HWNEWS
Three late 2025 deals promise more than 10 gigawatts of AI compute. The memory, optics, cooling, and power chain that has to build it does not yet exist at that scale.
A 10-gigawatt AI promise is not a 10-gigawatt AI factory. Nvidia and OpenAI have announced a strategic partnership to deploy 10 GW of Nvidia systems (Nvidia newsroom, OpenAI blog). OpenAI has separate definitive agreements with AMD for 6 GW of GPUs and with Broadcom for 10 GW of OpenAI-designed accelerators and networking. None of those gigawatts is the same document. A contract is not a rack, a rack is not a cluster, and a cluster is not yet an AI factory: the operational unit of memory, optical networking, cooling, and power that has to come up around the chip.
That is the gap the partnerships signed since late 2025 have opened. Nvidia says it will invest as much as $100 billion in OpenAI as the systems deploy (Nvidia newsroom). Broadcom disclosed in a filing that it can lend Anthropic up to $42 billion to help finance computing infrastructure (Reuters). AMD has issued warrants to major customers whose vesting is tied partly to gigawatt-scale GPU purchases. The dollar stack is real. The unit the public is asked to compare across deals is not.
The unit that translates the headline into physical hardware is the reference AI factory. EE Times uses Nvidia's published Vera Rubin reference design, the chip vendor's next-generation GPU platform, as a yardstick: 100 MW of AI factory capacity corresponds to roughly 40,000 Rubin-class GPUs and 12 petabytes of HBM4, the next generation of high-bandwidth memory stacked on the same package as the processor (EE Times). That is a single reference block, not a universal conversion factor, but the author walks the linear scaling: a gigawatt implies something on the order of 400,000 GPUs and 120 PB of HBM4. Three announced 10-GW deals, taken at face value, would point to roughly 12 million GPUs and 3.6 exabytes of stacked memory before counting anything else.
The chip is the part the wire is built to cover. The systems around the chip are what decide which announcements become operable capacity. High-bandwidth memory, or HBM, is the closest pinch point. Samsung has guided that HBM will account for nearly 30% of industry DRAM wafer capacity in 2027, up from about 20% today, which means the next memory cycle is being sized for AI before it is sized for the commodity dynamic random-access memory (DRAM) that feeds servers, graphics cards, and consumer devices (EE Times). A 30% HBM share in 2027 is a vendor forecast, not a confirmed allocation, but the order of magnitude is the story: the wafer line that builds HBM4 for Vera Rubin is the same wafer line that would otherwise build the standard DRAM in a phone or a PC, and that trade is being made ahead of the deals being stood up.
Then networking. A 10-GW build at the reference density means hundreds of thousands of GPUs that have to talk to each other at training latency, which translates into hundreds of thousands of optical transceivers and switch ports, plus the fiber plant that carries the signal. Liquid cooling is the next layer. Air cooling stops being economic somewhere around 60 to 80 kW per rack, and a 400,000-GPU reference site is by definition a high-density site. The supply chain for direct-to-chip cold plates, coolant distribution units, and the heat-rejection loop that ties them to a utility is real and is being expanded, but the expansion is being measured in years.
Power is the binding constraint at the end. A 10-GW AI factory is not a 10-GW power purchase agreement; it is a 10-GW substation, with interconnect queues, transformer lead times, and behind-the-meter generation that the regional grid operator has to accept. Substation transformers are still on multi-year backlogs in most U.S. ISO footprints. The same gigawatt that fits on a press release has to clear the queue.
The partnership announcements can still be read as honest bets on demand. Spyglass has raised structural concerns about the Nvidia-OpenAI deal, including the circularity of the $100B commitment and the timing of equity versus GPU deliveries (Spyglass). That critique is a real counterweight to the partnership language, and it is one a reader should keep in view. It does not change the systems problem. Even a deal whose structure a skeptic finds uncomfortable still has to stand up the same memory, the same optics, the same coolant, and the same substation, or it does not produce AI factory capacity at all.
The next test is timing, not magnitude. Nvidia has tied parts of the 10-GW deployment schedule to Vera Rubin generation cycles. Samsung's 2027 HBM wafer share guidance gives a public calendar for when stacked-memory supply has to clear. OpenAI's separate Broadcom collaboration targets 10 GW of OpenAI-designed accelerators and networking, not Nvidia parts, so the OpenAI build is in effect two parallel supply chains at once. Each of those clocks is now a public commitment. The question for the next four quarters is not whether the partnerships are real. They are signed. The question is how many of the named supply pieces arrive on the same schedule, and which one misses first.