Aug Concentration: Cloud Top-3 Eases to 57% — Dual-Hub Still 17%; Pipeline Hits 915

Our 2026 concentration print answered the first distribution question: how thick is the tip across chip ownership, hyperscale capacity, regional power draw, and token throughput — Google 25% of H100e, Top-3 owners 54.8%, cloud Top-3 57%, US regions ~45% of AI DC capacity, and ByteDance leading tokens at 29.2%. The Q3 concentration lens then re-scored after Synergy’s hyperscale ledger: cloud Top-3 ticked to 58%, US pipeline seats printed 54.4%, and the token Top-1 flipped to Google at ~27%. The August location rankings update refreshed the geography tape again — pipeline 915, cloud Top-3 57%, dual-hub 17%, top-20 markets 60%. This Aug concentration lens asks the desk follow-up those vintages imply: how concentrated is the system at the top when you re-score Top-1 / Top-3 after the location rankings, watch the cloud tip ease, and see where new seats land?

The interactive dashboard above is built as that Aug concentration lens. Toggle Scoreboard, Sites & cloud, Ownership, and Tokens vs chips. The punchline is deliberately multi-sided. On chip ownership, the tip is still stuck: Google ~25%, Top-3 ~54.8%, Big-5 71.4% with within-Big-5 HHI ≈ 2,421. On hyperscale cloud capacity, AWS + Azure + Google Cloud print 57% — −1 pp versus the Q3 capacity tape. On markets, Northern Virginia + Greater Beijing alone hold 17% of hyperscale capacity, the top-20 markets clear 60%, and the United States occupies 15 of 20 largest-market seats. On pipeline, the Aug mix expands to 915 sites (+112 versus Q3’s 803 hyperscale ledger) while US pipeline share holds near 54%. Token Top-1 stays carried at Google ~27% — the Aug location print does not restate the brand cohort.

The Aug concentration scoreboard

Read the table as a family of market shares, not one slogan. Ownership concentration is thick at the hyperscaler tip and still leaves nearly three-tenths of world H100e outside the Big-5. Cloud concentration is a capacity story among operators who rent to frontier labs — and in Aug it softened a point even as the pipeline grew. Market concentration says where capacity already sits: two metros equal 17%, twenty markets equal 60%, forty markets equal 79%. Token concentration says who processes the prompts — still Google-led in the carried Q3 cohort, while ByteDance and OpenAI sit far above their chip stock.

Ownership tip: still sticky after the location print

Toggle Ownership. Epoch’s Chip Owners frame (Q4 2025 anchors, Aug explorer restatement) remains the carried print: Google near 25%, Microsoft 17.3%, Amazon 12.5%, Meta 11.3%, Oracle 5.3%. China as an aggregate owner is still near 5%. That is the same ladder the prior concentration companions and the Q3 lens published — because Epoch has not yet closed a Q1/Q2 2026 period delta that would force a restatement.

The concentration geometry therefore freezes: Top-1 = 25%, Top-3 = 54.8%, Big-5 HHI on the renormalized five-name perimeter ≈ 2,421 versus an equal-five benchmark of 2,000. The Lorenz panel still rises well above the equal-split line by rank three. What changed in the Aug location vintage is not the ownership tip — it is the geography of seats and the one-point cloud ease. Analysts who wait for the ownership tip to “break” before updating risk maps will miss the market-band and pipeline moves that already printed.

Ownership is still not usage. OpenAI, Anthropic, and other frontier labs rent most operational capacity. A lab can dominate the product narrative while owning almost none of the H100e stock. That is why the token scatter remains the corrective lens — and why Google’s dual lead (chips and tokens in the carried cohort) is a coincidence of vintages, not a law of nature.

Sites, markets, and clouds: the Aug tip that actually moved

Open Sites & cloud. Synergy’s August 19 location rankings — the same frame as the Aug theme update — put 915 pipeline sites on the board (+112 versus Q3’s 803 hyperscale-only ledger). Treat those as different universes when you need a clean delta; the concentration story is that US pipeline share held near 54% even as the absolute seat count grew. Ops seats remain more geographically plural (US ~42.6% of 1,360 live hyperscale sites) while new seats still concentrate into the US corridor — including inland markets that entered the top-20 (Indiana, Tennessee) as coastal and APAC names churned out.

