MEMORANDUM

Issue 03 · September 18, 2026
The story. The business stakes. Five minutes.

Who really controls AI’s computing power?

Google leads the estimates. By 2030, the biggest shift may be how much computing power each new chip delivers—and who keeps it.

01 — WHERE WE ARE NOW

The AI arms race has moved well beyond last year’s numbers.

In September, Oracle reported delivering more than 300,000 GPUs and 850 megawatts of additional capacity since May 31. CoreWeave ended June with 1.5 gigawatts of active power, up from roughly 850 megawatts at the end of 2025. These are physical expansion figures, although they measure different things. Oracle, CoreWeave

So who has the largest fleet now?

Epoch AI’s September model puts global capacity at roughly 32 million H100 equivalents. Its central estimates give Google a substantial lead:

Millions of H100 equivalents. Epoch model evaluated September 9, published September 16; ranges are its 5th–95th percentiles. SpaceX/xAI uses the notebook’s site-based adjustment. Chinese owners and other firms are grouped separately in the source, so this is not a complete company census. September research and model

These are deployment-adjusted estimates, not audited inventories. Epoch extrapolates its last complete chip-ownership data, from the end of 2025, and applies a median installation delay of roughly three months. We checked its estimates against 2026 company disclosures and data-center records; those disclosures do not reveal every company’s complete fleet. The overlapping ranges also make close rankings uncertain.

An H100 equivalent compares a chip’s advertised computing speed with Nvidia’s H100, using the same numerical precision. Think horsepower: useful for comparing capacity, but it does not measure how quickly a real application finishes. Google’s custom TPUs count too. This is not a count of Nvidia chips.

02 — WHAT CHANGES BY 2030?

We built three scenarios using that September starting point, current expansion disclosures, announced deliveries, and assumptions about how fast capacity grows.

Here is our central scenario, with companies ordered by their modeled 2030 capacity:

Millions of H100 equivalents. Memorandum scenarios, not company guidance or Epoch forecasts. Starting from our 42m end-2026 planning estimate, global capacity doubles in 2027; annual growth then slows to 80%, 60%, and 40%. The starting point assumes about 33% growth from September to December as more hardware becomes usable. That is our deployment judgment, not announced fleet capacity. Company allocations are also assumptions informed by the evidence below. Full assumptions, annual figures, and scenarios are documented in Memorandum’s research notes.

The slower and faster cases put global capacity at 140 million to 594 million in 2030. Google spans 34–133 million, Amazon 21–98 million, and Microsoft 23–84 million. These are scenario outcomes, not statistical confidence limits. Amazon passing Microsoft is a possibility in this model, not a settled prediction.

03 — WHY THE NUMBERS CAN GROW THAT MUCH

New chips change the arithmetic. Nvidia’s preliminary Rubin specification advertises roughly nine H100s’ worth of dense 8-bit computing throughput per GPU. That is a specification, not a measured ninefold application-speed gain. Still, future capacity can grow much faster than the physical chip count. Nvidia specifications

The delivery pipeline is becoming concrete. AWS announced two million additional Nvidia GPUs for 2027–2028. CoreWeave targets at least eight gigawatts of active power by 2030. Meta’s Hyperion campus is expected to begin operating in 2028 and eventually scale to five gigawatts. These support continued expansion; they do not guarantee our precise company totals. AWS, CoreWeave guidance, Meta

Microsoft is expanding too: it announced a roughly two-gigawatt campus in Pecos, Texas, with investment over five to seven years. That is a multiyear plan, not capacity all arriving by 2030. Our Amazon–Microsoft crossover reflects assumed shares of future growth, not a like-for-like comparison of contracted deliveries. Microsoft

Power is the limiting test. Buildings, grid connections, memory supply, and cooling must arrive alongside the chips. Our central case requires substantial improvement in computing power per watt. Delays or weaker gains push the outcome toward the slower case. The IEA expects total data-center electricity consumption to nearly double by 2030, with supply bottlenecks shaping the buildout. IEA’s 2026 assessment

04 — THE TWIST: WINNING THE CHIP RACE CAN CHANGE WHO OWNS THE FLEET

Google plans to deliver TPUs directly into some customers’ own data centers. Its chip business could grow while some of the resulting capacity sits outside Google’s ownership. “Google-designed” and “Google-owned” will become increasingly different measurements. Alphabet’s April earnings call

Rental creates another split. Anthropic can use SpaceX’s machines while SpaceX develops a competing AI business. OpenAI can have enormous computing access without appearing near the top of an ownership table. Count the owner and renter together and you count the same capacity twice. SpaceX’s disclosed arrangement, Epoch’s ownership and usage distinction

For investors, that creates different businesses: selling chips, owning infrastructure, renting capacity, and building products on somebody else’s fleet. A company can win one and lose another.

Our read: Google starts ahead in estimated ownership. The next four years turn on who gets newer chips running, secures enough electricity, and retains the right to decide what those machines do. The leaderboard matters. The business model behind it matters just as much.

SOURCES

Epoch: September model · ownership and use
Oracle results · CoreWeave results · 2030 guidance
AWS GPU deliveries · Microsoft Pecos · Meta infrastructure
Alphabet earnings · SpaceX prospectus
Nvidia specifications · IEA energy assessment

Research cutoff: September 17, 2026.

MEMORANDUM · Issue 03 · September 18, 2026
THE DAILY BRIEF FOR INVESTORS AND BUSINESS LEADERS