Particle.news

Download on the App Store

OpenAI Taps Broadcom for 10 GW of Custom AI Chips, With Deployments Starting in 2026

The pact underscores OpenAI's bid to control performance, cost, supply by embedding model lessons into its own accelerators.

Overview

  • Under the multi‑year deal, OpenAI designs the accelerators and systems and Broadcom builds, integrates, and deploys full racks using its Ethernet‑based networking.
  • Installations begin in the second half of 2026 and continue through 2029 across OpenAI facilities and partner data centers.
  • Broadcom’s shares jumped, adding roughly $150 billion in market value following the announcement.
  • The 10‑gigawatt scale highlights a vast power footprint, with reports noting it is comparable to the electricity usage of more than 8 million U.S. households.
  • The agreement extends a rapid series of capacity moves that include a 6‑gigawatt AMD supply with warrants that could give OpenAI up to about a 10% stake, a Nvidia letter of intent tied to 10 gigawatts and up to $100 billion, and a large Oracle cloud commitment that multiple outlets have reported at roughly $300 billion, drawing scrutiny over financing, concentration, and execution risks.