# AI Compute Tracker

> The price of compute, tracked. Independent tracker of AI compute economics: GPU rental prices, accelerator specs and cost-efficiency, token prices, and the datacenter buildout.

Conventions: rental rows retain their published billing model in USD per
GPU-hour. Default history comparisons use provider-list on-demand rates;
marketplace, capacity-block, serverless, and secondary rows stay separate.
FLOPS figures are dense, never vendor sparsity numbers; blended token
price = (3 × input + output) / 4. Forward curves are reconstructed from
Kalshi contract ladders — indicative market-implied levels, not
current spot or executable quotes. The first point is a future exact-time
weekly anchor when a usable ladder exists, otherwise a labeled front-month
monthly-average anchor; each point carries its bid/ask bracket and a
quality verdict, and the methodology key names the estimator. Every
feed payload and page carries its own freshness
stamp (the market snapshot and the core-data survey are stamped
separately).

License: The site’s own text and data compilations are licensed CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/); reuse them with the attribution “AI Compute Tracker (aicomputetracker.com), CC BY 4.0; includes data from Epoch AI (epoch.ai), CC BY 4.0”. Third-party source data — such as Kalshi market data and providers’ list prices — remains subject to its sources’ terms.

## Data feeds

- [Current GPU quotes (JSON)](https://aicomputetracker.com/quotes.json): every current quote with its billing model, observation time, evidence fields and source, plus each citable class aggregate's latest dated point from the price-history ledger
- [Current GPU quotes (CSV)](https://aicomputetracker.com/quotes.csv): the same quote rows for spreadsheets, each with its source and attribution
- [GPU price history (JSON)](https://aicomputetracker.com/price-history.json): observations and legacy estimates with billing, comparison, quality, and raw-evidence fields
- [GPU price history (CSV)](https://aicomputetracker.com/price-history.csv): the same long-form ledger for spreadsheets (UTF-8 BOM, formula-neutralized)
- [Raw-evidence references (CSV)](https://aicomputetracker.com/evidence/refs.csv): every raw-evidence reference behind the price-history ledger, with the payload digests it pins and the sources behind it; identifiers of payloads in a private archive, not links
- [Forward curves (JSON)](https://aicomputetracker.com/forward-curves.json): instrument-specific implied anchor plus monthly tenors, method identity, and per-point quality evidence
- [Forward curves (CSV)](https://aicomputetracker.com/curves.csv): the same snapshot for spreadsheets (UTF-8 BOM, formula-neutralized)
- [Marketplace ask tape (JSON, withdrawn)](https://aicomputetracker.com/marketplace-tape.json): withdrawn on 2026-10-20 with its only source: served empty, with the reason, until the next major contract version removes it
- [Reference prices for citing sites (JSON)](https://aicomputetracker.com/reference-prices.json): what each chip page and provider page headlines: on-demand range and cloud count, the reference price with its label, coverage, dates and citation sentence, the spec strip, and each provider's cheapest on-demand price per chip
- [Current token prices (JSON)](https://aicomputetracker.com/token-prices.json): LLM API list prices per million tokens — input, output, cache-hit and batch prices where recorded, long-context tiers, blended — with review dates and sources, plus the dated GPT-4-class price-floor history
- [Current token prices (CSV)](https://aicomputetracker.com/token-prices.csv): the same table for spreadsheets
- [Accelerator specs (JSON)](https://aicomputetracker.com/accelerators.json): dense tensor TFLOPS by precision, memory, bandwidth and power per chip, with est flags, review dates and vendor documents
- [Accelerator specs (CSV)](https://aicomputetracker.com/accelerators.csv): the same spec table for spreadsheets
- [Frontier buildout ledger (JSON)](https://aicomputetracker.com/buildout.json): provenanced frontier-cluster site ledger — every capacity figure carries dated public-source claims
- [Compute deals ledger (JSON)](https://aicomputetracker.com/deals.json): contracts for AI computing capacity since 2023 — cloud, chips, data-center leases, financing and power — each with the basis of its stated value and dated public sources
- [Buildout financials (JSON)](https://aicomputetracker.com/financials.json): big-tech capex by year with each company's basis notes, and NVIDIA data-center revenue by fiscal quarter
- [Market & quote events (RSS)](https://aicomputetracker.com/feed.xml): the current desk note, the newest market and quote events from the desk log (quote moves, venue changes, floor lows, anchor moves, market-structure milestones) and the newest corrections, each pointing to its full record; archive housekeeping stays on the desk log
- [Evidence digest manifest (JSON)](https://aicomputetracker.com/evidence/digests.json): every SHA-256 evidence digest the published rows cite, by UTC day, with its evidence reference, the series it backs and a per-day Merkle root — digests only, no payloads
- [Freshness status (JSON)](https://aicomputetracker.com/status.json): the three freshness clocks side by side (market snapshot, live quotes, core data), each with its own stamp and build-time age, plus every core dataset's review
- [OpenAPI description](https://aicomputetracker.com/openapi.json): OpenAPI 3.1 document of every open endpoint, generated from the versioned data contract
- [Full dataset (plain text)](https://aicomputetracker.com/llms-full.txt): every current number inlined as tables — quotes, forward board, token prices, specs, citable series, buildout — one section per freshness clock

