How every number on the board is made

Rules first: nothing appears without a named public source and a date. Moving figures are arithmetic on sourced numbers, never telemetry: flows tick at their filed pace; a count ticks only as a clearly labeled pace measured between two dated observations and run forward; shares never tick. Anything we could not source got cut, and this page says what that was. Each metric also has its own page with the full table, sources and downloads — linked throughout. Board last re-based October 20, 2026.

The hero figure

The headline counter is big-4 capex only: $513.8B (Microsoft, Alphabet, Amazon, Meta, four quarters to Jun 30, 2026, as filed) ÷ 31,556,952 seconds = $16,281.93/second, about $1.41B/day. It deliberately excludes NVIDIA's data-center revenue to avoid double-counting; NVIDIA appears separately. The combined, overlapping flow ($25,082/s) remains the Buildout Index's basis and is documented below. Changed Oct 20, 2026 after panel review; logged in the changelog.

Plain-language conversions — the arithmetic

• PJM serves "more than 67 million people across 13 states and DC" — PJM's own pages and releases (2026).
• Capacity price ≈ 1.1¢/kWh: $269.92/MW-day ÷ 24h ÷ 1,000 kW = 1.12¢ per kWh if usage were flat around the clock ($28.92 → 0.12¢). A simplification: real bill impact depends on region, peak usage and state pass-through.
• Bills rose an estimated 1.5–5% in the first repriced PJM year — PJM figures as reported by WHYY, Jul 2025; utilities pass wholesale costs to customers.
• A US home uses ~30 kWh/day: EIA FAQ — 10,791 kWh per residential customer in 2022 ÷ 365.
• 47 → 108 GW: 415 and 945 TWh/yr ÷ 8,760 hours. 129 reactors: 107.9 GW ÷ 0.835 GW (Three Mile Island Unit 1's announced capacity).
• ND flaring ≈ heat for ~1 million homes: 173.3 MMcf/d ÷ ~163 cf/day per gas home (derived from EIA 2024: 4,389,004 MMcf across 73.7M residential customers).
• 31x: 233 GW of ERCOT large-load requests ÷ 7.48 GW under construction (CBRE) — requests vs real construction.
• ~4M gal/day: ICPRB's 4 MGD average-day consumptive figure for Potomac-basin data centers (~9% of the area's consumptive use).

How opinions are marked

Narrative framings on the board (chapter sentences, the Bakken juxtaposition) are ours and are written to be checkable against the data shown; the one standing editorial product — the Reckoning — is explicitly tagged EDITORIAL, with its conditions and thresholds fixed on Oct 20, 2026, before any pending outcome. The Reckoning carries no Dataset markup and is excluded from the citable datasets.

The wall — every live counter's formula

• Hero (big-4): see above — $16,281.93/s. The combined overlapping flow ($25,081.89/s trailing-year; $32,435/s at the latest-quarter pace) is retained here and on the index page.
• VC into AI: $102B (Q3 2026, Crunchbase) ÷ 92 days = $12,832/s at the Q3 pace.
• New ChatGPT users: (1.2B − 900M) ÷ 214 days between OpenAI's Feb 27 and Sep 29, 2026 statements = 16.2/s — the measured pace, run forward (an extrapolation).
• Data-center electricity: IEA's 415 TWh (2024, all data centers) ÷ seconds in a year = 13,151 kWh/s.
• Stargate: shown as a static $500B commitment. The former per-second counter ($3,961/s if spent evenly) was removed Oct 20 as a hypothetical, not a measured pace.
• Models published: (3,157,435 − 2,678,385) ÷ 228 days between archived Hugging Face counts = one repository every ~41 s.
• Tape: one arXiv cs.AI listing every ~8.2 min (51,361 in 292 days of 2026); one announced AI-cited US job cut every ~3.3 min (120,136 in 273 days); 46.3 gal/s of Potomac-basin data-center consumptive water (4 MGD average, ICPRB); 1,534 days to Jan 1, 2031 as of this re-base.

The Buildout Index

As of Q2 2026 the BRAIAIN Buildout Index stands at 647.7 (base: Q1 2022 = 100) — a 6.5x rise in 18 quarters, 55.2% compounded annually, including a dip to 97.7 in Q1 2023.

Definitions as filed: Microsoft and Alphabet report cash purchases of property and equipment; Amazon reports gross purchases (not netted for proceeds); Meta includes finance-lease principal, matching its own release figures. NVIDIA is data-center segment revenue; its fiscal quarters end ~1 month after calendar quarters and are mapped to the nearest one. Quarters not reported directly are derived by subtracting YTD figures within or across the same company's filings — each derivation is recorded in the data file. The components overlap (chips the big 4 buy from NVIDIA appear in both their capex and NVIDIA's revenue), so the sum is a gross, overlapping flow — a measure of buildout intensity, not a de-duplicated market total. Quarterly table figures are rounded; totals and index values are computed from unrounded filings. Full table and FAQ: the index page; component sources: the capex page.

Capital

Power

Compute & research

Infrastructure & the map

Flared gas (the waste line)

Society

The Reckoning model, v1.0

The board's one editorial layer: our estimate of when AI becomes the default condition of the economy. Not a measurement, a market, a poll, or anyone's forecast but ours — it exists to be argued with, which is why every input is here. It is kept visually and structurally separate from the sourced data, and it carries no Dataset markup.

