Concentration: Google Owns 25% of AI Compute — Top-3 Holds 55%; Tokens Tell a Different Story

The concentration question, not the stock question

The theme’s July research essay answered who owns the chips and where the watts sit. The August explorer update and Q3 / Aug location vintages restated Microsoft’s H100e share and Synergy’s site rankings. This post asks a sharper distribution question: how concentrated is the system at the top? Top-1, Top-3, and HHI — across ownership, hyperscale capacity, regional power draw, and token throughput.

The interactive dashboard above is built as a concentration lens. Toggle Scoreboard, Ownership ladder, Sites & regions, and Tokens vs chips. The punchline is deliberately multi-sided. On chip ownership, Google alone is about 25% and the top three clear ~55%. On hyperscale cloud capacity, AWS + Azure + Google still hold 57%. On geography, the United States hosts ~45% of AI data-center capacity by power draw, and fifteen of twenty largest hyperscale markets. On tokens, ByteDance leads at ~29% and China-origin brands clear ~62% of June 2026 volume — while China as a chip owner is still near 5%. Same theme, four different tops.

The 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. Regional concentration says where the megawatts live. Token concentration says who processes the prompts — and it does not match the ownership ladder. Analysts who quote only “Big-5 own 71%” understate how top-heavy the inside of that 71% is; analysts who quote only ByteDance’s token lead understate how little chip stock that usage owns.

Ownership: Google is Top-1, the tip is sticky

Toggle Ownership ladder. Epoch’s Chip Owners framework (Q4 2025 anchors, Aug explorer restatement carried through the August 202608 location update) puts Google near 25% of cumulative world AI compute — mostly custom TPUs. Microsoft sits at 17.3% after the explorer restatement that lifted it roughly +2.3 pp versus the July residual. Amazon 12.5%, Meta 11.3%, and Oracle 5.3% complete a Big-5 bloc of 71.4%. China as an aggregate owner remains near 5%.

That produces the ownership concentration print: 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 shows cumulative mass rising well above the equal-split line by rank three. Pair this with the research baseline: the Big-5 share path from 63% in Q1 2024 to 71% by Q4 2025 was already a concentration story; the Aug carry freezes the level while this lens ranks the tip.

Ownership is not usage. Frontier labs — OpenAI, Anthropic, xAI — rent most of their operational capacity from Microsoft, Oracle, Amazon, Google, and neoclouds. A lab can dominate the product narrative while owning almost none of the H100e stock. That is why the token scatter exists.

Sites and clouds: Top-3 still dominate, markets stay tight

Open Sites & regions. Synergy’s 19 August 2026 hyperscale location rankings — the newest site vintage in the Aug update — put AWS + Azure + Google Cloud at 57% of hyperscale capacity, down 1 pp from the Q3 58% print. That is still a concentrated operator oligopoly. The market-band chart is sharper on place: top-20 markets hold 60% of capacity; Northern Virginia + Greater Beijing alone are 17%; top-40 reach 79%. The United States fills 15 of 20 largest-market seats. Pipeline sites rose 803 → 915 (+112) while concentration at the tip barely budged.

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 therefore clear about 77%. Texas operational capacity growing +71% YoY against a +36% world average is a growth concentration story inside an already US-heavy stock: the inland corridor is where new megawatts land, not where the ownership tip breaks.

Geography and ownership can diverge. US-HQ operators dominate hyperscale revenue and chip stock, but China-origin token brands can lead usage while sitting inside a thin ownership slice. The build tracker keeps the campus-level map; this lens ranks how much of the capacity distribution sits in the first few seats.

Tokens: ByteDance is Top-1 — ownership does not follow

Switch to Tokens vs chips. The June 2026 slice of the major AI brands token series totals about 18,500 trillion tokens/month across tracked brands. ByteDance (Doubao) leads at ~29.2%, Google follows at ~19%, Alibaba (Qwen) at ~11.3%. Top-3 brands therefore clear ~59.6%. China-origin brands as a group hold about 61.8% of that month’s volume; US-origin brands about 37.7%.

That is the divergence that makes concentration analysis useful. Google is #1 in chips (~25%) and #2 in tokens (~19%) — roughly aligned. ByteDance is #1 in tokens (~29%) with a negligible ownership seat inside China’s ~5% aggregate. OpenAI holds mid-single-digit token share with no disclosed chip ownership. Microsoft and Amazon sit ownership-heavy relative to first-party token share because Azure and AWS rent capacity to other brands. The scatter’s upper-left is usage without chips; the lower-right is owns more than it first-parties.

Do not average token Top-1 with ownership Top-1. They answer different questions: who processes prompts this month versus who holds the depreciating capital stock. Policy that targets “AI concentration” without naming the perimeter will pick the wrong instrument.

Who is exposed under a concentrated tip

More exposed: enterprises and governments whose AI roadmaps assume a plural supplier set when three cloud operators still clear ~57% of hyperscale capacity; power planners who underweight Northern Virginia / inland Texas / Midwest corridors while 15 of 20 largest markets sit in the US; chip-allocation desks that treat China-origin token growth as evidence of China chip stock (it is not — ownership stays near 5%); and investors who read OpenAI product share as OpenAI balance-sheet compute.

Relatively better positioned: operators already inside the Top-3 cloud perimeter with contracted power and fiber in dual-hub and inland metros; buyers who multi-home inference across Google / Azure / AWS and China MaaS surfaces when latency and data residency allow; and analysts who keep separate ledgers for ownership, sites, regions, and tokens.

What would change the story: a new Epoch period print that drops Google below ~20% or the Big-5 below ~60%; Synergy Top-3 cloud falling through 50% as neoclouds and sovereign builds scale; US regional AI DC share falling materially below 35%; or token Top-1 dispersing so no brand holds more than ~15% of the June-style series. None of those appear in the mid-2025 to mid-2026 vintages summarised here.

Caveats and methodology

• Perimeters are not interchangeable. Ownership is H100-equivalent stock; cloud Top-3 is hyperscale capacity; regional shares are AI-relevant power draw; tokens are monthly throughput across vendor surfaces.
• Ownership shares for residual non-Google entities below the Big-5 aggregate are staff-aligned estimates that sum to Epoch’s disclosed hyperscaler total; treat Top-3 as order-of-magnitude concentration, not a prospectus table.
• Synergy market capacity hints inside the dual-hub / top-20 bands are illustrative within disclosed concentration prints (17% / 60% / 79%).
• Token figures include internal workloads (Search AI Overviews, recommendation, moderation). US and China headlines inflate the same way, so ratios remain useful; absolute “economy tokens” are overstated.
• Open-weight routing can double-count tokens between model authors and serving clouds.
• Chinese text tokenization differs from English; equal token counts are not equal work.
• Epoch period print for Q1/Q2 2026 is still openAug 202608 carries Big-5 ownership; the location vintage does not restate H100e.
• This post is a concentration lens. For the ownership essay use July research; for explorer deltas use the August update; for site rankings use the Aug 202608 update; for tokens use the brand series.

The shareable takeaway

AI compute demand is concentrated at the top of the distribution — but the name of the top depends on the meter. Chip ownership Top-1 is Google at 25%; Top-3 is ~55%; Big-5 hold ~71% with within-cohort HHI ≈ 2,400. Hyperscale Top-3 cloud is still 57%, top-20 markets 60%, and the US ~45% of AI DC capacity by power draw. Token throughput tells the other story: ByteDance ~29% Top-1, Top-3 brands ~60%, China-origin ~62% — while China ownership stays near 5%. For the stock essay keep research open; for sites keep the Aug location update; for usage keep the token series.