Who Buys AI GPUs? The Hyperscalers Nvidia Gets Paid By First
AI GPU buyers are the companies that pay for accelerators. Nvidia split $48.7B hyperscale and $40.3B ACIE in the July quarter. Who pays, and where cash sticks.
EXPLAINED · Players · Last reviewed: 19 Oct 2026 · Next review: after NVIDIA's fiscal third-quarter 2027 results
An AI GPU buyer is the company that pays for accelerators used to train and run models. Nvidia splits those buyers into two groups: hyperscale — public clouds and the largest consumer-internet companies — and ACIE, short for AI clouds, industrial and enterprise. In the quarter to 26 July 2026, hyperscale accounted for $48.7 billion of Nvidia's $89.0 billion of data-center revenue.
In short. Who buys AI GPUs is a roles map on the hyperscalers layer, not a ranking. Nvidia's books split data-center revenue in two: hyperscale paid $48.7 billion in the quarter to 26 July 2026, 54.7% of the $89.0 billion total (our arithmetic). ACIE — AI clouds, industrial and enterprise — paid $40.3 billion. One unnamed direct customer was 16% of Nvidia's $96.2 billion of company revenue, about $15.4 billion (our arithmetic). Nvidia gets paid when the chips ship. The buyer earns only once those chips run in a powered hall. That hall is the stuck step.
The same GPU sits on two cash clocks. Nvidia books the sale when a rack ships. Microsoft, Amazon, Google, Meta and Oracle spend first as AI capex, then earn only when cloud hours or ads bill. This page maps who pays Nvidia. It sits beside what a hyperscaler is and who makes AI chips.
Who buys AI GPUs?
Two groups, in Nvidia's own language. Hyperscale is revenue from public clouds and the world's largest consumer-internet companies. ACIE is everyone else: AI clouds (the specialist GPU clouds this site calls neoclouds), industrial sites, enterprises and sovereign projects.
Jensen Huang, Nvidia's chief executive, said on the 20 May 2026 call that the easiest go-to-market is the hyperscaler "because they are only… 5 or 6 of them." The rest of the industry is "250 thousand companies around the world." Nvidia does not publish a named list. Second Order's hyperscalers layer maps five public buyers: Amazon, Microsoft, Google, Meta and Oracle. Meta is in the group because its ads-funded halls buy at hyperscale size — see what Meta does. The same scarce finish lines sit behind every rack: TSMC packaging and high-bandwidth memory. Who makes AI chips covers those roles.
| Role | What it does | Who does it |
|---|---|---|
| Hyperscale buyer | Pays for GPUs, halls and power; earns on cloud, ads or both, only after a hall is live | Public clouds and the largest consumer-internet companies. Second Order maps Amazon, Microsoft, Google, Meta, Oracle |
| ACIE buyer | Pays for GPUs for specialist clouds, enterprises, industry and sovereign projects | AI clouds / neoclouds, industrial and enterprise customers. Nvidia named CoreWeave and Nebius among Vera Rubin partners |
| Indirect buyer | Pays a cloud company for GPU hours; Nvidia still counts some of that as its revenue concentration | Model labs and other customers that buy through Nvidia's direct customers. Nvidia said one AI research and deployment company was "meaningful" in the July quarter |
| Seller | Designs the GPU, books revenue when racks ship | Nvidia, with AMD as the other volume GPU designer hyperscalers also buy |
Roles only, not a ranking. Sources under Sources.
Who are the hyperscalers in Nvidia's numbers?
Nvidia's 20 May 2026 results release defined hyperscale as "revenue from the public clouds and the world’s largest consumer internet companies." It does not name them in the revenue table. Second Order tracks five public buyers on the hyperscalers layer: Microsoft and Amazon convert cloud backlog into hours; Google sells GPU and TPU hours and also designs its own TPU (who makes custom AI chips); Meta buys at hyperscale size for ads and models; Oracle delivers AI-cloud contracts and gets paid as capacity is accepted (remaining performance obligation).
Nvidia's 26 August 2026 release named where Vera Rubin racks were already running: CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius. Two of those five are hyperscalers on this map; two are AI clouds, which Nvidia books in ACIE.
| Buyer | How it earns after the GPU ships | Dated capex / GPU print |
|---|---|---|
| Microsoft | Azure and Copilot hours, once halls are live | Q4 FY2026 capex $41B, ~two-thirds CPUs and GPUs; CY2026 ~$175B after lease reclass (29 Jul 2026) |
| Amazon | AWS hours, once halls are live | Company-wide 2026 capex guide ~$200B (5 Feb 2026); TTM net PPE $169.0B at 30 Jun 2026 |
| Google Cloud hours, including TPU and GPU | Q2 2026 capex $44.9B; FY2026 guide $195–205B (22 Jul 2026) | |
| Meta | Ads and some sold-on capacity, once halls are live | Q2 2026 capex $31.08B; FY2026 $130–145B (29 Jul 2026) |
| Oracle | OCI AI cloud, as customers accept capacity | RPO $664B; >300,000 GPUs delivered since end-Q4 FY2026; Q1 FY2027 capex $28B (10 Sep 2026) |
Company prints, not a share table. Nvidia does not name which of these is the 16% direct customer. Sources under Sources.
