The headline: the US leads on frontier capability and outspends everyone (~$650bn 2026 capex vs China's ~$98bn). China leads on price-performance and open weights — models at 75–85% of GPT-4o quality for 10–15% of the cost. Europe leads on sovereignty and compliance. Your choice depends on which of those three you're actually buying.
The three blocs at a glance
| United States | China | Europe | |
|---|---|---|---|
| Leads on | Frontier capability, compute | Price-performance, open weights | Sovereignty, compliance |
| 2026 AI capex | ~$650bn (hyperscalers) | ~$98bn | Fraction of either |
| Data centres (2025) | ~5,427 | ~449 | Limited |
| Flagship labs | OpenAI, Anthropic, Google, Meta, xAI | DeepSeek, Alibaba, Moonshot, Zhipu, MiniMax | Mistral |
| Jurisdiction | US (CLOUD Act) | China (data stored in China) | EU (GDPR-native) |
The US: spending at staggering scale
America's main hyperscalers — Alphabet, Amazon, Meta, Microsoft — plan around $650–700 billion of data-centre spend in 2026 alone. Microsoft spent ~$80bn on AI capex in 2025; Amazon projects ~$200bn for 2026. The US had ~5,427 data centres in 2025 and consumed ~45% of the 415 TWh of electricity data centres used globally. This is the capability and infrastructure lead — and most enterprise AI runs here, under US jurisdiction (including the CLOUD Act).
China: less spend, astonishing efficiency
China's total AI capex could reach ~$98bn in 2025 (Alibaba alone is committing $53bn+ over three years). The investment gap with the US is enormous — yet Chinese labs ship models at 75–85% of GPT-4o quality for 10–15% of the cost. That ratio is the story. The how is Mixture-of-Experts architecture, and the twist is that US chip export controls forced the efficiency that became the advantage.
Europe: a sovereignty play, not a capability race
France's Mistral is the EU's serious entrant — but at a ~$14bn valuation, roughly 2% of OpenAI's. Big Tech's projected $700bn 2026 AI spend exceeds the total valuation of all European AI startups combined. Europe isn't trying to win on raw capability; it's positioning as the GDPR-native, sovereignty-first, government-trusted option. For regulated EU businesses, that can matter more than benchmark scores. See the Europe landscape.
The editorial line
The cheapest AI model in the world costs $0.10 per million tokens. Here's the real price: data stored under Chinese jurisdiction, built-in content censorship, and self-hosting complexity to deploy safely. The low number is real. The full cost is what nobody else shows you — laid out in the risk assessment.
Make it a decision
- Buying capability? US frontier labs — see the comparison table.
- Buying price-performance? Chinese models guide — with the risk framework.
- Buying sovereignty? Mistral and the EU, or self-host open weights.
- Unsure where data goes? The data sovereignty comparison.
Running open models yourself? You'll need infrastructure — our sister site Best VPS Match covers that decision.