Update: Microsoft +2.3 pp of World AI Compute — Big-5 Still ~71%, US Pipeline Hits 45 GW

What changed since the July research vintage

In late July we published the theme’s baseline: five hyperscalers own ~71% of global AI compute, with Google alone near 25%, China as a whole near 5%, and the United States hosting roughly 45% of AI data-center capacity by power draw. That post answered the stock question — who owns the chips, and where do the watts sit? This update answers the flow question markets and planners actually trade on: what moved in the newest official vintage, and does the ownership–location story still hold once individual H100e levels and the site pipeline are restated?

Three information events force a refresh. Epoch’s Chip Owners Explorer published individual Big-5 H100-equivalent levels that replace several of July’s estimated residual splits (Google and the Big-5 aggregate remain the disclosed anchors). Synergy Research (23 Jul 2026) printed a US data-center capacity update with an explicit ~45 GW IT pipeline and an inland investment shift that revises our hub panel. Token meters through June 2026 (same theme’s brand token series) keep widening the gap between who owns silicon and who burns tokens. The dashboard above is built as a vintage delta — prior bars, new bars, hub MW revisions, geography concentration deltas, and a tokens-versus-ownership scatter.

The headline table: Jul research vs Aug explorer restatement

The concentration slogan did not break. Five US hyperscalers still own about seven-tenths of world AI compute on the same ~20 million H100e world stock implied by Google’s quarter-share. What changed is the internal ranking: Microsoft’s explorer level (~3.45M H100e, “just under 3.5M”) is the cleanest upward restatement in the set, while Amazon and Oracle’s July residual estimates were too high. Meta’s move from a round ~10% / 2.0M narrative to ~2.25M is a smaller upward tick, not a regime change.

Treat the Microsoft +2.3 pp carefully. It is mostly a measurement correction against July’s staff residual, not a claim that Microsoft alone installed 450k H100e between the two posts. True period momentum still lives in Epoch’s next finalized quarterly share print (Q1/Q2 2026 were explicitly excluded from the 71% chart when that analysis froze). Until that quarter closes, the shareable frame is: Big-5 concentration stable; intra-Big-5 ranking restated toward Microsoft and Meta, away from Amazon and Oracle.

Microsoft’s restatement is the cleanest ownership delta

July forced a residual math problem. Epoch disclosed Google (~25%) and the Big-5 aggregate (~71%); Meta’s ~10% was a narrative anchor. The leftover ~46 percentage points had to be split across Microsoft, Amazon, and Oracle. We put Microsoft at 15% / 3.0M. The explorer coverage that landed in the spring — and that we are promoting into this update — puts Redmond closer to 3.5M H100e, almost entirely Nvidia with a thin AMD slice. On a 20M world stock that is ~17.3%, a +2.3 pp vintage delta versus the prior print.

That restatement matters for two adjacent stories. First, ownership versus rental: OpenAI’s access still rides Microsoft, Oracle, and CoreWeave campuses; a higher Microsoft owned stock raises the ceiling on how much of that rented pool is parent-owned silicon rather than neocloud offtake. Second, capex intensity: the same firm that just printed FY26 intensity at 26.8% is also the firm whose H100e stock was understated in our July residual. Absolute spend and absolute ownership are converging on the same name.

Amazon’s explorer level (~2.5M) revises the July 2.8M estimate down. That does not mean AWS stopped building — Trainium and Nvidia continue to ship — it means the July residual over-allocated the Big-5 leftover. Oracle’s drop from 1.4M estimated to just over 1.0M restated is the largest downward confidence correction; Stargate headlines still inflate announced MW faster than Epoch’s ownership meter. Meta’s +0.25M is consistent with continued GPU/AMD installs plus cloud deals that rent additional capacity without always showing up as Meta-owned H100e.

Synergy’s site ledger: 45 GW US pipeline, inland +20 pp

Ownership answers who. The July post’s geography half leaned on an industry-tracker ~45% US share of AI data-center capacity by power. That AI-specific share is unchanged in this update — we do not invent a new AI-only percentage without a new tracker vintage. What did print in July 2026 is Synergy’s broader capacity ledger, and it is loud enough to revise the site side of the theme.

Synergy’s 23 July release puts the known pipeline of future large data centers near 1,500 worldwide, with almost half in the United States. That US half represents about 45 GW of IT capacity across 74 companies — seven hyperscalers driving AI campuses plus 67 other large builders. Hyperscaler-owned US capacity is expected to double in about two years; total US data-center capacity to double in about three. Over the next five years the US is expected to remain well over half of the world’s operational data-center capacity. Separately, Synergy’s end-2025 hyperscale footprint shows the US at 55% of worldwide hyperscale operational capacity (up from 52% three years earlier), with hyperscale operators already 48% of all world data-center capacity (on a path toward 67% by 2031).

