Dell’s 757% Surge Puts Rack Value to the Test

In fiscal 2027 first-quarter results reported May 28, Dell Technologies posted $16.1 billion in AI-server revenue, up 757% year over year, while OpenAI, xAI and CoreWeave were arranging access through cloud rentals, GPU-linked financing and capacity agreements rather than simply buying servers. The apparent contradiction is the clue: every accelerator still needs cooling, networking, power and delivery to arrive together.

An accelerator without sufficient cooling cannot run as designed; a rack without networking cannot act as a cluster; and a server delivered before power is available becomes expensive furniture.

Cloud rentals, GPU-linked financing and capacity contracts put GPU access, power and cooling, networking, operations and delivery timing under connected obligations. For Dell, that creates a potential role as the rack-scale execution and supply-chain layer behind AI infrastructure. Whether the company earns more for that coordination remains unproved.

Microsoft and OpenAI also show why scarcity alone cannot explain these commitments. Semafor reported on March 21, 2025, that Microsoft declined to exercise a nearly $12 billion option for additional CoreWeave data-center capacity, which OpenAI then took up. OpenAI’s commitment therefore reflected a transferable commercial option as well as demand for available capacity.

Long contracts make availability the deliverable

OpenAI has reportedly planned to spend roughly $100 billion on backup servers rented from cloud providers through 2030, on top of $350 billion already projected for server rentals. OpenAI’s plan puts redundancy on the same long-dated procurement schedule as primary capacity, with roughly $100 billion reserved for backup access years in advance.

OpenAI also reportedly agreed to pay Cerebras more than $20 billion to use its server chips, potentially while receiving equity in the chipmaker. OpenAI’s agreement binds access, expenditure and strategic participation in one contract.

xAI’s planned capital raise approached $20 billion in equity and debt tied to Nvidia GPUs for Colossus 2. Nvidia was considering an investment of as much as $2 billion, while xAI intended to rent the chips. OpenAI’s cloud rentals, its Cerebras agreement and xAI’s GPU-linked financing use different structures, but each directs capital toward usable access, with equipment ownership secondary.

These deals fit the emerging contracted megawatt market without establishing a single procurement model. Buyers can substitute among ownership, rental and strategic investment, but every structure must still provide availability at the required cost and schedule.

A kilowatt chip makes the rack the design unit

Nvidia explains the expanding commercial boundary through its product architecture. Its GB200 Blackwell AI server packs 72 chips into one system and, for certain mixture-of-experts models, can deliver ten times the performance of H200 servers. Nvidia later unveiled a liquid-cooled rack containing 256 Vera CPUs, each with 88 custom cores and memory bandwidth of up to 1.2 terabytes per second.

As high-performance chips such as Nvidia’s GH200 exceeded one kilowatt in 2023, manufacturers began adopting direct liquid cooling. Cooling hardware entered the deployment’s critical path because the compute system could not operate at its designed performance without it.

Operators must design accelerators, interconnects, memory, coolant distribution and facility power against one another. An operator running a homogeneous deployment can optimize across those boundaries, while a buyer assembling each subsystem through separate purchasing cycles inherits every interface failure. The facility becomes the useful computing unit because each local optimization can idle something more expensive beside it.

Nvidia can define the architecture, but a cloud operator still needs a supplier to source, integrate, test and install the physical system. Silicon may produce benchmark scores, but plumbing and power determine whether the rack stays up. AI infrastructure has become unusually effective at reminding software companies that water has opinions.

Dell now competes on rack-scale execution

CoreWeave supplied the clearest operating example when it became the first cloud provider to install Nvidia Blackwell Ultra systems using liquid-cooled servers supplied by Dell. CoreWeave provided the cloud-access layer, Nvidia provided the accelerator platform and Dell supplied the installed rack-scale system.

Dell was also reportedly nearing a $5 billion agreement to provide Nvidia GB200 servers for xAI’s Memphis supercomputer project. The same report said Hewlett Packard Enterprise had won a reported $1 billion AI-server deal for X in late 2024 after Dell and Super Micro Computer also bid. Large AI projects still create consequential hardware contests in which each bidder must secure scarce components and execute a complex deployment.

Dell and HPE must translate an accelerator roadmap into coolant manifolds, network topology, component orders and installation schedules. Buyers may contract for finished compute, but OEMs still have to make the physical system arrive and operate as promised.

Lenders now allocate capacity alongside chipmakers

AI infrastructure companies need AI infrastructure finance because they pay for components and facilities before customers generate utilization. These companies borrowed more than $100 billion in 2025, while smaller borrowers paid higher interest rates as investors questioned unproven AI businesses.

Reuters reported on June 10, 2026, that Morgan Stanley expected global AI-linked debt issuance to more than double to nearly $570 billion that year as hyperscalers sought alternatives for funding capital expenditure. The forecast belongs inside the broader AI infrastructure commitment stack: builders need contracts, collateral and capital to fund the physical system behind a capacity promise.

CoreWeave had already demonstrated the link between chips and credit. Reuters reported on August 3, 2023, that the company raised a $2.3 billion debt facility collateralized by Nvidia chips after raising $421 million in equity that year. CoreWeave used the hardware it needed to deploy as security for financing more capacity.

General Compute pushed the mechanism one layer further when it received a $400 million loan described as the first transaction to use inference-specific chips as collateral. Upper90 attached credit to the hardware expected to produce compute revenue.

Lenders participate in capacity allocation alongside chipmakers and cloud operators. A well-financed builder can reserve components and carry construction costs before utilization begins. By raising borrowing costs or withholding credit, lenders can delay which builders secure hardware and deliver capacity.

Dell’s margins will reveal its rack-scale leverage

In the same fiscal first-quarter results, Dell raised its fiscal 2027 AI-server revenue forecast to $60 billion from $50 billion. The increase establishes strong demand for Dell’s systems, but revenue alone does not show how much economic value the company retains from integration.

The cited results did not disclose a separate gross-margin figure for AI servers or a comparable margin for box-only sales. Order conversion, backlog and margins must therefore test the thesis: if rack-scale coordination creates durable leverage, Dell should convert demand without surrendering the economics to component suppliers, customers or financing costs.

Dell’s $16.1 billion quarter resolves only half the opening contradiction. Buyers can contract for usable capacity and still drive OEM revenue because someone must source, cool, connect and install the racks. Dell’s margins must now show whether the obligation around the box has become a profitable product.