The Great Pyramid, finished around 2560 BC, was the tallest structure built by human hands for the next 3,800 years — built in two decades on a claim against Egypt's entire Nile grain surplus, a feat of sovereign-class power no merchant or private fortune of the era could have matched. China's Great Wall took centuries and several dynasties instead of two decades, but drew on the same kind of power: a state granary deep enough to outlast generations. Forty-five centuries after the pyramid, the United States organised a hundred and thirty thousand workers and roughly two billion dollars — close to thirty billion today — to build the first atomic bomb in under three years. A pharaoh, an emperor and a wartime president all reached for the same two things to build at that scale: a treasury large enough, and a throne with the authority to spend it. A treasury has always meant a flag.
Frontier artificial intelligence is the first one where the state never held that position to begin with. Anthropic and OpenAI alone have raised more than $300 billion in private capital to date — the kind of sum that, a generation ago, only a state could have raised — and it is coming almost entirely from venture funds, sovereign wealth funds, public listings and multi-year hyperscaler partnerships, racing each other to be first. The institutions building the most consequential capability of this century answer to shareholders and term sheets. Not one of them was elected. Not one of them governs by constitution.
States have noticed, and they are responding the only way the structure of this moment usually allows: from the outside in. Export controls on chips. Subsidies for fabrication plants. National strategies, procurement, security reviews. Last week the US government went further still, and for the first time put a frontier model itself under export control, ordering Anthropic to suspend all foreign access to its Mythos system over an unexplained national-security concern. Even that reach stopped at an off-switch. It could not touch how the model was built, trained, or governed. The labs still hold the thing itself.
The conversation about artificial intelligence still happens mostly in the comfortable language of augmentation — a copilot, an assistant, a tool that makes a lawyer faster or a doctor more accurate. That description is accurate for now, and it is also a description of a system still mid-lesson. The honest forecast is not that AI keeps making every profession marginally more productive. It is that AI will, within a horizon nearer than most professions are prepared for, perform — without exception — the tasks inside every job on Earth: the legal reasoning of lawyers and the judgment of judges, the diagnosis of doctors and the bedside care of nurses, the composition of designers and the calculations of architects, the persuasion of sales and the coordination of operations, the campaigns of marketing and the policy work of human resources, the wiring of engineers, and the pipe-fitting of plumbers too. No occupation has a durable argument for exemption.
There is a second moat forming underneath all of this, and it has nothing to do with chips. It is memory. The more of a person's history, preferences and half-formed thinking lives inside one AI's context, the more expensive it becomes to leave — the same lock-in as one bank or one inbox, except sharper, because what's stored this time isn't a transaction history but a working model of how someone thinks. Switching providers will start to feel less like changing software and more like starting a relationship over with a stranger.
Push that lock-in to the scale of a population and something new appears. Citizens of one country already see events differently depending on which news they trust; add an AI that doesn't just deliver information but reasons alongside each of them — drafting arguments, answering a child's homework question, shaping what counts as a reasonable conclusion — and the divergence stops being about which channel someone watches and becomes which mind is doing half their thinking. Half a nation routing its daily judgment through a system trained and aligned in one country, the other half through a system built somewhere else, shares a passport and little else. Call it an AI-sphere: a population grouped by whose model it thinks with, not where it lives.
And yet none of this — the jobs, the memory, the new lines humanity may end up organising itself along — is being built by a sovereign. No pharaoh, no emperor, no president. Entrepreneurs and private capital are building it instead.
Every sovereign-class power in history was forged by a state. AI is the first one built without one.
Humanity may end up coalescing less around the borders it has always organised itself by, and more around the handful of foundation models it has chosen to think with — doing every job, holding everyone's memory, sorting populations along new lines, all of it built by people who answer to a cap table rather than a constituency.
Governments still have a narrow window to change which side of that sentence they're on. Owning a credible stake in sovereign AI infrastructure is no longer a research-funding line item. It belongs in the same category as currency, military, or food supply, and should be treated with the same urgency. Owning none of it is not a competitive disadvantage. It is existential.
P.S. — It isn't only governments missing this moment. The trillion-dollar incumbents built over the last two and a half decades — Microsoft, Google, Meta, Amazon — hold more capital and more compute than almost anyone alive. But a company with a hundred thousand employees and a decade-old stock price it is still defending is rarely where the next disruptive idea gets built.
My prediction: the Big Tech of the last three decades will not be the Big Tech of the emerging AI century. Anthropic is already valued near a trillion dollars and OpenAI close behind it, both built in roughly a decade, both generating more revenue per employee than any public technology company on Earth with a fraction of the headcount. DeepSeek did something comparable from a standing start, on a fraction of the budget. PhotonAI is attempting the same thing for an entire continent. None of them have a legacy product line to protect, a board to talk out of moving fast, or twenty years of organisational weight to carry into the next decade. Big tech may turn out to be exactly as exposed to this transition as the nation-states this letter has been describing — just slower to notice, because its balance sheets are still going up.