Servers and systems

The pass-through business

An AI server is mostly other companies’ silicon. The accelerators, the high-bandwidth memory beside them and the networking chips account for most of the bill of materials, and they are priced by suppliers with far more pricing power than the companies assembling them. So when GPU prices rise, server revenue rises too, but the assembler’s profit grows only with the work it adds: integration, liquid cooling, testing, installation and service.

The numbers show it plainly. Over two years the companies in this layer grew revenue +56% to $297B, but their combined gross margin went from 17% to 16%. Of the $93.9B they added in revenue, $12.8B stayed as gross profit. The rest went straight to the chip and memory makers.

This isn’t a criticism of the companies. Integration at the scale of a 72-GPU, liquid-cooled rack is hard, and customers pay for speed and reliability. But it means the layer’s economics are about volume, working capital and execution, not about owning a scarce input.

Dell: the scale player

Dell reports its AI hardware inside its Infrastructure Solutions Group, which brought in $94.5B of revenue over its latest four quarters, +90% on the year before. Within it, AI-optimized servers and networking were $16.4B in the latest quarter alone, against $24.7B for the whole fiscal year to January 30, 2026.

What makes Dell interesting is that the segment’s operating income has grown faster than its revenue: $12.5B over the latest four quarters, +107%, an operating margin of 13%. Scale gives Dell purchasing power, a services business attached to each sale, and enterprise customers who buy storage and traditional servers alongside AI racks. It is the best evidence in this layer that volume can turn into profit.

Dell also reports $132B of remaining performance obligations, much of it AI orders not yet shipped. Backlog in this layer is a mixed signal: it guarantees revenue, but it also means inventory and receivables that must be financed until the racks are delivered and paid for.

Supermicro, HPE and the contract manufacturers

Supermicro grew revenue +78% to $39.1B over its latest four quarters at a gross margin of 11%, and its operating cash flow was −$6.8B. Negative cash flow while revenue grows that fast is the working-capital problem in its purest form: the company buys GPUs and components ahead of delivery and waits to be paid, so growth consumes cash.

HPE’s server segment brought in $20.7B over its latest four quarters, +14%, slower than Dell. HPE’s AI story leans increasingly on networking after its acquisition of Juniper, and it has given Oracle a warrant tied to deploying HPE networking in Oracle’s AI data centers, a sign of how hard suppliers in this layer compete for the largest accounts.

Celestica, Jabil and Flex build switches, servers and racks to the hyperscalers’ own designs. Celestica grew +62%, Jabil +21% and Flex +12%, at gross margins between 9% and 11%. Their advantage is that hyperscalers increasingly design their own hardware and need someone to build it; their limitation is that the customer owns the design and can move the work.

How to think about the layer

For an investor, three numbers decide whether a server company is creating value: gross margin (is it adding more than it passes through?), operating cash flow (is growth funding itself or consuming cash?) and customer concentration (could one hyperscaler’s design change take away a large share of revenue?).

The layer tends to do well when AI spending accelerates, because revenue scales with the price of what it ships. It is also the first to feel a slowdown, because inventory bought for orders that shrink becomes a write-down, and its thin margins leave little cushion. Treat its revenue growth as a reading of the build-out’s pace, and its margins and cash flow as the test of the business.

The companies, side by side

Latest four reported quarters for US filers; latest fiscal year for 20-F filers (marked). Underlined names have their own company page; the others link to the company’s filings on SEC EDGAR. Margins are left blank where a company doesn’t report gross profit or cost of revenue.

Where the money comes from and goes

This layer is paid by AI clouds. Its own spending flows to chips, memory, networking. For the dollar amounts and who keeps them, see following the money.

Financing links that involve this layer

Quoted from the filings. Numbers match the money-loop map.

• 21HPE → OracleCustomer warrant

  Warrant for up to 4.2M HPE shares, tied to networking in Oracle’s AI data centers

  “The Warrant was issued in connection with the Holder entering into a strategic collaboration with the Company to deploy the Company’s networking solutions in the Holder’s AI data centers.”

  HPE 10-Q for the period ended July 31, 2026, filed September 3, 2026

What to watch in the next filings

• Gross margin against revenue growth: a margin that holds or rises as AI revenue grows is the sign of real value added.
• Operating cash flow and inventory. Rising inventory with falling cash flow is the warning sign in this layer.
• Dell’s AI-optimized server revenue and backlog, the most direct read of enterprise and neocloud demand.
• Customer concentration disclosures (any customer over 10% of revenue) in each 10-K.
• New accelerator generations. Transitions can delay orders and leave older inventory to clear.