Inference Economics
Underwrite the facility as a token factory.
Levelized cost per million tokens, tokens per watt, cost shares, and breakeven utilization, computed from the assumptions you submit.
Method
Built from inputs you supply
The result comes from a visible calculation, not a forecast.
- 1
Submit assumptions
You provide accelerator count, power draw, utilization, throughput, energy price, and lifetime. The calculation does not use hidden defaults.
source_mode: submitted
- 2
Check the calculation
The same inputs always produce the same output. There is no model in this path and nothing to drift between runs.
Pure function, no inference step
- 3
Return the breakdown
The result includes cost per million tokens, tokens per watt, cost shares, and breakeven utilization.
Full breakdown, not a single figure
- 4
Show the sources
The result labels submitted assumptions and dated reference data separately.
source_mode + freshness declared
Outputs
What comes back
Levelized cost per million tokens
Total lifecycle cost divided by lifetime token output, expressed in the unit an inference buyer negotiates in.
lcoi_usd_per_million_tokens
Effective tokens per watt
The efficiency figure that lets a facility be compared against another on energy terms instead of headline accelerator counts.
effective_tokens_per_watt
Cost share breakdown
Where the cost actually sits across hardware, energy, and operations, so the sensitivity that matters is visible.
Cost shares returned per component
Breakeven utilization
The utilization rate at which the deployment reaches breakeven.
breakeven_utilization_pct
Token price snapshot
A dated snapshot of published token prices, labeled with its source and as-of date. It is not a live feed.
Dated and attributed, not streamed
Inputs clearly identified
Each result separates your assumptions from dated reference prices.
source_mode on every response
Scope
Unit economics run on submitted assumptions and are labeled as such. The token price snapshot is a dated, attributed reference, not a live market feed, and nothing on this surface should be read as a price forecast.
Questions
Scope and limits
Is this a forecast of token prices?
No. The calculator uses assumptions you submit. The token-price reference is dated and attributed, and is not a forecast. Illustrative output. Not investment, financial, or tax advice.
Why does this matter for a facility, not just a model provider?
Because AI facilities are being underwritten as real estate when they behave like token factories. Cost per million tokens and tokens per watt give a facility and a workload a shared basis that megawatts alone do not.
Can I call it programmatically?
Yes. Unit economics and the token-price reference are available through the Data API.
Read the API documentationHow does this relate to the LCOE calculator?
LCOE prices the energy going in; levelized cost of inference prices the tokens coming out. They use the same project-finance discipline at opposite ends of the same facility, and the LCOE calculator is free and needs no account.
Open the LCOE calculatorRun it against a deployment you are pricing.
Bring your accelerator mix, power terms, and utilization assumptions. The calculation shows which component dominates cost.