The name is the metric.

Every other unit — racks, GPUs, capex dollars, lease months, list prices — is an input. Tokens are the output someone actually pays for. A plan that minimises chip count or looks great on a vendor quote but lands at $0.80/M-tok when the market clears at $0.40 is a losing plan, however elegant the spreadsheet.

And the number moves constantly: models reprice monthly, new silicon reshuffles the build-side economics, a lab's intro pricing quietly expires. A cost answer that was right in March is wrong by June. That's why Mintok is built as measurement infrastructure, not a calculator — catalogue watchers track every model and chip continuously, price history accumulates into time series, and benchmarks measure real cost, speed, and quality on real prompts rather than quoting list prices.

• Measured beats quoted. List prices tell you what a vendor hopes; a benchmark on your prompts tells you what you'll pay. Everything Mintok publishes is computed from live data, with the methodology in the open.
• Three constraints, not one. Compute, memory, and bandwidth all bind at different points across model + workload combinations. A sizing that ignores any of them ships the wrong fleet.
• Contract is part of the chip. The same accelerator on a 36-month lease, a cloud hourly rate, or owned capex is three different economics. Treating them as one chip is a six-figure mistake.
• One kernel, every surface. The $99 benchmark, the $699 sizing report, and the full platform run the same deterministic engines — buying a single answer and running the whole practice never disagree with each other.
• Answers should ship as artifacts. A shareable report, a versioned brief, working agent code — analysis is only useful if it flows to the next decision without re-keying.

Two ways in, one measurement engine underneath:

• Self-Service — buy the answer. Benchmarks on your prompts, instant sizing reports, rack teardowns, agent cost designs with working starter code, and the monthly AI Compute Cost Index — produced by agents, priced from $99, delivered in minutes. Browse the catalog.
• Platform — run the practice. For fleet operators and forward-deployed engineers: sizing anchored four ways, orders with allocation, delivery against real supply, and watchers that flag drift — invite-first while we scale onboarding. How it works.

The public face of the measurement layer is free: the model lineuptracks every major provider's ladder with recent price moves, and the calculator runs the sizing kernel with no signup.

Every number traces to a published methodology; modeled estimates say so, measured ones show their runs. Tenant isolation is enforced in the database itself — the details are on the security page. Reports are research, not investment advice.

General questions about Mintok? info@mintok.ai