Charted: Five Hyperscalers Own 71% of the World’s AI Compute

Token charts answer who serves how many prompts. Capex decks answer who spends. The question that still gets blurred is physical and proprietary at once: who owns the chips that do the work, and where do those watts sit? Epoch AI’s Chip Owners hub puts a hard ownership number on the first half. As of Q4 2025, five US hyperscalers — Amazon, Google, Meta, Microsoft, and Oracle — collectively hold an estimated 71% of the world’s dedicated AI compute, measured in Nvidia H100-equivalents. That is up from 63% in Q1 2024. Google alone is about one quarter of the global stock — roughly 5 million H100e, most of it from custom TPUs, not rented Nvidia clusters.

The geography half is a different meter. Industry trackers citing Gartner, IDC, and grid filings put the United States at roughly 45% of global AI data-center capacity by power draw. Gartner’s 1Q26 capacity forecast has world data-center power demand at 104 GW in 2025 and 132 GW in 2026 (+27%), on a path toward ~290 GW by 2030, with AI-optimised servers already 31% of data-center power consumption in 2026. Ownership concentration and site concentration reinforce each other: the firms that own the chips also dominate the campuses where those chips plug in.

This post is deliberately not a retread of our global AI data-center build tracker (site-by-site MW and status) or our major AI brands token-consumption series (routed tokens by brand). Tokens can surge in China while chip ownership stays US-heavy. Announced campuses can inflate MW maps while energised IT load lags. Here the meters are H100e ownership, regional AI capacity share, and Gartner power capacity — with an explicit split between owning compute and using it.

The 71% number, not the “AI arms race” slogan

The Big-5 aggregate (71%) and Google’s (~25%) figures are the Epoch anchors. The Microsoft / Amazon / Oracle / Meta split of the residual 46 percentage points is a staff-aligned estimate that sums exactly to that aggregate — useful for ranking, not for litigation-grade precision. Meta’s ~10% ownership is separately consistent with Epoch’s narrative that Meta owned on the order of ~2 million H100e at end-2025 before newer cloud deals fully ramp.

What the table forces into view: China’s entire ownership stack (~5%) is smaller than Google alone. That does not mean Chinese models or token volume are small — our token post documents the opposite on API routing. It means the silicon balance sheet is still overwhelmingly held by US hyperscalers and their custom-chip programs. Huawei can lead China’s domestic FLOP/s mix after H20 export controls and still leave China far behind on the global ownership pie.

Ownership is not usage — frontier labs rent the stack

Epoch’s follow-on analysis of frontier labs is the corrective to “OpenAI has the most GPUs” folklore. OpenAI disclosed enough data-center electric power capacity for Epoch to convert access to roughly 1.7 million H100e — large, but still a fraction of Google’s owned stock, and mostly rented from Microsoft, Oracle, and CoreWeave. Anthropic likely clears more than 1 million H100e, again mostly via Google and Amazon. xAI is the odd case that owns a large share of what it uses (Colossus Memphis / Southaven). Inside Google and Meta, frontier labs compete with Search ranking, ads, Reels, and cloud customers for the same parent pools — Epoch’s first-pass guess is that DeepMind / Meta Superintelligence each use on the order of half of their parent’s owned compute, not all of it.

Put differently: the companies that train and serve frontier models are not always the companies that own the accelerators. The dashboard’s Own vs use scatter separates access (horizontal) from an ownership proxy (vertical). OpenAI and Anthropic sit far right on access but low on ownership; Google DeepMind sits high on both because the parent owns the TPU fleet.

That split matters for industrial policy and for markets. Export controls bite owners and fabs. Cloud contracts and offtake agreements move users. Capex intensity charts (see our hyperscaler capex intensity and spend map posts) track the dollar flow that creates ownership — but dollars authorized are still one step upstream of H100e installed and one step further upstream of tokens served.

Where the watts sit: US ~45%, hubs that strain grids

Ownership answers who. Power-draw geography answers where. Trackers synthesising Gartner/IDC/LBNL put the United States near 45% of global AI data-center capacity by power. China, Europe, the Middle East, and the rest of Asia-Pacific fill the remainder in our residual regional panel — those non-US shares are order-of-magnitude estimates, not a second Epoch disclosure.

Inside the US, the map is no longer “Northern Virginia plus leftovers.” Traditional cloud metros still host enormous live load, but the AI-relevant build wave is Midwest and Sun Belt heavy: Microsoft Fairwater (Wisconsin), AWS Project Rainier (Indiana / Mississippi), Meta Hyperion (Louisiana) and Prometheus (Ohio), OpenAI/Oracle Stargate campuses in Texas and beyond, xAI Colossus in Tennessee/Mississippi. Announced Middle Eastern programs (UAE Stargate with G42, Saudi Humain, NEOM DataVolt) can look enormous on paper — announced MW ≠ energised MW — which is why the hub panel tags status as live-heavy, building, or announced.

Europe’s story is denser and more constrained: Ireland, the Nordics, the UK, and new French/Portuguese campuses face permitting, power-price, and political pushback even when national grids look fine on average. Singapore and similar city-states show the same pattern at smaller scale — local interconnection, not national generation, is the binding constraint. Our separate US data-center power vs grid piece zooms into the American transmission gap; this post keeps the global ownership + location frame.

