Tokenomics

GPU inference profitability plots by Experiential Labs.

Open the dashboard

Benchmark credit

Throughput and cache-hit data come from SemiAnalysis AgentX and InferenceX. Benchmark credit belongs to SemiAnalysis and its contributors.

This is an independent visualization, not an official SemiAnalysis product. We add GPU rental, pricing and utilization assumptions to model profitability.

Model

The plots show monthly revenue minus GPU rent across utilization and blended token prices. Choose a model, GPU and serving setup; drag the $/GPU-hour slider to update the chart. The x-axis also shows total tokens/month, including cached input.

The dashboard includes 8 models, 10 GPU types and 73 serving curves; GLM-5.2 is excluded. Each curve uses its highest measured throughput meeting 50 output tokens/sec/user (p90); curves with no qualifying point are visibly marked. This is not a complete historical archive. Larger multi-node runs retain their actual GPU count. Smaller serving units are replicated to eight GPUs and labeled as replication assumptions. GLM-5.3 has its own measured results.

B300 defaults to $5.50/GPU-hour. Other defaults are provider list-price references checked October 3, 2026, not negotiated quotes or guaranteed availability. Vera Rubin has only a labeled SemiAnalysis rental estimate. All rates are editable and retained per GPU during the session. Sources and terms: lib/rental-rates.mjs.

TPU Ironwood and Trillium are listed as unavailable in the benchmark dropdown: Google publishes rental prices, but this AgentX snapshot has no TPU measurements. There are no invented TPU throughput curves.

Benchmark selection: lib/benchmarks.mjs. Raw public data: lib/agentx-snapshot.json. Refresh with node scripts/refresh-agentx.mjs, then run tests and review any new hardware/rates. Results assume 730 hours/month, before other operating costs; measured capacity is not a demand forecast. Source links, benchmark dates and price dates are in the dashboard’s Assumptions & sources.

Explorer: every catalog model on every GPU

/explorer puts every priced model from the public catalogs on a set of GPUs and draws:

• break-even utilization for every model, one line per GPU;
• one model's monthly profit against GPU price and against utilization, one line per GPU;
• a sortable table of profit and break-even utilization per model and GPU.

Catalogs (pick any in catalog.sources):

Source	API	Prices
experiential	https://api.experientiallabs.ai/api/models	Platform-funded lanes, in serving order (lane: "lead" is what Experiential charges)
huggingface	https://router.huggingface.co/v1/models	Hugging Face Inference Providers, open weights only
openrouter	https://openrouter.ai/api/v1/models	OpenRouter list price; open weights when it links a Hub repo

Throughput, cache hit rate and GPU prices are generic placeholders, not measurements. Everything lives in one JSON scenario, private/default-scenario.json:

• workload: input and output tokens/sec per GPU at throughput multiplier 1, plus cache hit rate.
• gpus: id, name, $/GPU-hour, and a throughput multiplier. Patch one with a map: {"gpus":{"h100":{"hourlyRate":1.5}}}; null removes a GPU.
• fleet: GPU count, hours per month, paid utilization, other monthly cost, fee rate.
• models: per-model overrides keyed by source:slug (or bare slug): prices, throughput, GPU count, and per-GPU measured numbers under gpus. See scenarios/measured-b300.json.
• catalog: sources, price lane (lead, cheapest, max), include/exclude slug patterns (* wildcard), openWeightsOnly.
• sweep: GPU price and utilization ranges for the curves.

Every input is a partial scenario deep-merged over the defaults, so the same JSON drives all three entry points:

# HTTP API (JSON; add ?format=csv for a spreadsheet, ?curves=1 for the curve data)
curl -s -X POST https://tokenomics-one-delta.vercel.app/api/report \
  -H 'content-type: application/json' -d '{"gpus":{"h100":{"hourlyRate":1.5}},"fleet":{"utilization":0.3}}'
curl -s https://tokenomics-one-delta.vercel.app/api/scenario   # defaults
curl -s 'https://tokenomics-one-delta.vercel.app/api/catalog?sources=experiential,huggingface'

# CLI (files merge left to right)
node report.mjs scenarios/measured-b300.json scenarios/cheap-hopper-busy.json --csv > report.csv

The explorer's JSON editor takes the same patch, and Copy link / Copy API call reproduce a view exactly.

Open in the benchmark heatmap (under the selected model in the explorer) opens the original dashboard on that catalog model and GPU with the current scenario: /?catalog=<source:slug>&gpu=<id>#s=<scenario>. The heatmap is unchanged; the model is one extra picker entry served by GET /api/heatmap-model?id=<source:slug>&gpu=<id>&scenario=<json>, labelled as generic throughput unless the scenario pins measured numbers for that GPU.

Run locally

npm ci
npm test
npm start

Open http://127.0.0.1:4319. No credentials required.