Calculation methodology, assumptions & exclusions +
Cloud costs use uncached input, cached input, and output token rates per million, multiplied by tasks per month, plus platform and operating fees. Cached tokens are a subset of input tokens.
Owned costs include device purchase price multiplied by count and allocated over useful life; setup allocated over its stated period; software, support, power, and operations. All fees except purchase price are fleet totals. Unknown software or support values are excluded from the subtotal and block a savings claim.
Measured qualifying tasks per device-hour determines capacity and active hours per device. An unmeasured scenario uses entered active hours and does not establish feasible savings. Multiple devices assume independent work divided equally.
Electricity = (active watts × active hours per device + idle watts × idle hours per device) ÷ 1,000 × price per kWh × device count. Cash-flow payback uses upfront hardware plus setup divided by positive operating savings; monthly amortization is excluded from that denominator.
Taxes, financing, residual value, cache-write/storage fees, network charges, hardware downtime, model maintenance and quality differences are not modeled unless incorporated into entered costs. No named provider pricing is fetched. This is a cost-only scenario, not proof of equivalent task quality.
Prepare the evidence behind your scenario with the Inference Cost Audit Kit, a Happiest Labs resource from Inference Savings. Its blank worksheets help connect actual costs, task acceptance criteria, and removable expense.
Ownership is about
more than a cost line.
Privacy, control, and traceability are operating requirements. Their relevance depends on your documents, your deployment, and the decisions you need to make.
Inspect the Trust Center ↗