Where Is The Compute?

Tools for people who follow the watts

Everything on this site is free to read. These tools are for analysts, journalists, investors and builders who need to model, compare, present and track the numbers. Every tool uses the same sourced data and keeps every range.

Plain terms: after checkout you get unlisted links (not DRM; please don’t share them). Supporter is $9/month; cancel any time. No accounts, no trackers. The underlying Epoch AI data is CC BY 4.0: you pay for curation, tools and updates.

Compute Brief · $29 Speculation

Scenario modeler

Set your own chip deliveries, power caps, utilization and retirement; get your own 2026–2028 compute path with a low–high band, which constraint binds (chips or watts), and JSON/CSV/SVG/PNG export. Labeled speculation on screen and in every export.

Get the brief — $29

Example scenario: OpenAI with starter assumptions (+1.0M H100e/yr, no power cap)2025: 1.74M (1.25M–2.19M); 2026: 2.74M (1.95M–3.48M); 2027: 3.74M (2.65M–4.79M); 2028: 4.74M (3.35M–6.08M)02M4M6M8M20252026202720281.74M2.74M3.74M4.74MSPECULATION — YOUR ASSUMPTIONS, NOT A FORECAST BY US OR ANY COMPANYExample only · starter values are arbitrary · 2025 = Epoch median + 90% CI

Compute Brief · $29

Lab-vs-lab comparator

Pick two to four labs or parent companies; ranges animate side by side, a plain-language verdict says whether 90% intervals overlap, and the view exports to SVG/PNG/CSV/JSON with the citation baked in.

Get the brief — $29

Example comparison: OpenAI vs Google DeepMindOpenAI 1.743M (1.251M–2.185M); Google DeepMind 1.583M (1.008M–2.549M)0600K1.2M1.8M2.4M3MH100e used, end-2025 · whiskers = 90% CI · verdict: ranges overlapOpenAI1.74M (1.25M–2.19M)Google DeepMind1.58M (1.01M–2.55M)

Compute Brief · $29

Chart & embed kit

Every chart as a standalone SVG and a 2× PNG with its citation baked into the image, plus copy-paste iframe and <wic-chart> web-component snippets.

Get the brief — $29

Compute actually used by each lab (export with citation)Who owns the world’s AI compute (export with citation)Follow the watts: capacity over time (export with citation)Cluster scale: one square = 10,000 GPUs (export with citation)

Compute Brief · $29

Slide-ready briefing deck

A 16:9 HTML deck (arrow-key navigation) plus a printable PDF, generated from the current data at build time: the answer, each view, ownership, clusters, gigawatts, caveats and sources.

Get the brief — $29

  1. The answer
  2. Compute used
  3. Compute owned
  4. Who owns the world’s compute
  5. Cluster scale
  6. Gigawatts over time
  7. Announced vs installed
  8. Caveats
  9. Discrepancy log
  10. Sources

Supporter · $9/mo

Data feed + 'what moved and why'

Versioned JSON and CSV snapshots, a JSON Feed and RSS changelog of figure changes, each with a 'what moved and why' note tied to its source.

Subscribe — $9/mo

{
  "id": "2026-10-20-D4",
  "title": "OpenAI future capacity: what moved and why",
  "content_text": "Was: Roughly 30 GW by 2030 through Stargate. Why: Shown as '~30 GW total commitments (all partners)'; '2030' flagged as unverified timing."
}

Supporter · $9/mo

Provenance trail

Per-figure history: the value in Iggy's source doc, earlier source values (e.g. Epoch Q4 2025 vs preliminary Q1 2026), the current value, and the source, date and check date for each revision.

Subscribe — $9/mo

Show data table
Example provenance: OpenAI compute used
StageValueDateSource
Iggy’s source doc~1.74M (1.25M–2.19M)2026-10source doc
Verified1,743,000 (1,251,000–2,185,000) H100e2025-12-31Epoch AI

Briefing call · $299

Custom scenario run

On the $299 call we run the scenario modeler on your assumptions together and you keep the JSON/CSV export plus follow-up notes with sources.

Buy a call — $299

Bring a question; leave with an exported scenario and sourced notes.

Free, for everyone

Close-up of copper cold plates and braided coolant hoses on an accelerator board
Illustration (AI-generated), not a real facility.