Charted: US Holds 45% of AI DC Capacity — China Origin Clears 62% of Tokens

Our theme research answered who owns AI compute — Big-5 hyperscalers near 71% of world H100e stock, Google alone near 25%. The concentration lens then measured how thick the tip is across ownership, cloud capacity, regional power draw, and token brands. The August location update printed Synergy’s live market rankings: top-20 markets 60%, Northern Virginia + Greater Beijing 17%, pipeline 915 sites. This post asks the desk question that sits beside those prints: where does activity, risk, and capacity sit on the map — and do token geography and site geography agree?

The interactive dashboard above is built as a geography lens. Toggle Regions, Metros, Tokens, and Pipeline. On regions flip share versus GW draw; on metros flip capacity, YoY growth, or pipeline weight and filter the capacity×growth scatter; on tokens filter brand origin. The punchline is deliberately multi-map. On AI DC capacity, the United States leads at about 45% of global power draw. On pipeline sites, the US is still heavier at ~54% of 915 known hyperscale facilities. On live dual-hub capacity, Northern Virginia + Greater Beijing alone clear ~17%. On token origin, China-origin brands clear ~61.8% of the June 2026 major-brands series — while US chip ownership HQ still owns ~71% of the H100e stock. Capacity, tokens, sites, and ownership are four different maps.

The headline ladder: regional and origin shares that disagree

Read the table as a family of maps, not one slogan. Capacity geography tells you which regions absorb AI-relevant power draw. Metro geography tells you which campuses densify live hyperscale. Token geography tells you which brand origins dominate routed inference and product surfaces. Pipeline geography tells you where the next sites are being planned. Ownership geography tells you which corporate HQs hold the chips. Averaging these rows into “AI compute is an American story” or “AI compute is a China story” is a category error — both can be true on different ledgers.

Regions: US capacity lead, Middle East announced-heavy

Open Regions. Against a ~41 GW AI-relevant DC perimeter, the United States rolls to about 18.5 GW (~45%), China about 7.4 GW (~18%), Europe about 5.8 GW (~14%), the Middle East about 4.1 GW (~10%), rest-of-APAC about 3.7 GW (~9%), and LatAm & other about 1.6 GW (~4%). Flip the metric to GW draw and the same ladder appears in absolute power rather than share — useful when desks need MW for interconnect models rather than percent for narrative slides.

The capacity pie is therefore US-weighted without being a monopoly. A system where one country holds a little under half of estimated AI DC capacity still leaves China + Europe + MEA + APAC with a material majority of regions and a large absolute MW tip. That is the geographic twin of the concentration scoreboard: Google remains the largest single owner; rolling Google, Microsoft, Amazon, Meta, and Oracle onto the same ownership sheet makes the US the thickest HQ region without claiming that “most AI tokens are American.”

The stacked path from 2022→2026 is an editorial regional share track, not a new microdata extract. The US share drifts from the high-40s toward ~45% as Middle East announced campuses thicken and China holds the high-teens. Treat the path as a shape check against the live 2026 snapshot — useful for seeing that MEA’s rise is real on announcements, not proof that live FLOP/s already relocated.

Metros: dual-hub densification and inland growth premiums

Switch to Metros. Synergy’s 19 August 2026 location rankings put Northern Virginia near ~9% of hyperscale capacity and Greater Beijing near ~8% — together the 17% dual-hub print that also appears in the August update. Oregon, Iowa, Ohio, and Dallas–Fort Worth fill the next band. Dublin remains the sole European seat in the top-20 — a reminder that “Europe” as a regional capacity story is not the same as “Europe” as a Synergy top-20 metro story.

Toggle the metro metric to YoY growth. Texas ops aggregate prints about +71% versus a world ops average near +36%. Iowa and Ohio sit in the +40% band; Dublin is softer near +18%. The capacity×growth scatter makes the inland shift visual: NoVA and Beijing sit right on capacity with mid-20s/low-30s growth; Texas sits mid-capacity with extreme growth; Dublin sits left-and-low. Bubble size tracks pipeline weight — the desks that only watch live capacity underweight where the next interconnect queues are forming.

Filter the scatter to United States and the inland corridor story sharpens further: Midwest and Texas nodes pull share from coastal densification without yet matching NoVA’s live tip. That is facility geography, not a full company segment breakout — and capacity hints inside disclosed Synergy bands should not be quoted as audited MW extracts.

