Particle.news
Download on the App Store

OpenAI Broadens Inference Chip Sourcing as Reports Highlight Push for Lower Latency

OpenAI is seeking SRAM‑rich accelerators to speed responses for select customers.

Overview

  • Reuters reported OpenAI is dissatisfied with some Nvidia inference accelerators and has evaluated alternatives with large on‑chip SRAM to cut latency.
  • Sam Altman publicly praised Nvidia and denied a rift, responding to the report with a statement that OpenAI aims to remain a large customer.
  • The previously floated $100 billion investment proposal has not materialized, and Nvidia CEO Jensen Huang has said he does not plan to invest that amount.
  • OpenAI has lined up rival supply, including a reported five‑year AMD purchase totaling six gigawatts and a Cerebras Wafer Scale Engine agreement through 2028 for 750 megawatts.
  • Nvidia licensed Groq’s technology in December 2025 and hired much of its design team, and separate reporting says very low‑latency accelerators could account for about 10% of OpenAI’s inference fleet for premium users.