How much of Nvidia's revenue do they account for?
Use Nvidia's split, not a guessed list.
In the quarter to 26 July 2026, data-center revenue was $89.023 billion of $96.221 billion total, 92.5% (our arithmetic). Hyperscale was $48.710 billion, 54.7% of data center; ACIE was $40.313 billion, 45.3% (our arithmetic). Hyperscale rose 13% from the prior quarter; ACIE rose 25%. For the first half of fiscal 2027, hyperscale was $91.761 billion of $164.269 billion of data-center revenue, 55.8% (our arithmetic). The "hyperscalers buy all the GPUs" story is already wrong on Nvidia's own page.
Concentration is still real. One unnamed direct customer was 16% of total revenue in the July quarter, primarily in Compute & Networking — about $15.4 billion (our arithmetic). For the first half, three unnamed direct customers were 16%, 15% and 13% of total revenue, 44% combined, about $78.2 billion (our arithmetic). Five direct customers held 22%, 14%, 13%, 11% and 10% of accounts receivable at 26 July 2026.
Nvidia also warns that some indirect customers — companies that buy cloud from Nvidia's direct customers — individually represent 10% or more of revenue. In the July quarter it estimated that "one AI research and deployment company contributed a meaningful amount of our revenue by purchasing cloud services from our customers." A model lab can be a large GPU buyer without appearing on Nvidia's invoice. And the split itself can move: Nvidia reclassified a company from ACIE into hyperscale in the July quarter "due to a change in their business model" and recast the prior period.
Who else buys — the other half?
ACIE is the other $40.3 billion — "diverse AI purpose-built data centers and AI factories across industries and countries," Nvidia said on 20 May 2026. July-quarter growth, the 10-Q said, came from "end-demand from AI natives, enterprises, and sovereign customers, as well as hyperscalers utilizing AI clouds."
That last clause matters. A hyperscaler that cannot stand a hall up fast enough still buys GPU hours from a neocloud. The GPU is the same. The invoice is not. Nvidia can book that sale in ACIE even though the end user is a hyperscaler. CoreWeave and Nebius, named as Vera Rubin partners on 26 August 2026, are that overflow door. GPU cloud providers maps the fleets; neocloud vs hyperscaler is the versus; who finances GPU clusters covers how they are paid for. Sovereign projects sit here too: a government rarely buys the GPU itself. Sovereign AI is the mechanism.
Who gets paid — and where is the stuck step?
Nvidia gets paid first, at shipment. Data-center revenue of $89.0 billion in the July quarter is that invoice. Cash can lag: the 10-Q says investment-grade buyers can take 90 days and up to a year on large builds, and five customers held 22%, 14%, 13%, 11% and 10% of receivables at 26 July 2026 — 70% together (our arithmetic).
The buyer gets paid later. Microsoft, Amazon and Google earn when cloud hours bill. Meta earns when ads run. Oracle earns when an AI-cloud customer accepts the hall. Between Nvidia's invoice and that first dollar sit packaging, racking and a powered hall. Cash capex is the cash that leaves before those hours bill.
The stuck step is still the hall. Nvidia can ship. The buyer cannot earn until the rack is plugged into megawatts that have cleared the grid. The hyperscalers stack has that bind as gas and grid reaching the hall. A shipped GPU in an unpowered aisle is cash Nvidia has booked and the buyer has not yet earned. Packaging and HBM still gate how many GPUs can ship at all — who makes advanced packaging, who makes HBM. If packaging catches up, where the stuck step moves is the next bind: through memory, to power.
What to check next
Nvidia's fiscal third-quarter 2027 results, for the hyperscale versus ACIE split and the 16% customer print. Late-October 2026 earnings from Microsoft, Amazon, Alphabet and Meta, for capex and hall-delay language. Oracle's next remaining-performance-obligation print against $664 billion. The money side sits on the hyperscalers stack.
Common misconceptions
- People assume hyperscalers buy all the AI GPUs. Actually Nvidia's July-quarter split put hyperscale at 54.7% of data-center revenue and ACIE at 45.3% (our arithmetic).
- People assume Nvidia names Microsoft, Amazon and Google as the 16% customer. Actually it does not name the customer. Treat names on this page as a roles map, not as Nvidia's list.
- People assume a model lab that buys GPU hours is absent from Nvidia's books. Actually Nvidia said some indirect customers individually represent 10% or more of revenue, and that one AI research and deployment company was "meaningful" in the July quarter.
- People assume a hyperscaler only buys Nvidia. Actually they also buy AMD GPUs and their own custom chips. Google's TPU, Amazon's Trainium and Microsoft's Maia sit in that second path.
- People assume Nvidia's sale means the buyer is already earning. Actually Nvidia is paid at shipment. The buyer earns only once the chips run in a powered hall.
