Methodology

1. The Clock

The Clock shows a modeled estimate of installed global AI power capacity, in watts:

W(t) = base_gw × 10⁹ × g^((t − base_date) / year)

with base_gw = 30, base_date = 2025-12-31 and annual growth factor g = 2, all read from clock.json — the ticking rate is derived (dW/dt = W·ln g / year), never hardcoded.

Epoch AI chip-sales-based estimate with ~2.5x infrastructure multiplier; growth factor is an assumption pending first recalibration. Capacity, not usage. Note: Epoch's facility-level hub separately tracks 11.9 GW of IT power across 75 sites — a tracked subset, not the same quantity as this 30 GW estimate; see methodology.

Basis: Epoch AI's chip-sales-based power estimate (CC-BY): cumulative AI accelerator shipments converted to power draw, multiplied by ~2.5× for the surrounding infrastructure (cooling, networking, power delivery, non-accelerator hardware). This measures capacity, not usage — a data center's peak supportable draw, not what its chips consume at any given moment.

30 GW vs 11.9 GW — two different quantities

Epoch also runs a facility-level tracking hub reporting 11.9 GW of IT power across 75 tracked sites. That number is smaller for two reasons: it covers only the tracked subset of facilities (not all installed AI capacity), and it counts IT power (before the infrastructure multiplier). The Clock's 30 GW base is a top-down estimate of the whole; 11.9 GW is a bottom-up census of the identified part. They are not comparable and this site never mixes them. The Pipeline is our own bottom-up ledger and inherits the same caveat: its totals will always be smaller than the Clock.

Recalibration log (public)

The Clock is recalibrated monthly against new data; every change to the base or growth factor is logged in clock.json and listed on the changelog.

No recalibrations yet. The first monthly recalibration will be logged here and in clock.json.

2. The Pipeline

Stage definitions, in order of progression:

• Announced — a public statement of intent; no known binding commitment.
• Contracted — a signed, reported agreement (compute contract, lease, or JV) covering the capacity.
• Under construction — physical build-out underway.
• Energized — powered and operating. Only this column describes the present; only energized watts move the clock. everything else is a claim about the future.

Site de-duplication rule

Sites are the unit of capacity truth; deals link to sites. A site's GW counts once, no matter how many deals reference it, at the furthest-progressed stage attested by the site record or any linked deal. gw_by_stage holds exclusive tranches, never cumulative figures — when a fleet discloses "0.9 GW energized of 3.5 GW contracted", the contracted tranche is stored as 2.6 GW. A tranche recorded as null means "stage attested, magnitude undisclosed": it renders as "—", is excluded from sums (marked †), and is never silently estimated.

One deliberate exclusion: power-generation deals (e.g. the Microsoft × Constellation Three Mile Island PPA) appear in the Ledger's Offtake tab but contribute no Pipeline GW — generating capacity is not data-center capacity, and counting both would double-count the same watts.

3. The dollars taxonomy — why unlike units never sum

Every dollar figure carries a value_type:

• contract lifetime — total value over a multi-year compute contract.
• capex — construction capital expenditure.
• PPA — power purchase agreement value.
• ceiling (up to) — an "up to" figure; an option, not a commitment.
• equity — investment for ownership, not a purchase of compute.

These measure different things over different time horizons ("$300B over five years of compute" + "$27B of construction capex" + "up to $27B someday" is not a number). Aggregates on this site are computed within a single value_type only, and every aggregate states its unit and exclusions. There is deliberately no "total AI deal value" figure anywhere.

4. Build–Buy Index (beta)

score = 100 × (energized self-owned GW) / (energized GW consumed)

Computed quarterly per compute consumer (hyperscalers and labs), at build time, from Sites + Ledger rows only — never hand-entered. GW-only basis: dollars never enter. A consumer with zero attributable energized GW, or any energized site of undisclosed magnitude, renders "insufficient data" rather than a guess. Known limits: labs cluster near 0 or 100 on tiny denominators; capital-structure outliers (xAI's affiliate ownership, DeepSeek's parent-funded compute) blur the self-owned/bought boundary. Beta — do not cite without these caveats.

5. Update cadence

• Clock — recalibrated monthly (logged above).
• Ledger — rolling, as deals and builds are reported and verified.
• Build–Buy Index — quarterly.

All data lives as versioned JSON in the public repository; git history is the official snapshot record. Month-over-month diffs and the changelog derive from commits, not from a database.

6. Verification & provenance

Every displayed figure traces to a data-file row. Each number carries a confidence marker — measured (instrumented/officially filed), disclosed (stated by a party to the deal, incl. via press), estimated (third-party estimate) — and a verification state: rows without at least one source URL on file render with a hollow marker and "unverified" tooltip sitewide. Missing data renders "—", never 0.

7. License & corrections

This site's own data files are licensed CC-BY 4.0: reuse freely with attribution to computeclock.com. Upstream Epoch AI data is CC-BY and credited above.

Correction policy: errors are corrected in the data files via pull request, land as commits (therefore appear in the changelog permanently), and materially wrong figures are flagged in the commit message with correction:. To report an error, open an issue on the repository with a source URL.

8. Embedding

The Clock can be embedded from /embed/clock — a standalone, dependency-free page under 50 KB, cacheable at the edge. The iframe snippet is on the homepage. Every figure also ships as a PNG with the domain and data date burned in.