The AI Power Paradox

In 2018, OpenAI's most advanced language model had 117 million parameters. Today, less than six years later, the most powerful models exceed a trillion. To give you an idea: the required computing power has grown tenfold every year for ten years in a row. We are facing a technology that is growing faster than any other in history, and the incredible thing is that the very speed of this growth is changing who is in charge in the world. It is always thought that only states have the power to regulate and manage decisive technologies. But with artificial intelligence, it's the opposite: it's companies — not governments — that have real control. And this is the real geopolitical revolution of AI. Until a few years ago, tech companies were considered mere service providers, tools in the hands of governments. Today, however, companies like OpenAI, Google, Meta, and a handful of others decide what their models can do, who can use them, how, and for what purposes. They have a power that previously belonged only to states. Mustafa Suleyman, one of the authors of the article and co-founder of DeepMind, was among the first to say that AI developers are "new geopolitical actors," with a sovereignty that derives not from the vote or the social contract, but from technical capability. And when U.S. Senator Ted Cruz says that Congress "has no idea what it's doing" about AI, he recognizes a huge power vacuum. The paradox is that the more powerful and accessible AI becomes, the more difficult it is to control. AI is not like nuclear energy: you don't need secret laboratories or years of research; all it takes is a code leak, and anyone can download advanced models from the web. Meta saw its Llama-1 model end up online a few days after its launch, and today versions that are almost as powerful run on a computer that you can rent for a few dollars an hour. Soon, sophisticated models will run directly on smartphones. This means that the slightest flaw in a country, a company, or a regulation becomes an open door to global risks — from mass disinformation to evolutionary digital weapons to threats on an existential scale, such as an AGI that is out of control. But the real question is: who can really govern this race? The classic answer — states — is no longer enough today. In the West, private companies have a freedom of action that no government can really limit. In China, where the state and companies are intertwined, control is tighter, but technology is moving too fast for it to really be kept in check there either. Meanwhile, the United States and China treat AI as a zero-sum challenge: whoever wins this race thinks they can dominate the century. But while they compete, technology spreads everywhere, and most countries will be forced to depend on whoever controls the most advanced models and chips. And here comes the twist: the solution is no longer to make rules between states, but to build a system in which even technology companies — whether we like it or not — sit at the global decision-making table. For the first time, the governance of a critical technology must be designed with those who develop it, those who own it, and those who use it, together with governments, experts, and civil society. The model of negotiations between states, as for nuclear power, is no longer sufficient, because AI is too easy to copy, too fast to spread, and too decentralized. We need "techno-prudent" governance, made up of agile, precautionary rules that are impervious to flaws and capable of adapting as AI evolves. There are three strong ideas: a global scientific task force, similar to the UN climate panel, to define and monitor the risks of AI; a non-proliferation regime — as for weapons — to prevent dangerous models from spreading uncontrollably among states, companies, and individuals; and a technological stability body, along the lines of the world's financial institutions, that can intervene in crises and impose universal rules. But none of this will work without the direct and responsible involvement of the companies that are the real referees of the game today, for better or for worse. The contrarian? We are used to thinking that the real danger is the big tech companies taking power. But there is an even greater risk: that no one really takes it, and we end up with technology without brakes, without rules, and without anyone able to respond when something goes wrong. In a world where just one "out-of-control" model is enough to create global damage, the danger is not only companies that are too strong, but also companies that are too weak or left to their own devices. If you want a phrase to remember: AI is not just changing what we can do — it is changing who is really in charge. If this idea has made you see AI in a new light, you can press I'm In on Lara Notes: it's the gesture to say that this perspective is now part of your way of thinking. And if in a few days you find yourself talking about it with someone while discussing who really decides the future of AI, on Lara Notes you can tag that person with Shared Offline — so there's a record of the conversation that matters. This Note comes from Foreign Affairs and saves you 21 minutes.