Renting AI from foreign providers is a national security risk, warns Cohere CEO

“This sentiment of renting AI from someone rather than owning it is a national security risk,” Aidan Gomez said. “You need to fully control it.”

Aidan Gomez, CEO of Canadian AI company Cohere, has spent years planning for the possibility that foreign governments might abruptly cut off access to frontier AI models. Earlier this month, when the U.S. government moved to restrict foreign access to Anthropic’s Mythos, it was a striking example of exactly what he’d been warning about.

“I think it’s woken everyone up to the reality, which is that centralized dependence on a single entity is a structural risk,” he told Fortune. “You can absolutely just have access revoked … and services shut down.”

The U.S. has largely led the global AI race since it began; controls vast amounts of the world’s AI computing power; and is home to the majority of leading frontier labs. China is its closest rival, and together the pair hold 90% of the world’s computing power, according to data from AI research firm Epoch AI.

Outside the two nations, the field of frontier AI developers is more narrow: Cohere, founded in Toronto in 2019, and Paris-based Mistral are among the only prominent companies building and deploying large-scale models from outside the U.S.-China axis. As AI becomes more deeply embedded in critical systems, such as hospitals, financial markets, military logistics, and public services, the question of who controls that technology and who can switch it off is becoming a matter of national strategy.

Gomez argues that democracies need to stop renting critical AI from a handful of foreign providers and instead build sovereign systems they can control end‑to‑end. In practice, that means owning or having some control over the full chain of components an AI system depends on: the data centers, the chips inside them, and the models themselves, rather than paying another country’s company for access to those things via an API (the software interface that lets you tap into someone else’s model over the internet).

“This sentiment of renting AI from someone rather than owning it is a national security risk,” Gomez said. “You need to fully control it.”

Cohere already sells its own models in a way that lets governments and companies run them entirely on their own infrastructure, Gomez said. Rather than sending data to Cohere’s servers, customers receive the model and the software needed to deploy it inside their own data centers, including on systems that are physically disconnected from the internet and to which Cohere has no remote access and no ability to revoke service.

Sovereign AI

“Sovereign AI” has become something of a catch-all term, often used to describe many different levels of independence. Some use it to mean controlling the data an AI system is trained on or processes; others mean the ability to build and train new frontier models from scratch; then some mean simply ensuring that sensitive information never leaves a country’s borders. The lack of a settled definition has become part of the political debate and can allow governments and companies room to claim progress on sovereignty without necessarily committing to any of its more demanding forms.

For Gomez, the central part of sovereignty starts with physical infrastructure.

“Domestically controlled infrastructure is the first piece of the puzzle,” he said. “Each country needs an infrastructure champion.”

Only when governments have a domestic player that actually controls the chips, power, and data centers a country’s AI systems run on, can they be sure that no foreign firm—or foreign regulator—can unilaterally pull the plug, he said.

However, the dream of fully sovereign AI inevitably runs up against hard constraints around compute, capital, and energy. Building and running frontier‑scale systems is extraordinarily expensive and resource‑intensive, and most countries will never be able to own every layer of the stack outright. Realistic sovereignty, Gomez said, has to be pursued through alliances. 

Rather than every government trying to replicate the whole ecosystem at home, Gomez wants democracies to pool their bets around a small number of “champions” and share capacity through trusted partnerships.

“It’s not the case that every country is going to have each layer of the AI stack inside their own country … Instead they need to find multiple partners for those layers,” he said. “If we try to spread our resources, it just won’t work. We have to back champions and not spread our bets.”

Cohere has been attempting to partner with other “middle power” nations and companies. In April, Cohere announced a deal to acquire Germany’s Aleph Alpha, with the aim of creating a transatlantic AI company anchored in both countries. 

Two months prior to the deal, Canada and Germany had signed a Joint Declaration of Intent on Artificial Intelligence and launched a sovereign technology alliance. The Cohere–Aleph Alpha deal was positioned as a way to put that political commitment into commercial practice. The combined company, which would retain the Cohere name and be led by Gomez, will be valued at around $20 billion, according to the Financial Times.

At last week’s G7 discussions, Gomez said leaders and AI executives largely agreed on the need for democratic countries to coordinate on standards and regulation. Now, he wants them to go further and invest in shared capacity, so their critical infrastructure isn’t dependent on a few U.S. or Chinese firms.

The politics around AI, he argues, are not that different from the technology itself.

“Centralized dependence on a single entity is a structural risk; we need to avoid these single points of failure,” Gomez said. “I’m a computer scientist, and that’s one of the first things you learn: Don’t build a system with a single point of failure.”

This story was originally featured on Fortune.com