The ten conditions

1. Machines outnumber humans on the web — met 2024 (51% automated traffic; Imperva-network scope).
2. A grid capacity auction clears short of its requirement — met Dec 2025 (PJM 2027/28, 6,623 MW short; repeated Jul 2026 for 2028/29 at 6,831 MW).
3. One AI product passes a billion weekly users — met ~May 2026 (interpolated between OpenAI's Feb and Sep 2026 statements).
4. Big-4 infrastructure spend passes $500B a year — met with the four quarters to Jun 2026 ($513.8B, filings).
5. AI takes the majority of venture capital — met Q3 2026 (64%, Crunchbase).
6. 100,000+ job cuts in a year citing AI — met 2026 (120,136 by September, Challenger).

Pending:

7. First reactor restarted for a tech buyer — Crane targeted 2027 (Constellation; sourced projection).
8. Data centers take a tenth of US electricity — 4.4% in 2023; LBNL projects 6.7–12% by 2028. High case crosses 10% by 2028; extrapolating the low case's implied growth (8.77%/yr relative) puts 10% near 2033. This condition sets the range.
9. 10 GW of hyperscaler nuclear announced — 7,754 MW to date; at the ~3.6 GW/yr pace measured Sep 2024–Oct 2026, crossing ~mid-2027 (our extrapolation; no new deal since Jan 2026).
10. A frontier lab reports crossing its autonomy threshold, as defined in published scaling policies (e.g. autonomous replication in Anthropic's RSP — v1.0 cited; the policy has later versions) — no sourced projection exists; we extrapolate the overall crossing pace (6 conditions in 2.25 years ≈ 2.7/yr) to ~2028. The model's weakest input, and we say so.

The estimate

What moves it: a pending condition crossing, a met condition being revised away, LBNL/IEA updating projections, or a versioned model revision logged below. The conditions, thresholds and midpoint rule are editorial choices; different conditions give a different year — that is why the machinery is published.

The ChatGPT-era strip

• $1.09T: Microsoft, Alphabet, Amazon and Meta capital spending summed over the filed quarters Q4 2022 – Q2 2026 (big-4 only, no NVIDIA — no overlap). The earlier combined figure ($1.62T incl. NVIDIA revenue, overlapping) was retired from the board on Oct 20 and lives only here for the record. ChatGPT launched Nov 30, 2022 (OpenAI blog), one month into Q4.
• ×31 models: 101,574 archived Hugging Face repositories on Dec 15, 2022 → 3,157,435 on Oct 16, 2026.
• ~$64 per person: $513.8B big-4 trailing-year capex ÷ the 8 billion people the UN counted on Nov 15, 2022. Population has grown since, so today's per-head figure is slightly smaller. (The earlier ~$99 figure used the overlapping combined flow and was retired.)

The scorecard — we grade our own extrapolations

Every forward-looking figure this site publishes is logged here with its date, and graded publicly at later re-bases: right, wrong, or revised, with the arithmetic. Sourced third-party projections (IEA, LBNL) are theirs to grade; these are ours.

The archive — the board remembers itself

Every re-base is frozen at a permanent URL under /archive/ — the board and its data files exactly as published that day. Old citations stay checkable forever, and the record of what this site said, and when, is itself public. Snapshots are never edited after the fact; corrections happen on the live board and in the changelog.

What we cut, and why

• Live stock prices / market caps — audited Stooq, Alpha Vantage, Finnhub, Twelve Data, Cboe, Nasdaq/NYSE and IEX: no free route permits public ad-supported display without an agreement (IEX's delayed feed permits redistribution but has no web API; Wikidata's CC0 market caps run months stale). The watchlist shows sourced events only until a licensed route exists.
• Global AI query energy — Google publishes 0.24 Wh per median Gemini text prompt (Aug 2025), but global query volume is unpublished; any global figure would be invented.
• Global data-center water — only incomparable per-company reports exist; we show the Potomac-basin study instead.
• Summed tokens/day across providers — one-off disclosures with incompatible definitions.
• % of new content that is AI-generated — rests on AI-detector studies with known false-positive problems.
• Crowd forecasts of AGI — Metaculus's terms require written permission for commercial reuse; nothing of theirs appears here.
• USPTO AI patent counts and proposed-data-center map pins — no current official count; the best proposed-sites dataset is non-commercial-only.

Embed a counter — free

Any site may embed BRAIAIN counters at no cost. Copy one line, swap the c= value:

Counters: spend (building AI, $/s) · vc (VC into AI) · users (ChatGPT weekly users pace) · kwh (data-center electricity) · index (the Buildout Index level) · flared (North Dakota flaring). Each card links back here and carries its source and date; values are baked at each re-base, and moving ones run the measured paces described above. We may block embeds on sites that misrepresent the figures.

About & corrections

BRAIAIN is researched, written and maintained as a solo project (a dnilgis project). Corrections are welcome and logged: open an issue on the project's GitHub repository (dnilgis/braiain) with the figure, the problem, and a source. Cite any figure as "via braiain.com" with its original source named — the original source is always the authority. Compiled datasets on this site are CC BY 4.0; underlying sources keep their own terms.

Changelog

© 2026 BRAIAIN, a dnilgis project. Not affiliated with any company named. Map data © OpenStreetMap contributors (ODbL).