OpenAI Expects AI Compute Shortages Through 2027, Plans Nvidia Vera Rubin Training Run This Fall
OpenAI expects shortages of AI computing capacity to remain severe through 2027 as demand continues to outstrip supply, according to remarks attributed to Chief Financial Officer Sarah Friar. Friar said the company’s next major training run is expected this fall on Nvidia’s Vera Rubin platform, even as it works with AMD and Cerebras and develops a custom chip with Broadcom.
The comments add to OpenAI’s broader effort to reduce dependence on a single chip supplier as it hunts for scarce capacity. The company already has early access to Vera Rubin systems and is building software to run AI workloads across multiple chip providers; Friar also said GPT-5.5 is more efficient to serve than GPT-5.4, which she described as a much larger and costlier pre-training run.
From the sources (4 posts)
@wallstengineOpenAI is developing internal software to run AI workloads across chips from multiple providers, not just $NVDA, per The Information. Compute chief Sachin Katti described a future with heterogeneous AI infrastructure and hinted OpenAI coul
@theinformationOpenAI already has early access to Nvidia’s Vera Rubin systems, but it’s also building software designed for a future where AI runs across many chip providers. Full story:
@hangsiin@deredleritt3r @OpenAI @thefriley Yes, along with that, her explanation of GPT-5.5 and 5.4 also stood out to me. Even though the price increased with GPT-5.5, the way she described it made it sound as if 5.4 was the heavier model to serve
@hangsiinOpenAI(@OpenAI) CFO Sarah Friar(@thefriley) said the following during a conversation: -Computing is currently an extremely scarce resource, and OpenAI is seeing demand rise almost like a vertical wall. If you want to buy significantly more