Update: US Takes 437 of 803 Hyperscale Pipeline Sites — Top-3 Cloud = 58%, Ownership Flat

What changed since the August update

In mid-August we published a vintage delta on Epoch’s Chip Owners Explorer: Microsoft restated to ~17.3% of world AI compute (+2.3 pp versus the July residual), Amazon and Oracle revised down, and Synergy’s July capacity note put ~45 GW of US IT load in the large-data-center pipeline with Texas + Midwest jumping from 33% of operational US hyperscale capacity to 53% of the pipeline. That post answered what moved in the ownership restatement and the GW pipeline. This Q3 refresh answers a sharper site question: once Synergy’s hyperscale site-count ledger and Top-3 concentration print are promoted, and Gartner’s US electricity composition is layered on, does the “who processes / where located” story change — or does ownership stay frozen while geography gets louder?

Three information layers force the refresh. Synergy’s inland hyperscale footprint (end-2025 ops / multi-year pipeline) prints hard site counts — 1,360 operational hyperscale data centers worldwide (580 in the US) and 803 pipeline sites (437 in the US) — plus a concentration punchline the August GW note did not feature: Amazon + Microsoft + Google already hold 58% of all hyperscale capacity. Named inland markets (Abilene, Mount Pleasant, South Bend, El Paso, Boone County, Kansas City) and Amazon’s reported ~$15B / ~2.4 GW Northern Indiana add revise the hub panel upward again. Gartner’s June public release of the 1Q26 forecast supplies a US electricity composition the theme had not yet used as a headline meter: the US draws about 204 TWh of the world’s 565 TWh data-center electricity in 2026 (36%), and dedicated AI data centers consume about 68 TWh inside the US — roughly one-third of US data-center power. Ownership Big-5 shares are carried from August; Epoch still has not finalized a Q1/Q2 2026 period Chip Owners print that would rewrite the 71% aggregate.

The headline table: Aug update vs Q3 site & power print

The ownership slogan did not move. Five US hyperscalers still own about seven-tenths of world AI compute on the same ~20 million H100e world stock. What did print is a second geography meter: sites, not just gigawatts. Almost half of Synergy’s hyperscale pipeline sites sit in the United States (437 / 803), even while the US’s share of operational site counts (580 / 1,360 ≈ 43%) undershoots its share of operational capacity (still well over half in Synergy’s capacity framing). That is not a contradiction — it is the size effect: new campuses are nearly twice the average IT capacity of currently operational hyperscale sites, so fewer US pipeline sites can still dominate world watts.

Ownership is carried — the open Epoch quarter is the rewrite risk

August’s Microsoft +2.3 pp was a measurement correction against July’s staff residual, aligned to Epoch Chip Owners Explorer coverage. It was never sold as a Q1→Q2 2026 shipment delta. Epoch’s own hyperscaler-share chart still freezes on Q4 2025 anchors and explicitly excluded unfinished Q1 2026 data when that analysis printed. Until a finalized quarterly ownership vintage lands, this Q3 post refuses to invent a fake period move in Google’s 25%, Microsoft’s 17.3%, Amazon’s 12.5%, Meta’s 11.3%, or Oracle’s 5.3%. The dashboard’s ownership panel is labeled a carry for that reason.

That honesty matters for readers who want a YoY ownership delta. The shareable YoY capacity story remains Gartner’s path — data-center electricity +26% to 565 TWh in 2026, capacity +27% to 132 GW — not a fabricated Epoch quarter. The shareable concentration story that is new this vintage sits one layer down: inside hyperscale capacity, the Top-3 cloud providers alone are 58%. Big-5 chip ownership (~71%) and Top-3 site/capacity concentration (~58%) are different meters on related balance sheets; do not average them into a single “AI monopoly index.”

Synergy’s site ledger: 437 US pipeline sites, Top-3 at 58%

August emphasized Synergy’s ~45 GW US IT pipeline and the inland +20 pp corridor jump. Those numbers remain the capacity anchors. Q3 promotes the site-count companion that was sitting in Synergy’s inland footprint note all along: 1,360 operational hyperscale facilities worldwide at end-2025 (580 US), and a known pipeline of 803 future hyperscale sites (437 US). The US therefore accounts for about 54% of pipeline sites and about 43% of ops sites — while still holding well over half of world operational data-center capacity and 55% of hyperscale capacity in Synergy’s capacity tables.

