01 · Nordic economies 🇫🇮 🇸🇪 🇳🇴 🇩🇰
Low power cost meets efficient cooling
Finland, Sweden, Norway and Denmark combine relatively inexpensive industrial electricity, cool climates, reliable grids and access to imported frontier hardware.
The new cost of national competitiveness
Which country has the cheapest AI?
The Kimi K3 moment made the thesis concrete: model monopolies will not endure. As capable models converge in quality, the cost of inference infrastructure becomes key to future national prosperity. The AI Cost Index compares that cost across 45 countries on a like-for-like basis. Methodology and data updated monthly.
Cost of frontier inference by country
Switch between two deployment sizes, three production workloads and annual economic conditions from 2010 to 2026.
General inference
8k input · 2k visible output · ≤5 s TTFT · 100 MW data center · 2026 conditions
Years before 2026 reprice the fixed 2026 service and hardware basket under historical economic conditions. Read the methodology
Fixed-basket AI ownership conditions
A fixed 2026 hardware and service basket isolates electricity, financing, construction, grid and operating conditions. Accelerator and memory market cycles are excluded from this view.
Lowest-cost year2020Median index 93.1
Highest-cost year2023Median index 100.9
2026 median cost$25Per million output tokens
Change since trough+7.4%Cross-country median
45 countries Cross-country median All countries
What the first backcast says
The lines show how the modeled cost changes within each country. The cross-country median falls through much of the 2010s, reaches its low point in 2020, and then rises as financing, electricity and construction become more expensive. The modeled median peak is in 2023. Conditions in 2026 remain about 7% more expensive than the 2020 trough.
Median cost components in 2026
- Hardware
- $18
- Electricity
- $2
- Site and grid
- $2
- Operations
- $3
Constant 2026 USD per million output tokens. Hardware capability and real purchase price are fixed, so accelerator and HBM scarcity cycles are excluded. Historical methodology
Why countries rank where they do
Electricity matters, but hardware access, financing, cooling, utilization and construction can matter more.
02 · United States 🇺🇸
Deep capital and hardware markets
The United States benefits from inexpensive power, accelerator availability, vendor support, construction experience and deep capital markets. This cost ranking does not measure its much larger installed compute base.
03 · China 🇨🇳
Competitive, with greater uncertainty
China combines inexpensive power and lower operating costs with state support for domestic hardware. Its result depends on an unverified Huawei-equivalent capital and throughput normalization, so it receives the widest uncertainty grade.
04 · United Kingdom 🇬🇧
High electricity cost weakens the result
The United Kingdom ranks 36th of 45 in the default scenario. Its modeled large-load electricity input is the highest in the dataset. Its absolute disadvantage is less dramatic than the rank suggests: $26 per million tokens, compared with $23 in Finland.
05 · Brazil 🇧🇷
Capital cost dominates the estimate
Brazil's result combines a 25% landed-hardware premium, the highest financing assumption in the dataset, warmer-climate cooling, lower modeled availability and higher site and grid annualization.
The central hypothesis
AI access will become critical national infrastructure.
Models are converging in quality. As the gap closes, inference becomes a utility like electricity, cloud infrastructure and telecommunications. Price, abundance, reliability and distribution will determine national competitiveness.
Countries that can deliver abundant, reliable and low-cost intelligence will outgrow those that cannot. The AI Cost Index measures the production cost of that intelligence.
Read the national competitiveness argument