On capacity among cloud operators, the Aug print puts AWS + Azure + Google at 57%, down 1 pp from the Q3 capacity tape of 58%. That is a small move with a large meaning: the tip did ease slightly as the pipeline expanded and non-Top-3 operators absorbed a point of share — but it did not dilute into a plural market. Market bands still put top-20 markets at 60%, Northern Virginia + Greater Beijing at 17%, markets 21–40 at 19%, and the long tail outside the top-40 at 21%. Seat composition is the US story: 15 US / 4 APAC / 1 Europe (Dublin) — non-US seats in the top-20 fell to 5 (from 6 YoY / 7 two years ago). Texas operational hyperscale capacity grew +71% YoY versus world +36%, so the inland corridor is not just adding seats — it is adding faster capacity.

The Gartner electricity slice is the watts twin of the site ledger. World DC power prints near 565 TWh in 2026; the US alone is 204 TWh (36%), with roughly 68 TWh dedicated AI inside the US (~1/3 of US DC power, ~12% of world DC TWh). Regional AI DC capacity by power draw — the synthesis carried from theme research — still puts the United States near 45%, China near 18%, and Europe near 14% (Top-3 regions ≈ 77%). Site counts, capacity shares, and electricity TWh are three meters on the same geography — they will not match digit-for-digit, and they should not be averaged.

Tokens: carried Q3 cohort — mismatch geometry intact

Toggle Tokens vs chips. The Aug location rankings do not restate brand token shares, so this lens carries the Q3 cohort: Google near 27%, ByteDance near 21%, OpenAI near 15%, token Top-3 about 63%. That keeps the prior June → Q3 flip visible (ByteDance 29.2% → Google 27%) without inventing an Aug token discovery.

That does not mean ownership and usage finally agree. Google is the rare name that sits near the diagonal of the scatter: ~27% tokens and ~25% ownership. ByteDance still prints ~21% tokens against ~1.5% ownership. OpenAI prints ~15% tokens against ~0% disclosed chip stock. Microsoft, Amazon, and Meta remain ownership-heavy versus first-party tokens — Azure, Bedrock, and Llama routing dilute credit. The concentration lesson for Aug is therefore two-sided: the identity of token Top-1 is carried, but the mismatch geometry between rented usage and owned chips did not disappear just because metros re-ranked.

What the Q3 → Aug deltas actually say

The ownership tip is the stable risk factor. The cloud tip is the creeping one — and in Aug it crept down a point, not up. The market bands are the placement story: dual-hub and top-20 remain thick even as inland US seats proliferate. The pipeline count is the expansion story — more seats, similar US share. The token tip is the volatile one, but Aug does not restate it. Desks that collapse these into a single “AI is concentrated” slogan will mis-rank which perimeter actually moved this vintage.

Caveats and reading rules

1. Carried ownership is not a no-news print. It means Epoch has not closed a superseding period. Treat Google 25% / Big-5 71.4% as the last disclosed ladder, not as an Aug discovery.
2. Hyperscale-only ≠ location mix. Synergy’s Q3 803 pipeline seats are not the Aug 915 large-DC count. Do not subtract them into a fake “+112 hyperscale” claim without noting the scope change.
3. Token cohorts are editorial and carried. Vendor keynotes, China NDA stats, OpenRouter samples, and revenue proxies do not form a single audited market. Aug does not refresh them.
4. Dual-hub 17% is capacity, not seats. Northern Virginia + Greater Beijing as 17% of hyperscale capacity is not the same meter as US 15 of 20 market seats or US ~54% of pipeline sites.
5. H100-equivalent ≠ identical utility. Peak 8-bit FLOP/s conversions hide memory, networking, and software differences across TPU, Blackwell, Trainium, and Ascend.
6. Cross-perimeter averages are meaningless. Do not average owner Top-3, cloud Top-3, top-20 market share, and token Top-1 into one “concentration index.”

What would change the story

A finalized Epoch Q1/Q2 2026 Chip Owners period that moves Google off 25% or lifts China ownership well above 5%; a Synergy capacity print that pushes Top-3 cloud below 55% as neoclouds and Meta/Oracle absorb share; a location ranking where dual-hub falls under 14% or non-US top-20 seats rebound above 7; a pipeline mix where US seats fall under 45% as Middle East and APAC campuses clear; or a token series that restores China-origin brands above 50% without a matching ownership move. None of those supersede the Aug location bands and cloud 57% print in the sources we carry here.

Related reading: August location rankings update, Q3 concentration lens, 2026 concentration print, Q3 site-count update, and theme research baseline.