I hold professional affiliations and advisory roles with several organizations, I don’t receive compensation or instructions from them (or any other entities) regarding the views I express in essays, panels, or other public and semi-public forums unless otherwise stated in a specific piece or setting. In some cases, I might receive an honorarium from the publisher as payment.
There is an irony here. The AAEP's supporting tools — development finance, technical assistance, a Tech Corps housed in the Peace Corps — are analogs of the same soft power instruments the administration has dismantled or defunded elsewhere, most notably USAID. Reorienting U.S. development tools toward AI capacity-building is a defensible idea. But the disorderly dismantling of existing institutions has damaged the relationships and credibility that the new tools will need to succeed.