What does a token really cost to serve, and what does a megawatt earn?
Model one accelerator or one megawatt end to end: energy per token, fully loaded cost per million tokens, margin, payback, NPV and IRR, break-even utilization, carbon, and a side-by-side comparison of current hardware.
Set the assumptions. Results update as you type.
Pick an accelerator, choose whether to estimate throughput from a model or enter a measured figure, then set power, hosting and pricing. Everything here is an editable assumption.
One accelerator, one month.
Revenue less energy, hosting and other operating cost gives EBITDA. Depreciation on top gives operating profit. NPV and IRR use monthly cash flows over the useful life plus residual value, before tax and financing.
Energy, speed and carbon per token.
Joules per token is the cleanest cross-hardware efficiency measure because it folds throughput, utilization, idle draw and facility overhead into one number.
How low can utilization or price go?
Everything except utilization is held at the values above. Idle power is charged at any utilization.
What one megawatt of facility power yields.
Every accelerator on the same assumptions.
Each row runs the throughput estimator for that hardware with your model, precision, batch, pricing and operating settings. Click a heading to sort. Hardware that cannot run the chosen precision or fit the model is listed at the bottom.
Throughput is a first-order estimate, not a benchmark. Measure your own stack and switch to manual throughput for decisions.
Own it or rent it.
Reference prices and cloud rates are illustrative placeholders, not quotes. Replace them with real vendor and provider pricing.
How the numbers are built.
Hardware specifications are approximate public figures and may differ by configuration. Reference prices and cloud rates are illustrative and change often. Model estimates ignore prefill cost, speculative decoding, batching policy, network limits and software maturity. Results are for planning, not a quote or a guarantee.
NPV and IRR are pre-tax and unlevered. Carbon uses a single grid factor and excludes embodied emissions.
Want these numbers checked against real capacity?
We source powered, GPU-ready capacity and help teams compare options on cost per token, not on the pitch. Tell us what you are serving and we will bring back options.