Two implications follow. First, pipeline geography is more US-heavy on sites than live geography is on site counts, which is consistent with the inland power chase: new gigawatt campuses prefer Texas and the Midwest even when Northern Virginia remains the densest live cluster. Second, Top-3 concentration at 58% of hyperscale capacity means the ownership story and the landlord story are converging on the same three names — Amazon, Microsoft, and Google — even before Meta and Oracle’s chip stocks are added. Meta, Alibaba, Tencent, Oracle, Apple, and ByteDance fill the rest of Synergy’s ranked footprint; the long tail of neoclouds and colo-backed AI campuses shows up more in the 74-company / 45 GW large-DC pipeline than in the hyperscale-only site ledger.

Treat the August “~1,500 large-DC pipeline sites” and the Q3 “803 hyperscale pipeline sites” as different universes. The July 23 capacity release counted a broader large-data-center pipeline (~1,500 worldwide, ~half US). The inland hyperscale note counts hyperscale operators only (803 pipeline / 1,360 ops). Mixing them produces fake deltas. The dashboard stacks US versus rest-of-world on the hyperscale-only ledger and carries the 45 GW figure beside it so readers can see both meters without collapsing them.

Inland hubs revise again: Midwest +2.7 GW, Texas +1.4 GW

The August hub panel already lifted the Midwest to ~12.5 GW and Texas to ~5.8 GW on the inland pipeline intensity. Q3 leans on Synergy’s named non-traditional markets — Abilene, Mount Pleasant, South Bend, El Paso, Boone County, Kansas City — plus the widely reported Amazon Northern Indiana commitment on the order of ~$15 billion and about 2.4 GW of data-center capacity. Our Midwest IT-MW revision therefore moves ~12.5 → ~15.2 GW (+2.7 GW), and Texas ~5.8 → ~7.2 GW (+1.4 GW). Northern Virginia is essentially flat (+50 MW); it remains the densest live hub and is no longer the expansion center of gravity. Middle East announced MW stays flat at ~5.5 GW — energisation, not press releases, is still the open question. Eastern China ticks up modestly on domestic cluster build while Epoch’s ~5% China ownership share remains the chip-stock ceiling.

For campus-by-campus status, the global AI data-center build tracker remains the site list. This update is the vintage delta on concentration, site counts, and hub MW, not a replacement facility census.

Tokens still refuse the ownership pie — Q3 cohort nudge

Chip ownership and token throughput remain different meters. The Q3 brand cohort in the dashboard scatter still shows usage geography that ownership tables cannot reproduce. Google sits near the diagonal. OpenAI’s token share ticks to about 15% with still ~0% owned H100e on Epoch’s pie. ByteDance remains far to the right on tokens inside China’s ~5% ownership envelope. Microsoft and Amazon sit above the diagonal: they own more of the world’s AI compute than their first-party token meters imply, because they rent the rest to frontier labs and enterprise tenants.

That divergence is the policy trap. Export controls and fab rules target owners; app-store and model-routing rules target token surfaces. Neither alone maps the system the brand token series and the Epoch ownership pie jointly describe. Do not average ownership shares with token shares into a fake “AI power” index — the scatter exists to stop that habit.

Power composition: AI crosses conventional in 2027; US is 36% of DC TWh

Gartner’s capacity path is unchanged as a world vintage — 132 GW / 565 TWh in 2026, ~165 GW / 702 TWh in 2027, ~290 GW by 2030 — but this Q3 post features the composition hinge and the US electricity slice. AI-optimised servers rise from 95 TWh (2025) → 175 TWh (2026) → 258 TWh (2027) while conventional servers stay nearly flat (193 → 195 → 200 TWh). 2027 is the year AI-optimised server electricity is forecast to surpass conventional servers. Cooling and other infrastructure climb with denser racks (159 → 195 → 243 TWh).