Power capacity: 104 → 132 GW, AI servers at 31%

Gartner’s worldwide forecast is the cleanest public capacity series for this theme:

• 2025: ~104 GW data-center power demand; 447 TWh electricity.
• 2026: ~132 GW (+27%); 565 TWh (+26%).
• 2030: ~290 GW capacity; electricity on a path past 1,200 TWh in some Gartner extensions.
• AI-optimised servers: 31% of data-center power consumption in 2026, on track to surpass conventional servers in 2027.

Capacity (GW) is the grid’s peak obligation; terawatt-hours are the energy bill. Both are rising, but capacity is what decides whether a campus gets an interconnection agreement. The dashboard’s power panel plots GW bars against the AI server-share line so the composition shift is visible: the sector is not merely growing — the AI slice inside it is growing faster (industry trackers cite AI-focused electricity growing nearly 3× overall data-center electricity in 2025).

Do not confuse this with LBNL’s US-only data-center TWh path or with Goldman’s hyperscaler dollar capex. Gartner’s perimeter is global data-center power; Epoch’s perimeter is AI chip ownership. They answer adjacent questions and should not be averaged into a fake “AI share of GDP” mashup.

Training vs inference: the demand that relocates

Volume mix inside the AI slice is shifting. Industry syntheses put training as the majority of AI compute in 2023, roughly parity by 2025, and inference near two-thirds by 2026 as chat, API, and agentic workloads compound. Training clusters want contiguous, high-bandwidth fabrics — the multi-gigawatt campus. Inference wants latency, price, and power availability, which pulls capacity toward secondary metros, renewable-adjacent sites, and eventually more distributed formats.

That is why “who processes how much” diverges by workload. A training run for a frontier model still concentrates on a handful of owner campuses. A global inference fleet can be routed across many regions — and across rented neocloud capacity that never shows up as “OpenAI-owned” in Epoch’s ownership table. Token-leaderboards can therefore flip geography (China vs US API routing) without flipping the silicon ownership pie.

What would rewrite the map

Several observables would force an update to this ownership-and-location story:

1. Epoch Q1/Q2 2026 Chip Owners revisions show Big-5 share rolling over below 70%neoclouds and sovereign buyers catching up faster than hyperscaler installs.
2. Google TPU vs Nvidia mix shifts so Google’s H100e lead narrows even if its dollar capex stays high (conversion assumptions matter).
3. China ownership climbs materially above ~5% if domestic advanced-node supply and Huawei Ascend deployments compound faster than US export-control scenarios assume.
4. Energised vs announced MW in the Middle East and India: if UAE/Saudi/India programs connect multi-GW IT loads on schedule, the US 45% capacity share compresses; if they slip, the US share sticks.
5. Inference relocation: sustained agentic demand that cannot clear US interconnection queues would show up first in European Nordics, Canadian hydro corridors, and Middle Eastern gas-adjacent campusesnot in ownership tables.

Until those print, the shareable frame stays narrow: five US hyperscalers own about seven-tenths of the world’s AI compute; Google alone owns about a quarter; the United States hosts close to half of AI data-center capacity by power; and global DC capacity is racing from ~104 GW to ~132 GW in a single year while AI servers take nearly a third of the power.

Caveats and methodology

• H100-equivalent ≠ identical utility. Epoch converts chips on peak 8-bit FLOP/s. Memory bandwidth, software stack, and networking can make a TPU H100e more or less useful than a B300 H100e depending on the workload. Training comparisons are more reliable than inference.
• Ownership coverage is incomplete. Epoch tracks Nvidia, AMD, Google TPU, Amazon Trainium/Inferentia, and Huawei. Meta and Microsoft custom chips are largely omitted and believed small relative to Google/Amazon customs.
• Big-5 71% and Google 25% are disclosed anchors; intra-Big-5 splits except Meta’s ~10% narrative are estimated so the leaderboard sums cleanly.
• China ~5% is an Epoch ownership narrative, not a claim about Chinese model quality or token volume.
• Regional capacity shares outside the US ~45% headline are residual estimates; hub MW mixes live, building, and announced IT loads and will disagree with any single utility filing.
• Gartner GW/TWh figures are forecasts (1Q26 vintage via public reporting); they are not LBNL US historicals and not Goldman capex dollars.
• Workload train/infer splits are industry synthesis, not a single regulator series.
• Frontier lab access figures (OpenAI ~1.7M H100e, etc.) come from Epoch’s lab analysis and inherit its conversion assumptions.

The shareable takeaway

As of Q4 2025, five US hyperscalers own ~71% of global AI compute (H100e), with Google alone near 25% — mostly TPUs. China as a whole owns about 5%. The United States hosts roughly 45% of AI data-center capacity by power draw, while Gartner’s global DC capacity jumps from 104 GW (2025) to 132 GW (2026) and AI-optimised servers take 31% of DC power. Frontier labs like OpenAI and Anthropic use vast rented pools; they do not top the ownership table. For site lists see the build tracker; for who burns tokens see the token series.