Token Forbes

Token Forbes is a Vite + React app that ranks people and organizations by inferred AI token consumption.

Production:

• https://token-forbes.vercel.app

Background

In the AI era, tokens and compute have replaced traditional physical resources as the absolute measure of influence. Rankings now track cumulative burn since January 2025 and the average monthly burn rate behind it.

What It Does

• Shows an AI token-consumption leaderboard for individuals and enterprises
• Uses cumulative totals since January 2025 plus average monthly burn
• Supports proof-of-compute submissions and GitHub repo intake
• Builds a dynamic GitHub global ranking from:
  • GitHub Search API
  • recent contributor activity
  • benchmark repos such as microsoft/vscode, vercel/next.js, langchain-ai/langchain, and huggingface/transformers
• Builds a modeled global enterprise ranking using the same proxy methodology plus public-code calibration
• Stores leaderboard submissions in SQLite by default, with Vercel Blob as a fallback
• Uses OpenRouter for chat and data-engine AI features
• Supports Google and GitHub sign-in via Firebase Auth

Stack

• Frontend: React 19, Vite, Tailwind CSS, Motion
• APIs: Vercel serverless functions
• Persistence: SQLite by default, Vercel Blob fallback, browser localStorage fallback on the client
• Auth: Firebase Authentication
• AI: OpenRouter, default model stepfun/step-3.5-flash:free

Local Development

Prerequisites:

• Node.js 22.5+
• npm

Install and run:

npm install
npm run dev

Default dev URL:

http://localhost:3000

Environment Variables

Create local env values as needed. For production, configure them in Vercel.

Required for AI

OPENROUTER_API_KEY=...
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
OPENROUTER_MODEL=stepfun/step-3.5-flash:free

Recommended for GitHub Ranking

GITHUB_TOKEN=...

Without GITHUB_TOKEN, the global GitHub ranking route works in a degraded mode and can hit GitHub rate limits sooner.

Default / Optional Persistent Storage

SQLite is the default server storage path.

TOKEN_FORBES_STORAGE_PROVIDER=sqlite
SQLITE_DB_PATH=.data/token-forbes.sqlite

Optional Blob Fallback

BLOB_READ_WRITE_TOKEN=...

If TOKEN_FORBES_STORAGE_PROVIDER is omitted, the server tries SQLite first and falls back to Blob when SQLite is unavailable.

Optional Firebase Override

The app currently supports overriding the built-in Firebase project through Vite env vars:

VITE_FIREBASE_API_KEY=...
VITE_FIREBASE_AUTH_DOMAIN=token-forbes.firebaseapp.com
VITE_FIREBASE_PROJECT_ID=token-forbes
VITE_FIREBASE_STORAGE_BUCKET=...
VITE_FIREBASE_MESSAGING_SENDER_ID=...
VITE_FIREBASE_APP_ID=...

If all six values are present, the app uses that Firebase project instead of the fallback one baked into the repo.

Firebase Notes

For Google/GitHub sign-in to work, the active Firebase project must:

• enable Google Sign-In
• enable GitHub Sign-In
• include these domains in Authentication -> Settings -> Authorized domains
  • token-forbes.vercel.app
  • token-forbes-vercel.vercel.app

Vercel Deployment

This repo is configured for Vercel.

Deploy:

vercel --prod

Important deployment details:

• token-forbes-vercel.vercel.app is redirected to https://token-forbes.vercel.app/#rankings
• AI requests go through server-side api/ai, so the OpenRouter key stays on the server
• GitHub ranking APIs run on the server and use GITHUB_TOKEN when available
• On Vercel, SQLite defaults to /tmp/token-forbes.sqlite, which is cheap and functional for low traffic but not durable across cold starts; use Blob or a persistent volume if you need durable production storage

Main API Routes

• /api/leaderboard
• /api/proof
• /api/github-ranking
• /api/github-global-ranking
• /api/ai

Scripts

npm run dev
npm run build
npm run preview
npm run lint

Recent Product Changes

• Switched AI features from client-side Gemini to server-side OpenRouter
• Added GitHub login alongside Google login
• Changed leaderboard semantics to:
  • Total Tokens: cumulative total since January 2025
  • Avg Monthly Burn: average monthly token consumption since January 2025
• Added dynamic global GitHub ranking with up to 500 contributors
• Added a dedicated global enterprise token-consumption ranking derived from project proxy methodology
• Replaced Firestore-era persistence with SQLite-first storage and Blob fallback

TODO: Data Accuracy & Sustainability

Current implementation reality: the production leaderboard still blends demo seed entities, public GitHub proxy estimates, self-reported proofs, and a manually seeded enterprise model. To make Token Forbes objectively credible and sustainably evolvable, prioritize the following:

• Remove demo/mock entities from production rankings, or gate them behind an explicit demo mode so synthetic seed data never competes with live entities.
• Build a per-entity provenance ledger that stores source URLs, collection timestamps, raw snapshots, parser version, prompt/model version, and methodology version for every ranked estimate.
• Expand GitHub ingestion beyond the current stars > 1000 and recent-push sampling window so rankings are less biased toward a small set of famous public repos.
• Add identity resolution across GitHub logins, organizations, proofs, and future partner feeds so rankings stop merging or splitting people based only on display names and repo owners.
• Add stronger GitHub quality controls for bots, mirrors, vendored code, mass-formatting commits, monorepos, and smaller but high-intensity AI repositories that the current heuristic filters miss.
• Calibrate calculateAITokens() against real-world ground truth such as voluntary billing samples, provider receipts, IDE telemetry, or partner data instead of relying only on hardcoded multipliers, floors, and caps.
• Make confidence intervals evidence-based, using source coverage, freshness, and cross-source agreement, instead of mostly static hand-tuned percentages.
• Separate self-reported Proof of Compute submissions into verified and unverified tiers, and require verification artifacts or connectors before they can influence the main public leaderboard.
• Replace manually curated enterprise seeds, signal scores, and scale multipliers with an evidence-backed feature store built from official disclosures, cloud partnerships, infra announcements, product usage, and calibrated public-code signals.
• Keep direct disclosure, proxy inference, model estimation, and API-partner data in separate pipelines with explicit trust weights instead of using source tags as labels only.
• Persist citations and extracted numeric evidence for Data Engine evaluations so AI-generated enterprise estimates are reviewable, reproducible, and diffable over time.
• Add private-repo and private-usage opt-in connectors so Token Forbes can measure serious AI builders whose compute spend is invisible in public open-source activity.
• Introduce scheduled snapshotting, regression tests, anomaly detection, and manual review queues so methodology changes do not silently rewrite the leaderboard.
• Publish methodology changelogs and historical snapshots so users can understand why a rank changed and compare estimates across versions.

Repository

• Remote: https://github.com/wangxumarshall/Token-Forbes.git