The inland shift is the geographic punchline. At end-2025, Texas and the Midwest accounted for 33% of operational US hyperscale capacity — but 53% of Synergy’s pipeline of new US hyperscale capacity. That is a +20 percentage-point gap between where the live load sits and where the next gigawatts want to land. Northern Virginia remains the densest live cluster; Texas is the single most prominent state in the pipeline. Wisconsin, Indiana, Michigan, and Missouri all gain importance as Amazon, Google, Meta, Microsoft, OpenAI, and CoreWeave chase power availability inland. Our hub panel therefore revises Midwest IT MW from ~9.8 GW → ~12.5 GW (+2.7 GW) and Texas from ~4.2 → ~5.8 GW (+1.6 GW), while Middle East announced MW is held flat — energisation, not press releases, is still the open question. For campus-by-campus status, the global AI data-center build tracker remains the site list; this update is the vintage delta on concentration and pipeline.

Tokens still refuse to copy the ownership pie

Chip ownership and token throughput are different meters. June 2026 brand disclosures in our token-consumption series keep showing usage geography that ownership tables cannot reproduce. Google sits near the diagonal — large owner, large token surface. OpenAI and ByteDance sit far to the right on tokens with little or no owned H100e on Epoch’s pie. 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.

China’s ownership share remains about 5% of world AI compute — smaller than Google alone — even as Chinese apps and open-weight routing can dominate neutral token meters such as OpenRouter. That divergence is not a data error. Export controls, custom-silicon programs, and cloud contracts split balance-sheet ownership from routed inference. Policy that targets owners and fabs will miss a large slice of token demand; policy that targets apps will miss who actually holds the accelerators. The dashboard scatter is the reminder: do not average ownership shares with token shares into a fake “AI power” index.

Power capacity: Gartner path intact, 2027 is the composition hinge

Gartner’s 1Q26 worldwide forecast is unchanged as a capacity vintage — we are not restating 132 GW / 565 TWh — but this update features the 2027 step that July treated as background. Electricity rises from 565 TWh (2026) to 702 TWh (2027); AI-optimised servers are on track to surpass conventional servers on power consumption in 2027, after already taking 31% of data-center power in 2026. Capacity still races from 104 GW (2025) → 132 GW (2026) → ~165 GW (2027 est.) → ~290 GW (2030). Capacity (GW) remains the interconnection constraint; TWh remains the energy bill. Pair the global path with the AI power-grid concentration cut for where those watts concentrate, 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%a true period delta, not a residual restatement.
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 share — or slipping again, which sticks the US >50% ops story.
4. Token–ownership convergence: if OpenAI or Anthropic begin owning a material H100e stock on Epoch’s table, the scatter’s “renter” cluster collapses.
5. Gartner 2H26 / 2027 vintage revises the 702 TWh / AI-surpasses-conventional hinge.

Until those print, the live frame versus July is narrow: Big-5 ownership concentration is stable near 71%; Microsoft’s explorer restatement is +2.3 pp; Amazon and Oracle’s July residuals were high; the US site pipeline is ~45 GW with TX+Midwest taking 53% of new hyperscale capacity; and token demand still refuses to copy the silicon balance sheet.

Caveats and methodology

1. Explorer restatements ≠ quarterly installs. Microsoft’s +2.3 pp is primarily a correction to July’s estimated residual, aligned to Epoch Chip Owners Explorer coverage (via Epoch / Network World Apr 2026 reporting). It is not a measured Q1→Q2 2026 shipment delta.
2. 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.
3. World stock ~20M H100e is implied by Google ≈ 25% / 5.0M; if that anchor moves, every share moves.
4. Synergy meters are not Epoch meters. “US ~45% of AI DC capacity by power” (prior theme) is not identical to “US well over half of world operational DC capacity” or “US 55% of hyperscale capacity.” We keep them separate in the geography panel.
5. Hub MW figures mix live, building, and announced IT loads and are order-of-magnitude revisions, not utility interconnection queues.
6. Token shares in the scatter are illustrative slices of the tracked brand cohort in the token post (Jun 2026), not a regulator census; ByteDance’s meter includes surfaces that inflate token counts relative to chat-only Western APIs.
7. China ~5% ownership remains an Epoch ownership narrative, not a claim about Chinese model quality or token volume.
8. Gartner GW/TWh figures are the June 2026 public release of the 1Q26 forecast; they are not LBNL US historicals and not Goldman capex dollars.

Primary sources: Epoch AI Chip Owners / hyperscaler share chart (Q4 2025 anchors); Epoch Chip Owners Explorer individual levels as reported in Network World (Apr 2026); Synergy Research US capacity update (23 Jul 2026) and inland hyperscale investment analysis; Gartner Forecast: Data Center Power Capacity and Consumption, Worldwide, 2024–2030 (1Q26); prior theme baseline in compute demand research.

The shareable takeaway

Versus July, the Big-5 still own ~71% of world AI compute — but Microsoft’s explorer restatement lifts it to ~17.3% (+2.3 pp), while Amazon and Oracle revise down. Google remains ~25%. Synergy’s July site ledger adds a ~45 GW US IT pipeline and a +20 pp inland jump (Texas + Midwest from 33% of ops to 53% of pipeline). Token demand still diverges from ownership. For the baseline ownership essay see the July research post; for sites see the build tracker; for tokens see the brand series.