The real
cost of AI

The sticker price of a token is not the cost of an answer.

What it actually costs to build and run on AI in August 2026 — inference economics, GPU ownership, model selection, routing, fine-tuning, and the 24-month forecast — carried with its evidence and its boundaries.

Start from your decision Read the whole report
  1. orientationThe short version
  2. evidenceInference economics — five verified claims
  3. evidenceGPU ownership — three verified claims and a crossover model
  4. decisionIf you build with AI APIs
  5. decisionIf you run ML infrastructure
  6. decisionIf you budget AI spend
  7. forecastThe 24-month horizon — six dated bets
  8. decisionThe routing layer itself — thin fees, thin moats
  9. decisionIf you're evaluating fine-tuning vs retrieval
  10. boundaryEvidence boundaries
How to read thisTen sections, one report. Each has its own URL, its own bounded answer, and the boundary that reopens it.
Evidence reviewed through
August 20, 2026
Recheck by
November 20, 2026
Report sections
10
Audience routes
7

What this site is

One maintained synthesis of what it costs to build and run on AI, carried onto the public web with its evidence boundaries intact. Prices here are dated evidence of posted rates, not durable claims. Forecasts are stated as dated bets with the conditions that annul them, not as curves. Where the answer is unmeasured, this site says so rather than estimating.

Start from your decision

All routes

Each route composes the actual report sections in the order that decision needs them, and states the proof standard it holds them to.

The short version

All 10 sections

The four findings that change a cost decision before any per-model comparison starts.

What the evidence establishes

All 10 sections

The verified claims behind the headline numbers, notebook by notebook, with their corroboration.

What this means for decisions

All 10 sections

The same evidence, arranged by the decision it is supposed to inform.

Evidence boundaries

All 10 sections

What this evidence does not establish, and the conditions that reopen each answer.

What readers ask

The full ledger

Eight cross-cutting questions, each with what the evidence currently answers and what is still open.

The whole argument, in one document

The maintained synthesis: what a token costs, when owning GPUs beats renting, what routing is worth, when fine-tuning pays back, and six dated bets on the next 24 months. Published with its update log and a clean Markdown source for machine readers.

Read the report Markdown source