Emmanuel Macron is often accused of having a Jupiterian style of governance, but something different can be true of his manner in person. Watch him with his international peers and he is a figure of besuited bonhomie: a VP gladhanding at an offsite. Expect plenty of arm-clasping and thigh-slapping when he meets his British counterpart Keir Starmer today.
But watch him with Jensen Huang, as I did recently, and you will observe an unusual dynamic. The Nvidia chief executive acts not just as Macron’s peer, but even, in brief glimpses, his superior. When the two men were on stage together at a tech conference in Paris, there was a telling moment almost lost in an animated conversation. Huang, wearing his usual leather jacket and t-shirt pairing, talked over the President.
It was as if Huang were a statesman of slightly senior rank, an idea that seems less illusory with every year of Nvidia dominance. Arthur Mensch, the 31-year-old chief executive of Europe’s leading frontier AI company, Mistral, sat on the fringe of the conversation. Quiet, well-dressed, and intermittently talked up by Macron, he gave a stronger impression than Huang of being the President’s junior.
Huang’s assuredness is justified. Macron is keenly aware of the predictions that advanced AI will usher in a period of wealth creation unlike any the modern world has seen. He is also aware that with creations of wealth come redistributions of power. The French believe that if they are to be on the right side of these developments, they will need homegrown companies such as Mistral to compete with giants like OpenAI. And if Mistral is to compete, then it will need the Élysée to secure the physical hardware on which the company’s computation will take place and its data transferred.
It is Nvidia, of course, which sells that hardware: the silicon chips at the heart of this revolution. Imagine that, at the dawn of the fossil fuel era, there was a single company that was central to the piping and other infrastructure required to transfer oil, gas and coal and to burn them. Sure, Google and Amazon have their own serviceable chips; other companies are trying to develop their own. But Nvidia occupies a dominant position.
It is no wonder, therefore, that the California-based company has the second-highest market cap of any company in the world. And it is no wonder that Macron has championed Mensch and has helped broker a deal between Nvidia and parties including Mistral. As Macron said: “This is our fight for sovereignty, for strategic autonomy, which is to say: we want to keep that.”
The deal will ensure that Mistral is one of the beneficiaries of Nvidia’s Blackwell platform: the new generation of microchips on which high-level computation can take place. This will allow Mistral to handle the duties that will be increasingly outsourced to it by the French state: bureaucratic, logistical, even military. Blackwell will also be the training ground for forthcoming iterations of Mistral’s large language models. Sure, the Élysée could contract an American company — OpenAI has begun hawking its services to entire countries — but that would risk France’s independence. Such a deal could ultimately be the equivalent of filling a civil service with foreign nationals who could be recalled without warning.
Even using ChatGPT puts French consumers downstream of decisions made by American developers. It was American developers, famously, who were responsible for the Gemini model that gave the impression that Vikings and Nazis alike were black. That kind of overbearing cultural engineering, though accidental and swiftly retracted, is a poor fit for a country that has a heavy-handed cultural programme of its own. Take the Académie Française, whose role of language custodianship — did I say “role” again? I meant rôle — has no equivalent in the Anglophone world.
Sovereign AI, then, makes philosophical and pragmatic sense for a country that retains a defiant sense of selfhood. And as the data fed to AIs becomes increasingly voluminous and intimate, even as the AIs are given more comprehensive decision-making power in realms from utilities to the military, the task of implementing sovereign AI becomes ever-more urgent.
But is it really sovereign? There is an obvious irony in buying one’s sovereignty from an American company. Neither Macron nor Huang nor Mensch acknowledged that on stage. But the facts are plain: Donald Trump could, with a single executive order, prevent Nvidia from exporting chips even to Western allies such as Britain, France and South Korea. The world’s most powerful chips, those used within Nvidia equipment, are the product of near-superhuman manufacturing whose complexity and intricacy are unsurpassed beyond Taiwanese shores. Blackwell models include 208 billion transistors, each of which is smaller than a viral particle and etched by lasers with nanometre-scale precision. This is some of the most technically demanding engineering on the planet.
Very few countries, even with sustained national effort, could emulate the performance of those GPUs with an in-house substitute. So far, only China comes close. And the US and China are the only two countries that can already claim to be anywhere near AI sovereignty, developing both hardware and software either on their own shores or via domestically headquartered companies such as Nvidia. True AI sovereignty, for almost every other country, is therefore something approaching a mirage — at least within the current technological paradigm.
“True AI sovereignty is something approaching a mirage.”
The same is true, though, of almost any given technology. Prosperity depends on trade; trade depends on interdependence. The pipework of AI, so to speak, happens to be accessible only via an unusually small number of vendors. But the impracticability of perfect sovereign AI does not mean that Macron and his peers should not strive for it. There are relatively few vendors of oil and gas, and total energy independence is very difficult, but it is still worth aiming for. Partial dependence is much more desirable than total dependence.
There remains the uncomfortable question of what will become of the countries that fail to secure any meaningful kind of AI sovereignty. They are likely to outsource some amount of decision-making to foreign tech companies, and this decision-making will be inflected by someone else’s values. The decision-making would be based on data that might be held in foreign servers. One can imagine President Donald Trump Jr, a couple of election cycles hence, demanding access to a third-world country’s cobalt mines, threatening to turn off this country’s access to its own data. More subtly, a leading power might use other levers: limiting, for instance, a rival’s speed of access to the latest AI models. No less significantly for such a country, its selfhood will inevitably be eroded if it is governed by algorithmic values that are not quite its own. As Nathan Beniach, an investor and the author of the “State of AI” report, tells me: “For countries that opt out of building strategic AI infrastructure, the risk isn’t immediate subjugation but progressive loss of agency, what I’ve called a form of AI colonialism. These nations will increasingly find their economies, healthcare systems, security apparatus, and political discourse shaped by foreign models they neither control nor fully understand.”
The classical nation-state model, says Benaich, will struggle to adapt to these conditions. What emerges from the “sovereign AI” sales pitch, he suggests, is “a kind of neo-mercantilist compute empire, whereby Nvidia occupies the position of upstream monopolist, extracting rents across multiple national industrial strategies”.
Along the way, Huang, who is already worth £93 billion, will become even wealthier. Nvidia, whose stock price jumped on the day of the Paris event, is on course to become even more dominant than it is already; Apple and Microsoft will have to innovate rapidly in order to keep up. Anyone fearing an eternal monopoly, though, should note the rapid outmoding of the Google search engine, once seen as unassailable, by today’s youthful cadre of large language models. It is conceivable that, even without the advantage of Chinese IP theft, someone will make cheaper versions of Nvidia hardware.
And so the scramble for digital sovereignty begins. Last week, the British government called for “high-ambition researchers” to come up with uses for its new supercomputer, combining AI with robotics, biosciences, materials discovery and national security. Meanwhile, the President of South Korea has announced a $735-billion investment plan that will keep the country’s AI South Korean, and has been harvesting leading tech executives to steward the project.
Richer and more assertive countries, such as France, will retain some level of independence, even if intelligence, power and wealth begin to concentrate more intensely than they do already. Britain has catching-up to do; having allowed Google, as it was then, to buy DeepMind, we have no home-grown frontier AI company to rival Mistral.
If the doomers are correct, the ongoing concentration of intelligence and power is freighted with danger. The big fish will get bigger, the small fish will get smaller, and Starmer and Macron have a limited amount of time in which to find their countries’ place in this precipitous new hierarchy.
