🧠 GPU & compute — is the body of AI still available, and at what price?

Latest observations

"Not yet measured" means missing, not zero. Machine-readable: /us/api/gpu_us.json

Observation tables

Question this table answers: what are the values and dates in each row of "Capital expenditure reported in SEC filings by 4 major cloud operators, sum per calendar…"? (source or retrieval date on each row)

Question this table answers: what are the values and dates in each row of "Federal Register documents from the Bureau of Industry and Security matching "advanced c…"? (source or retrieval date on each row)

Question this table answers: what are the values and dates in each row of "Public on-demand price of one NVIDIA H100 SXM, per GPU-hour, USD (Lambda public pricing…"? (source or retrieval date on each row)

Question this table answers: what are the values and dates in each row of "Published inference token prices, tracked models (USD per 1M tokens, list price; weekly…"? (source or retrieval date on each row)

Question this table answers: what are the values and dates in each row of "Most recent BIS documents matching "advanced computing" (ledger of the last 15)"? (source or retrieval date on each row)

Question this table answers: what are the values and dates in each row of "DOE / LBNL 2024 United States Data Center Energy Usage Report - figures as quoted by DOE…"? (source or retrieval date on each row)

Candidates (not yet observed) — added to this chapter once verification completes

• Cloud GPU list prices from additional U.S. providers (CoreWeave and others) — Per-provider terms of use must be checked before any snapshot (CoreWeave AUP check pending per docs/規約確認_Phase2_20260829.md).
• Hyperscaler pricing APIs (AWS Price List, Azure Retail Prices) — AWS bulk price files are hundreds of MB per region; the Azure Retail Prices API is documented but the site terms of use for reproduction were ruled out for the token-price series on 2026-08-26 - not adopted without a fresh ruling.
• Semiconductor supply feature (Korea HS-level exports, Taiwan monthly totals) — Prepared as a separate feature page (docs/準備メモ_半導体サプライチェーン特集_20260903.md); Korea API key pending.

Method, scope, and why this chapter is not judged

• Observing, not judging: no frozen criterion exists for the U.S. compute layer. Numbers, dates and sources only; do not infer thresholds. Nothing here ranks or evaluates a company.
• Export controls (main series): Federal Register documents issued by the Bureau of Industry and Security (BIS) whose full text matches the exact phrase "advanced computing", counted by month of publication; the most recent rules are listed. The count is defined entirely by the query (documents mentioning the phrase incidentally are included). Public domain; no key.
• H100 cloud price (weekly snapshot): the public on-demand hourly price of one NVIDIA H100 (per GPU, USD) as displayed on one U.S. provider's public pricing page, fetched at most once a week with an identifying user agent (terms checked 2026-08-29: low-frequency snapshots permitted). Only the figure is extracted; no page text is copied. Three states: ok / pattern mismatch (page changed — needs review) / fetch error. One provider is a sample, not a market.
• Token prices: published list prices per million tokens for tracked models and the cheapest observed, same definition v1.0 as the Japan edition (global USD prices, shown here as world background).
• Capex ledger (quarterly): "Payments to acquire property, plant and equipment" (or "productive assets", as each filer tags it) reported by Amazon, Microsoft, Alphabet and Meta in SEC filings, taken from the SEC XBRL frames API for each calendar quarter. Reproduced as filed; the sum is arithmetic on public filings. Filers whose fiscal quarters do not align with calendar quarters are not included.
• DOE/LBNL ledger: the figures the U.S. Department of Energy quotes from the 2024 LBNL United States Data Center Energy Usage Report (verified on energy.gov 2026-09-04). Public domain; irregular.
• Boundaries: grid load belongs to the electricity chapter; rulemaking on AI generally to the regulation chapter (this chapter counts only BIS advanced-computing documents); construction and cooling to the data-center chapter.
• What this chapter can say: how many Federal Register documents from the Bureau of Industry and Security matched "advanced computing" each month, and which rules were the most recent; the public on-demand price of one H100 GPU-hour at one U.S. provider, week by week, and the published list prices per million tokens for the tracked models; how much capital expenditure Amazon, Microsoft, Alphabet and Meta reported to the SEC per calendar quarter (summed), and which figures DOE quotes from the 2024 LBNL data-center energy report.
• What this chapter cannot say: whether GPUs can actually be obtained (no inventory, lead-time or waitlist data is published or observed); prices actually paid under negotiated or long-term contracts (public list prices only), or the prices of providers outside the panel; whether any level is high or low, or which company is doing better (nothing here ranks or evaluates a company).
• Sources: Federal Register API, SEC EDGAR XBRL frames API, Lambda public pricing page, U.S. DOE (LBNL report announcement) — compiled by AI Lifeline.