Published cost model · Reviewed 23 August 2026

A million tokens costs
$0.2460 to serve.

That is the energy, hardware and infrastructure cost of one million inference tokens on geothermal power in Kenya. The same workload at an EU hyperscaler costs $0.4285, which is 74% more. Every input is below, so you can check our working rather than take our word for it.

Cost per million tokens

Where the money actually goes.

Three components, costed identically across markets. Energy is the line that separates them.

Market Energy Hardware Infrastructure Total kWh vs Navon
Navon Hells Gate $0.0215 $0.1510 $0.0735 $0.2460 0.239 baseline
EU Hyperscaler $0.0553 $0.1951 $0.1781 $0.4285 0.299 +74.2%
Emerging Market $0.0454 $0.2349 $0.1608 $0.4410 0.378 +79.3%

Figures are cost to serve, before margin. At our default 40% target margin the Navon price works out at $0.4100 per million tokens.

Show the working

Every input, stated.

Change any of these in the calculator and the numbers above move with them.

Input Navon Hells GateEU HyperscalerEmerging Market
Energy price ($/kWh) 0.090.1850.12
PUE 1.291.51.4
Hardware cost ($ per GPU) 150001800016000
Hardware life (years) 333
Power draw (kW per GPU) 0.70.70.7
Throughput (tokens/sec) 150015001200
Utilisation (%) 706560
Infrastructure ($/month) 200450300
Method and caveats

What this is, and is not.

This is a reference model, not a quote. It assumes steady-state serving of a single model class and excludes networking egress, support and anything specific to your deployment. The EU and emerging-market columns are representative figures for comparison, not a specific competitor's price book.

We publish it because almost nobody publishes cost-per-token for African compute, and an argument you cannot check is not worth much. If you think an input is wrong, tell us and we will correct it.