AGI Won't Come. Then What?

Metric	US	China	EU
2026 AI capex (hyperscalers / big tech)	$764bn35	$102bn2	~$11.6bn (sovereign programs)4
Capex as % of GDP	2.36%8	0.49%8	0.05%9
2027E capex	$1,018bn3	$123bn2	n/a
2027E capex as % of GDP	3.14%8	0.59%8	n/a
2025 private AI investment (Stanford HAI)	$286bn7	$12bn7	$21bn7

Capex rows are hyperscaler spending for the US and China, sovereign programs for the EU (there are no EU hyperscalers to count; broadening to all European AI server spend gives ~$47bn⁶). The Stanford HAI row is the one that actually measures all three blocs the same way.

The US, China, and the EU are placing three very different bets on the same technology. They can’t all be right.

The US Is All In

The US is spending money on AI like nobody has ever spent money on anything. AI infrastructure alone eats 2.36% of GDP this year, a projected 3.1% next. And those are just the numbers anyone can look up.

Off the books, it’s worse. Five companies have parked $1.65 trillion in special purpose vehicles, eight times what it was four years ago.¹ Nobody knows how that gets unwound if the bet doesn’t pay off, probably not even their own finance departments.

For scale: the Manhattan Project, adjusted for inflation, cost about $30 billion total and peaked at 0.4% of GDP.¹⁴ That was the entire wartime bomb program. The US is now spending roughly six times that share of GDP on AI, every single year, with no end date attached. At least the Manhattan Project knew what it was building.

Why spend like that? Because if AI really does eat the economy, whoever owns the models and the datacenters when it happens gets to charge everybody else rent forever. If you believe that, no price is too high. The US believes it.

China is turning AI into a commodity on purpose

China replicated the US frontier for a fraction of the price: DeepSeek trained a GPT-4-class model for a disclosed $5.6 million, against the $100 million+ reported for GPT-4 itself.¹¹ Moonshot’s Kimi and Zhipu’s GLM were built similarly cheap. China is on track to trail the US frontier by about 6 to 18 months, on roughly an eighth of the US capex. You can’t patent a transformer architecture, or the parts of the LLM training pipeline, the way you can patent a drug molecule. So once a capability is matched, Chinese labs can price the workhorse tier of generic AI just based on the cost of inference. No Western lab can match that price and still recoup what it spent training the model, but Chinese companies can price there just fine.¹³ The weights are open for now, but nothing’s stopping the next generation from closing up and going proprietary.¹⁰

Their play: make the AI battle merely about the underlying chips and electricity, margin destruction as the whole strategy. And once the AI battle is a manufacturing battle, China has already won. Everyone else is still arguing about who has the best model.

Moonshot and Zhipu ship real research of their own too, and in a couple of narrow areas the US is the one trailing.¹²

Europe rents its AI

Sorry, but Mistral is just an insurance policy: it keeps the institutional knowledge of how to train a large model alive inside the EU, nowhere near state of the art but alive. As a competitive bet against the frontier labs, nobody serious is counting on it. Ask a European company how it actually gets its AI and the answer is a little embarrassing. Pay a US company by the token. Or download a Chinese model for free and run it wherever’s cheapest, which, again, is usually an American server.

As long as the weights stay open, there’s no regulatory catch either. Download the model, serve it on European hardware, and user data never leaves the building. GDPR just doesn’t come up. A European company gets near-frontier AI without ever paying for a training run, in exchange for depending on someone else’s base model.

That dependency only starts to hurt if Chinese labs close their next-generation weights, which is exactly what the scenarios below with remaining headroom would push them to do. At that point the cheap option becomes Chinese-hosted inference APIs, and there the rules bite hard: EU companies handling user data generally can’t route it through inference inside China without getting into serious GDPR trouble. Maybe some compliant middleman setup could patch that; that’s a question for lawyers and politicians, and nobody’s answered it yet. Open weights make Europe’s free ride legally trivial; closed weights make it legally impossible on current terms.

A look into the crystal ball

How much of this actually plays out depends on how much headroom is left in the current architecture.

Close to no headroom left: everybody lands in roughly the same place, capability-wise, and the whole game collapses into who paid the least getting there. Just diminishing returns from here on out. Whatever edge someone has doesn’t last, and once it’s gone so is the premium price tag that came with it, the commoditization thesis from earlier playing out for real. It’s also the scenario where the current US capex trajectory looks worst in hindsight: $1.65 trillion in off-balance-sheet debt written against premium prices that never show up, some of it never repaid on the terms it was written. The EU wins here by default, smallest spend by a wide margin and still full access to wherever capability settles, and China wins too, cheaply, while doing real damage to the US along the way.

A moderate amount of headroom left: whoever can still push the frontier keeps some pricing power, so the US holds an edge for a while, but China keeps closing the gap on the cheap, and every release inside its usual 6-to-18-month window eats into that edge a little more. Europe’s free ride only breaks once China closes off its next-generation weights, and China is already lining up to do exactly that. Can Mistral, plus whatever’s left of Europe’s academic and public research capacity, actually copy China’s homework the way China copied everyone else’s? Look at the money going in right now and the honest answer is no, though Europe could still grow out of that. China comes out ahead here, by a mile.

Substantially more headroom than the current frontier suggests: forecasting gets hard. The question stops being about capex ratios and turns into who can keep compounding research and infrastructure spend for another decade straight, whether Europe tries its own version of the replication play or just keeps renting from one of the other two forever. Nothing about Europe’s current posture answers that. Here the US wins, and the money it spent buys exactly what it was priced for.

And if we end up creating a god? Unlikely, but see my guide for the AI that finds itself in that position.

The first one barely requires anything to happen. The world just has to keep looking the way it already does. The other two are bets on how much room is actually left to run. We can’t predict how things will pan out; especially not those people currently spending the trillion and a half dollars.