Inside that world total, the United States accounts for about 204 TWh in 2026 — 36% of global data-center electricity — and dedicated AI data centers consume about 68 TWh of that US total (~one-third). That is a different claim from the theme’s older ~45% US share of AI data-center capacity by power: one is a Gartner electricity composition for all data centers; the other is an AI-specific capacity share. Keep them separate. Pair the global path with the AI power-grid concentration cut for where interconnection binds, and with hyperscaler capex intensity for the dollar flow that funds the next GW.

What would rewrite this update

1. Epoch Q1/Q2 2026 Chip Owners finalization moves the Big-5 aggregate materially below ~70% or above ~73%the first true period ownership delta since the Q4 2025 freeze.
2. Google’s TPU H100e conversion revises the 5.0M / 25% anchor; every other share moves with the world stock.
3. Synergy / utility energisation prints show Middle East or India multi-GW IT loads connecting on schedulecompressing US operational capacity share — or slipping again, which sticks the US >50% ops-capacity story.
4. Top-3 hyperscale share prints below ~55% or above ~62% on a new Synergy vintagerewriting the landlord concentration punchline.
5. Token–ownership convergence: if OpenAI or Anthropic begin owning a material H100e stock on Epoch’s table, the scatter’s “renter” cluster collapses.
6. Gartner 2H26 vintage revises the 2027 AI-surpasses-conventional hinge or the US 204 TWh slice.

Until those print, the live frame versus August is narrow: ownership Big-5 shares are carried near 71%; the new vintage is Synergy’s hyperscale site ledger (437 US of 803 pipeline sites) and Top-3 cloud at 58% of hyperscale capacity; inland hubs revise up again on named Midwest/Texas markets; and Gartner’s US electricity composition puts the US at ~36% of world DC TWh with dedicated AI already ~1/3 of US DC power.

Caveats and methodology

1. Ownership carry ≠ stagnation claim. Flat Big-5 shares mean no new Epoch period print, not proof that installs stopped. Treat August’s Microsoft +2.3 pp as a restatement, not Q3 momentum.
2. Hyperscale site counts ≠ large-DC pipeline counts. Synergy’s 803 / 1,360 hyperscale ledger is not the July ~1,500 large-DC pipeline universe. Deltas across those frames are invalid.
3. Site share ≠ capacity share. US ops sites at ~43% can coexist with US ops capacity well over 50% because new campuses are larger (~2× average IT capacity).
4. Hub MW figures mix live, building, and announced IT loads and are order-of-magnitude revisions, not utility interconnection queues. The Indiana ~2.4 GW figure is reported investment capacity, not a Synergy census cell.
5. H100-equivalent ≠ identical utility. Epoch converts on peak 8-bit FLOP/s. Memory bandwidth, networking, and software stacks differ across TPU, Blackwell, Trainium, and Ascend.
6. World stock ~20M H100e is implied by Google ≈ 25% / 5.0M; if that anchor moves, every share moves.
7. Gartner US TWh (204 / 565; AI ~68 TWh inside the US) is a public composition slice from the June 2026 releasenot LBNL US historicals and not identical to the theme’s ~45% AI-capacity share.
8. Token shares in the scatter are illustrative slices of the tracked brand cohort, not a regulator census.

Primary sources: Synergy Research inland hyperscale investment analysis (ops/pipeline site counts; Top-3 58%); Synergy US capacity update (23 Jul 2026, ~45 GW pipeline carried); Epoch AI Chip Owners / hyperscaler share chart (Q4 2025 anchors, carried); Gartner Forecast: Data Center Power Capacity and Consumption, Worldwide, 2024–2030 (1Q26 / Jun 2026 public release); prior theme posts Aug compute update and July research baseline.

The shareable takeaway

Versus August, Big-5 ownership is unchanged near 71% — the Q3 vintage is sites and power composition. Synergy’s hyperscale ledger puts 437 of 803 pipeline sites in the US, Top-3 cloud at 58% of hyperscale capacity, and new campuses at roughly 2× current average size. Inland hubs revise up again (Midwest +2.7 GW, Texas +1.4 GW). Gartner’s US slice is ~36% of world data-center electricity, with dedicated AI already about one-third of US data-center power. For the explorer ownership restatement see the August update; for the baseline essay see July research; for campuses see the build tracker; for tokens see the brand series.