## Pages

- [Home](https://aicomputetracker.com/): current GPU offers and entry points for API prices, hardware and research
- [Research](https://aicomputetracker.com/research/): market-implied forward curves, snapshot replay, H100 rental history, training compute and capex
- [GPU Rental Prices](https://aicomputetracker.com/gpu-rental-prices/): current published quotes with evidence-aware history and billing retained per row
- [Hardware](https://aicomputetracker.com/hardware/): accelerator specs and rental price per PFLOP-hour
- [AI Model API Prices](https://aicomputetracker.com/ai-model-api-prices/): blended token prices and the price-per-token collapse
- [Buildout](https://aicomputetracker.com/buildout/): datacenter capex, power, and cluster scale
- [Open data docs](https://aicomputetracker.com/data/): the endpoint contract: field schemas, freshness clocks, license, stability policy
- [Desk log](https://aicomputetracker.com/log/): dated event stream (quote moves, curve sessions, anchor moves, methodology and venue milestones, corrections) re-derived from the archives and the corrections log on every build
- [Evidence ledger](https://aicomputetracker.com/evidence/): the provenance chain end-to-end: snapshot IDs, checksums, verified-row counts, and pinned raw-evidence Git references
- [Verify a digest](https://aicomputetracker.com/evidence/digests/): how to recompute a rawEvidenceSha256 from a retained raw file, and a day's Merkle root from the manifest
- [Methodology](https://aicomputetracker.com/about/): sourcing rules, freshness stamps and review targets, estimator notes, independence
- [Corrections](https://aicomputetracker.com/corrections/): every material correction to the site's own figures and claims, as dated history: what it said, what it was corrected to, where, and the evidence
- [Terms of use](https://aicomputetracker.com/terms/): information, not advice; posted prices, not quotes; what the CC BY licence covers; third-party notices
- [Embeddable charts](https://aicomputetracker.com/data/embeds/): every iframe-ready single-chart permalink under /embed/ (in-frame freshness stamps, third-party source credits and attribution), each with a copy-paste snippet whose source line links back and states the data's licence
- [Badges](https://aicomputetracker.com/data/badges/): live SVG price badges for READMEs, with Markdown and HTML snippets that link each badge to the page it stamps

## Per-GPU pages

- [B300](https://aicomputetracker.com/gpu-rental-prices/b300/)
- [B200](https://aicomputetracker.com/gpu-rental-prices/b200/)
- [GB200 NVL72](https://aicomputetracker.com/gpu-rental-prices/gb200-nvl72/)
- [H200](https://aicomputetracker.com/gpu-rental-prices/h200/)
- [H100](https://aicomputetracker.com/gpu-rental-prices/h100/)
- [A100 80GB](https://aicomputetracker.com/gpu-rental-prices/a100-80gb/)
- [MI355X](https://aicomputetracker.com/gpu-rental-prices/mi355x/)
- [MI300X](https://aicomputetracker.com/gpu-rental-prices/mi300x/)
- [L40S](https://aicomputetracker.com/gpu-rental-prices/l40s/)
- [RTX 5090](https://aicomputetracker.com/gpu-rental-prices/rtx-5090/)
- [RTX 4090](https://aicomputetracker.com/gpu-rental-prices/rtx-4090/)
- [V100](https://aicomputetracker.com/gpu-rental-prices/v100/)
- [TPU v6e](https://aicomputetracker.com/gpu-rental-prices/tpu-v6e/)
- [TPU v5e](https://aicomputetracker.com/gpu-rental-prices/tpu-v5e/)
- [TPU v5p](https://aicomputetracker.com/gpu-rental-prices/tpu-v5p/)
- [Trainium2](https://aicomputetracker.com/gpu-rental-prices/trainium2/)

## Per-GPU event feeds

One RSS feed per chip that has market or quote events: the chip's newest 20 — its quote and anchor events, and the market-structure milestones whose source names it — with the site feed's item markup and guids.

- [B300 market & quote events (RSS)](https://aicomputetracker.com/feed/b300.xml)
- [B200 market & quote events (RSS)](https://aicomputetracker.com/feed/b200.xml)
- [H200 market & quote events (RSS)](https://aicomputetracker.com/feed/h200.xml)
- [H100 market & quote events (RSS)](https://aicomputetracker.com/feed/h100.xml)
- [A100 80GB market & quote events (RSS)](https://aicomputetracker.com/feed/a100-80gb.xml)
- [RTX 5090 market & quote events (RSS)](https://aicomputetracker.com/feed/rtx-5090.xml)
- [RTX 4090 market & quote events (RSS)](https://aicomputetracker.com/feed/rtx-4090.xml)