Microsoft exits Nscale UK data center over 2030s power

The Microsoft-Nscale cancellation is the concrete trigger for a broader point about how AI infrastructure money is structured. The facility was meant to provide computing power for Azure, but the electricity to run it would not be available until the 2030s — a timing gap long enough for a major customer to walk away.

That distinction runs through the rest of the piece. Nvidia sells processors and networking equipment and generally recognizes revenue when the equipment is delivered rather than waiting years for the customer to earn a return. Broadcom designs custom AI accelerators and supplies networking gear, getting paid for the semiconductors and networking products it supplies. AMD has an agreement with OpenAI for deploying 6 gigawatts of AMD GPUs over multiple years, beginning with the MI450 series, and also granted OpenAI warrants to acquire up to 160 million AMD shares subject to purchasing milestones and other conditions.

On the other side, Oracle builds and operates data centers and rents computing capacity, and if a customer cannot meet its obligations it could be left with infrastructure built for a particular workload. CoreWeave operates GPU-powered cloud facilities, and Nvidia supplies the chips, owns a stake, and has agreed to purchase unused computing capacity under certain conditions through 2032, yet CoreWeave still has to finance infrastructure, service debt and deliver capacity on schedule. Nebius has long-term contracts but must pay for its infrastructure even if a project is delayed or a customer scales back. Alphabet sits on both sides through Google Cloud.