Tokens: China-origin lead against US ownership HQ

Open Tokens. The June 2026 major-brands series (~18,503T tokens/mo perimeter) rolls to China origin ~61.8%, United States ~37.7%, and a thin residual. ByteDance alone is ~29.2% of the brand ladder — Top-1 on tokens without Top-1 chip ownership. Google is ~19% of tokens while owning ~25% of chips. Alibaba (Qwen) ~11.3% and OpenAI ~9.2% complete the next tier.

Filter brands to China or US to see the origin stacks without the residual. The ownership-by-HQ bar beside the token pie is the post’s central hinge: US Big-5 HQ geography still clears ~71% of world H100e stock, China aggregate owners ~5% — almost the inverse of token-origin leadership. The token consumption series already showed brand ladders; geography makes the origin disagreement operational for desks that confuse “who owns GPUs” with “whose products route the tokens.”

Treat token origin as brand HQ / product-surface geography, not as a claim about where every inference FLOP physically runs. Cross-border routing, rented Azure/Oracle capacity, and open-weight dilution mean a China-origin brand can serve users worldwide — and a US owner can rent chips to a US brand that does not appear in the ownership ladder at all.

Pipeline: US overweight on sites versus already-live share

Toggle Pipeline. Known hyperscale pipeline facilities print 915 sites (+112 versus the Q3 ledger of 803). Desk regional roll-ups put the United States near ~54% of those sites versus ~45% of already-live AI DC capacity — a build overweight. China is closer to balanced (~16% sites vs ~18% live). Europe sits slightly underweight on sites versus live share. Middle East remains announced-heavy: high conversion risk relative to inland US corridors that already have interconnect and power programs in flight.

The grouped bars (pipeline sites % vs live capacity %) are the operational view for power and permitting desks. A region that is overweight on pipeline relative to live capacity is where interconnect queues, water permits, and labour markets will feel the next shock first. Pair this with the global AI data-center build tracker for campus-level colour; this geography lens keeps the regional mismatch readable in one chart.

Growth callouts stay sticky from the August vintage: Texas ops +71% YoY, 15 of 20 Synergy top-20 seats in the United States, top-20 markets still ~60% of hyperscale capacity, top-40 near ~79%. Pipeline geography does not rewrite live dual-hub densification — it tells you whether the next MW reinforce the same tip or migrate inland.

Five meters, one category error to avoid

Return to Regions and read the five geography meters together. AI DC capacity Top-1 is the US at ~45%. Dual-hub metros are NoVA + Beijing at ~17%. Token origin Top-1 is China at ~61.8%. Pipeline sites Top-1 is the US at ~54%. Chip ownership HQ Top-1 is US Big-5 at ~71%. Desks that average these into a single “AI geography is concentrated in X” sentence will mis-price power risk, export-control exposure, and product-surface competition in the same week.

The Q3 site-count update and August rankings already moved the live market and pipeline clocks. Geography does not replace those vintages; it asks which map a desk is actually using when it says “compute is concentrated.”

Caveats and methodology

• Regional capacity shares are industry-tracker power-draw synthesis carried from theme researchorder-of-magnitude GW, not a utility census. Cross-check against company disclosures before treating any single region cell as audit-grade.
• Metro capacity hints sit inside Synergy’s disclosed rank bands (dual-hub, top-6, top-20). Percent labels are illustrative within those bands, not audited market-share extracts.
• Token origin rolls the June 2026 major-brands series by brand HQ / product origin. It is not a traceroute of every inference GPU, and open-weight routing can dilute brand credit.
• Pipeline regional shares are desk roll-ups of the Synergy site ledger plus theme carry. Announced Middle East campuses carry higher conversion risk than inland US corridors with live interconnect programs.
• Ownership HQ shares are Epoch Chip Owners carries (Q4 2025 anchors + August explorer restatement). Ownership geography ≠ deployment geography ≠ token geography.
• Editorial 2022–2026 regional paths are shape checks, not a new panel dataset. Do not splice them into Synergy’s August point-in-time rankings as if they shared a methodology.

What to watch next

Watch three geography meters, not one. First, whether inland US growth (Texas +71% ops YoY, Midwest pipeline weight) erodes NoVA’s live tip without flipping the US regional capacity lead. Second, whether China-origin token share stays above ~60% even while US ownership HQ remains near ~71% of H100e — the ownership≠usage hinge from concentration, now drawn as a map. Third, whether Middle East announced MW converts into live power-draw share or stays a pipeline story. For top-of-distribution HHI keep the concentration series open; for live rankings keep the August update open; for brand ladders keep the token series open.

Related reading: theme research · concentration lens · August location rankings · Q3 site ledger · build tracker.