Allocators and lenders sizing the AI capital-spending chain
The whole rack
2026-11-05
For two years the story of AI hardware was told in graphics processors. That framing is now out of date. The April-quarter results from the market leader, reported in May, show a company whose fastest-growing lines are the parts around the processor: the switches and cables that join thousands of chips into one machine, and from this year a general-purpose processor of its own. The unit of sale is the rack. The unit of account, for anyone lending against a data centre, should be the same.
What the quarter showed
- Revenue of 81.6 billion dollars, up 85% on the year and the fourteenth consecutive quarter of sequential growth. Data-centre sales were 75.2 billion of that.
- Networking was the standout at 14.8 billion dollars, roughly tripling on the year. It is now about a fifth of data-centre revenue, up from an eighth a year earlier, and has grown faster than the compute line in four of the last five quarters.
- The customer base has split in two. Hyperscale cloud providers were about half of data-centre revenue. The other half came from AI clouds, industry and enterprise, with the AI-cloud piece more than tripling and sovereign deployments, now in nearly forty countries, up by four fifths.
- China contributed nothing to data-centre compute, against 4.6 billion dollars in the same quarter a year earlier, before export restrictions. The outlook assumes it stays at zero.
- The company guided the July quarter to 91 billion dollars with gross margin around 75%, and said it expects sequential growth to continue through this year and into next.
- Cash generation reached a record 48.6 billion dollars in the quarter, and the board added 80 billion to the buyback and raised the quarterly dividend from one cent to twenty-five.
The processor nobody priced
The detail that matters most for the next two years is the processor. The company disclosed visibility to about 20 billion dollars of standalone revenue this calendar year from its own CPU, sold at board level alongside the accelerators. At plausible selling prices that is several million processors, or a mid-single-digit share of the world's server CPU market, taken in a single year by a company that was not in that market before.
The point is not the share. It is the attach. Each generation of rack carries more of the company's own content: the accelerator, the network, now the host processor, and the software that ties them together. A customer who buys the rack buys all of it. The suppliers who used to sell those parts separately are now selling into a bundle they do not control.
Why this matters to the financiers
Most of the capital going into AI infrastructure is not equity in the chip company. It is debt and structured finance against the buildings, the power and the equipment. For that money, three things follow from the quarter.
- Residual value is a rack question, not a chip question. A lender who models the accelerator alone, with its two-to-three-year useful life, misses that the network and the host processor on the same invoice have different lives and different secondary markets.
- Concentration risk has moved. The equipment in a financed data centre now comes overwhelmingly from one vendor, with one supply chain behind it. Advanced wafers and high-bandwidth memory are the known bottlenecks, and the company has said its outlook is limited by supply rather than demand.
- The customer mix is broadening below the hyperscalers. AI clouds, sovereigns and enterprises are half the market, and they are the borrowers in most of the financing we see. Their credit is not the hyperscalers' credit.
Our read
The company has turned a component into a system, and the system into the unit that gets financed. That is good for the vendor and awkward for everyone pricing the risk around it, because the familiar comparables, server lease books and chip resale values, no longer describe the asset.
We spend our time on the structures behind this build-out, on how the equipment, the power and the tenancy are carved up between lenders and owners, and on access to the private companies in the chain. The lesson from this quarter is to read the invoice, not the ticker. Whatever the headline number does, the rack is the thing that gets delivered, financed and eventually written down.