Power map ยท Updated June 2026

The Global AI Power Map

Not "who's winning" — a procurement map. Where the models come from, what they cost, and what each origin means for your data. The cheapest model in the world is $0.10 per million tokens. Here's the real price.

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 StatesChinaEurope
Leads onFrontier capability, computePrice-performance, open weightsSovereignty, compliance
2026 AI capex~$650bn (hyperscalers)~$98bnFraction of either
Data centres (2025)~5,427~449Limited
Flagship labsOpenAI, Anthropic, Google, Meta, xAIDeepSeek, Alibaba, Moonshot, Zhipu, MiniMaxMistral
JurisdictionUS (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

Running open models yourself? You'll need infrastructure — our sister site Best VPS Match covers that decision.