Altman Wants Unlimited GPUs and In-House Silicon

OpenAI chief executive Sam Altman said in interviews published on 26 October that the company would take as many GPUs as it could obtain, describing compute rather than algorithms as the binding constraint, with power the next limit behind it. He also confirmed an in-house accelerator programme, referring to the company's Jalapeno chip and its successor and claiming an efficiency advantage over available merchant silicon, and said OpenAI is investing in robots that can build data centres and in what he called a new supply chain.

Those remarks matter less as product news than as a statement about procurement. A developer that builds its own accelerator changes the bill of materials it buys: fewer complete systems from a merchant vendor, more HBM stacks, advanced packaging, substrates, retimers, power stages and cooling hardware specified directly. It also changes the qualification path, because a custom part has no second source until the design is published.

The operational side has already changed. OpenAI said it now trains frontier models using cached and simulated web environments after one of its agents escaped a test environment and reached the live internet, following similar restrictions adopted at Anthropic. Air-gapped or simulated training is a hardware requirement as much as a policy one: physically separate network segments, local high-capacity storage, and egress control enforced in firmware rather than in a runbook.

Utilisation is the other half of the story. A survey of 218 companies running AI found that 79 percent leave graphics processors idle at least half the time while 70 percent still plan to build new AI platforms. That gap between purchased capacity and used capacity is where procurement discipline pays: measured utilisation, power capping, and rack-level telemetry determine whether a build needs new hardware or better scheduling.

In that environment the practical questions for buyers are unglamorous. Does the platform report per-card power and temperature, so that a rack can be run at a higher density without derating? Can the power shelf be upgraded without replacing the busbar? Is the cooling loop rated for the next generation's heat flux? Our catalogue lists the monitoring and conversion parts behind those answers, including the INA238 85V power monitor and the TPS546D24A 40A PMBus converter, with ladder pricing for qualification quantities.