FAQ
Who buys AI GPUs?
Two groups, in Nvidia's own split. Hyperscale buyers are public clouds and the largest consumer-internet companies. ACIE buyers are AI clouds, industrial and enterprise customers. In the quarter to 26 July 2026, hyperscale paid $48.7 billion of Nvidia's $89.0 billion of data-center revenue; ACIE paid $40.3 billion.
Who are the hyperscalers that buy AI GPUs?
Nvidia does not name them in the revenue table. It defines hyperscale as public clouds and the world's largest consumer-internet companies, and Jensen Huang has said there are five or six. Second Order maps Amazon, Microsoft, Google, Meta and Oracle.
How much of Nvidia's revenue comes from hyperscalers?
$48.7 billion in the quarter to 26 July 2026, 54.7% of $89.0 billion of data-center revenue (our arithmetic). For the first half of fiscal 2027, hyperscale was 55.8% of data-center revenue (our arithmetic).
Does one customer dominate Nvidia's GPU sales?
One unnamed direct customer was 16% of Nvidia's $96.2 billion of total revenue in the July quarter, about $15.4 billion (our arithmetic). For the first half, three unnamed direct customers were 16%, 15% and 13%. Some indirect customers also individually represent 10% or more.
Do hyperscalers only buy Nvidia GPUs?
No. They buy Nvidia in volume, and they also buy AMD GPUs and their own custom chips. Google's TPU, Amazon's Trainium and Microsoft's Maia are that second path.
Who else buys AI GPUs besides hyperscalers?
AI clouds, enterprises, industrial sites and sovereign projects — Nvidia's ACIE group, $40.3 billion in the July quarter. Nvidia named CoreWeave and Nebius among Vera Rubin partners on 26 August 2026.
Is this a list of stocks to buy or sell?
No. Second Order Players pages map who does what and how the money moves. They do not rank companies or give buy or sell calls.
Related terms
Hyperscaler; AI capex; neocloud vs hyperscaler; GPU cloud providers; who makes AI chips; who makes custom AI chips; remaining performance obligation.
Markets and stack doors
Layer: Hyperscalers. Also: Neoclouds, Semiconductors, AI Data Centers. Company models: Amazon, Microsoft, Google, Meta, Oracle. Markets: US, Taiwan.
Get the Wire — daily notes when hyperscaler capex, GPU shipments and hall clocks move.
How we check this
Nvidia's revenue split and customer-concentration figures come from its Form 10-Q for the quarter ended 26 July 2026 and its 26 August 2026 results release. Jensen Huang's "5 or 6" line comes from the 20 May 2026 call. Microsoft, Amazon, Alphabet, Meta and Oracle figures come from those companies' own results, dated below. Our percentages from Nvidia's totals are labelled "our arithmetic". This page maps who pays. Not a buy list.
Last reviewed: 19 Oct 2026 · Next review: after NVIDIA's fiscal third-quarter 2027 results
Spot an error? Tell us and we will correct it and note the change here.
Sources
- NVIDIA — Form 10-Q for the quarter ended 26 July 2026 — hyperscale $48.710B, ACIE $40.313B; 16% direct customer; first-half 16%/15%/13%; receivables mix; indirect-customer language; ACIE-to-hyperscale reclass
- NVIDIA — Second quarter fiscal 2027 results (26 Aug 2026) — revenue $96.2B; data-center $89.0B; Vera Rubin at CoreWeave, Google Cloud, Microsoft Azure, Oracle, Nebius
- NVIDIA — First quarter fiscal 2027 results (20 May 2026) — hyperscale and ACIE definitions
- NVIDIA — Fiscal Q1 2027 earnings call transcript (Motley Fool; call held 20 May 2026) — Huang: hyperscalers are "only… 5 or 6"; rest of industry "250 thousand companies"
- Microsoft — Fiscal year 2026 fourth-quarter earnings call (29 Jul 2026) — Q4 capex $41B, ~two-thirds CPUs and GPUs; CY2026 ~$175B; Q1 FY2027 capex over $50B
- Amazon — Fourth quarter 2025 results (5 Feb 2026) — 2026 company-wide capex expected about $200B
- Amazon — Second quarter 2026 results (30 Jul 2026) — TTM net PPE $169.0B; TTM free cash flow outflow $7.6B
- Alphabet — Second quarter 2026 earnings call (22 Jul 2026) — Q2 capex $44.9B; 2026 guide $195–205B
- Meta — Second quarter 2026 results (29 Jul 2026) — Q2 capex $31.08B; 2026 outlook $130–145B
- Oracle — Fiscal Q1 2027 results (10 Sep 2026) — RPO $664B; more than 300,000 GPUs delivered since end-Q4
- Oracle — Fiscal Q1 2027 earnings call transcript (Motley Fool; call held 10 Sep 2026) — Q1 capex $28B; net cash capex $18B
S
Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors. Focus: who gets paid, where money sticks, and which step is stuck.
