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 measures all three blocs identically.

Three blocs, three completely different bets on the same technology. Worth asking what each one is actually buying.

The US bet: spend to lead, worry about the bill later

America is deploying capital at a rate with no real precedent outside wartime industrial mobilization: 2.36% of GDP into AI infrastructure in 2026, heading toward 3.1% in 2027, plus $1.65 trillion in off-balance-sheet liabilities parked in special purpose vehicles behind five companies¹, about 5% of GDP and eightfold growth in four years. Put that next to the reference point people usually reach for when they want to describe a spending program as civilization-scale: the Manhattan Project cost about $1.9 billion in 1945 dollars, roughly $30 billion in today’s money, and consumed around 0.4% of GDP at its peak¹⁴. 2026 AI capex alone is already six times that as a share of the economy, sustained year over year rather than concentrated in a single wartime push, and that’s before counting the off-balance-sheet debt. The bet is straightforward: whoever owns the compute and the frontier models first captures the economics of the transition. China is now attacking that pricing-power assumption directly.

China’s replication play: turn software into a commodity

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, roughly a twentieth of the cost¹¹. Moonshot’s Kimi and Zhipu’s GLM have kept the pattern going since. The gap to the US frontier now runs roughly 6 to 18 months, not years.

The interesting part isn’t the speed, it’s the pricing. There’s no patent regime protecting a transformer architecture the way one protects a drug molecule, so once a capability is matched, Chinese labs can price the workhorse tier of generica AI just above the cost of inference: Kimi K2.6 runs $0.95 input / $4.10 output per million tokens¹³, cheap enough that no Western lab can price there and still recoup what it spent training the model. The weights are open for now, but that’s a choice, not a structural feature; nothing stops the next generation from closing¹⁰.

Strip away the “race to the bottom” framing and the actual plan is margin destruction: push AI pricing down toward the cost of inference and the industry stops looking like software and starts looking like chips and electricity, terrain where China’s manufacturing base gives it a structural edge over both the US and the EU.

None of that makes these labs mere generica factories. On raw capability, Chinese labs are close behind the US frontier, and in a few narrow areas ahead¹².

Europe: a user of AI, not a builder of it

Mistral exists less as a competitive bet against frontier labs and more as an insurance policy, keeping the institutional knowledge of how to train a large model alive inside the EU even though it isn’t close to state of the art. The practical European AI stack, for most companies, is increasingly: rent US compute, or download an open Chinese model and run inference locally.

That second option comes with a wrinkle worth naming plainly. Chinese open models carry the training data and alignment choices of their origin, including gaps and refusals around politically sensitive topics like Tiananmen Square. A European fine-tune to patch those gaps is a solvable engineering problem, not a moat. And as long as the weights stay open, there is no regulatory catch either: a model downloaded and served on European hardware never sends user data anywhere, so GDPR simply doesn’t enter the picture. A European company gets frontier-adjacent capability without having paid anything close to frontier-level development cost, at the price of depending on someone else’s base model.

The dependency only develops teeth if Chinese labs close their next-generation weights, which is exactly what the scenarios below with remaining headroom would incentivize. At that point the cheap option becomes Chinese-hosted inference APIs, and there the regulatory constraint bites hard: EU companies handling user data generally can’t route it through inference inside China without triggering serious GDPR exposure. Whether a compliant intermediary structure could exist is a policy question, not an engineering one, and it’s currently unresolved. Open weights make Europe’s free ride legally trivial; closed weights make it legally impossible on current terms.

How this ends depends on how much further the curve bends

If current performance is close to the ceiling. Everyone captures roughly the same capability, and the game becomes entirely about who paid least to get there: no further step-change, just diminishing returns on the current architecture and incremental efficiency gains. None of the capability turns out to be exclusive, so none of it commands a durable premium. This is the scenario the commoditization thesis above points toward directly, and it’s the one that makes the current US capex trajectory look worst in hindsight: $1.65 trillion in off-balance-sheet debt written against a pricing regime that never materializes, and some of that debt doesn’t get repaid on the terms it was written.

If there is a moderate amount of headroom left. Whoever can still push the frontier keeps some pricing power, and the US retains an edge for a while, but the gap keeps compressing on China’s release cadence, which has consistently landed within 6 to 18 months of the last frontier move. Europe’s free ride only breaks if China closes the next-generation weights, and China is already positioning to be able to do exactly that. If that happens, the real question becomes whether Mistral and European academic and public research capacity can actually execute a fast-follow strategy the way China has, whether Europe can replicate the replication. Nothing in the current spending trajectory suggests that capacity exists yet, but it’s a spending choice, not a law of nature.

If there is substantially more headroom than the current frontier suggests. This is the scenario where forecasting gets genuinely hard, and where the answer stops being about capex ratios and starts being about who can sustain compounding research and infrastructure investment over another decade. Whether Europe attempts its own version of the US-China replication dynamic, or accepts permanent dependency on one of the other two, becomes the central strategic question. Nothing about the current European posture answers it either way; the current spending level is consistent with a temporary holding pattern or a permanent one, and which it turns out to be is a policy decision that hasn’t been made yet.

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

Under a nothing-ever-happens prior, low headroom left, the EU comes out on top: smallest investment by a wide margin, full access to wherever capability settles. China is also a winner, partially by ending up in a good spot at still low cost, and partially by having substantially harmed the US. If moderate headroom remains, China likely comes out ahead: it keeps closing the gap at a fraction of US cost while eroding US pricing power with every release. If the headroom is substantial, the US wins, and the capex and the hidden debt turn out to have bought exactly what they were priced for. The last two calls are far more speculative than the first; the first only requires the world to keep looking the way it currently does.