Token margin

Live at https://tokens.ganeshkumarashok.com. What are the margins on inference? This models one NVIDIA NVL72 rack (GB200 or GB300) serving an open-weights model (GLM-5.3, FP4) for a day: what it costs to rent or own, how many tokens it serves, what they sell for, and what is left. Moving a slider recomputes the day, and the URL records every change, so each scenario is a shareable link.

Throughput is measured (InferenceX agentic benchmark), the token mix and prices are OpenRouter's, and rents are published list prices or filed contracts. Results are in docs/findings.md, including a comparison with DeepSeek V4 Pro, which the model supports but the site does not show. Every default lives in data/assumptions.json and traces to research/sources.md; the raw data behind them is in research/data/.

Run locally

npm install
npm run dev          # site at http://localhost:5173/
npm test             # engine tests, including checks against InferenceX's published points
npm run scenario     # scenario tables -> docs/findings-tables.md
npm run e2e          # browser checks, screenshots and public/og.png (first run: npx playwright install chromium)
python3 research/reference.py          # independent reference model the tests were derived from
python3 research/openrouter_ratios.py  # token ratios from the OpenRouter snapshot

Layout

• packages/engine: the model. Pure TypeScript, shared by the site and the CLI.
• apps/web: the static site (Vite + React): the day as one equation, the rack, and sliders.
• scripts: scenario tables and price snapshots.
• research: sources, raw data snapshots and the Python reference model.

License

• Code (packages/, apps/, scripts/, research/*.py): MIT, see LICENSE.
• Text, analysis and the curated assumptions (docs/, research/*.md, data/assumptions.json, the site's copy): CC BY 4.0. Reuse them with credit; the citation is in CITATION.cff and on the site.
• Third-party data in research/data/ and data/raw/ remains its owners' and is not covered by either license; see research/data/README.md.

Disclaimer

Personal side project, built from public data. It does not necessarily represent the views of my employer or of any company named here, and none of them sponsored it. The author may hold investments in companies mentioned here. Not financial, investment, tax, or purchasing advice. Estimates from public sources under stated assumptions; they can be wrong by large factors. No warranty.