OpenAI Hires $92 Million Gimlet to Optimize Codex-Spark for Cerebras Chips

11 articles · Updated · The Information · May 11

OpenAI has hired San Francisco startup Gimlet Labs to tune its models for Cerebras chips, which are being used to run Codex-Spark, a faster version of its coding tool.
The move reflects a broader push to reduce reliance on hard-to-get Nvidia chips by making AI models work across multiple server architectures.
Founded in 2023, Gimlet has raised $92 million and sells software that can split a model across different hardware and translate Pytorch code into lower-level chip-specific code.
Gimlet says its tools can improve inference efficiency by 3 to 10 times at the same cost and power, a potential lever as OpenAI projects $655 billion in AI training and inference spending through 2030.
The deal underscores how a new layer of infrastructure startups is emerging to help major AI labs pursue cheaper multi-chip strategies while custom chip efforts remain years and financing-dependent.

As AI's hunger for chips and power grows, is the industry racing towards an unsustainable resource wall?

AI can now autonomously hack and self-replicate. Are we building our greatest defense tool or our most unstoppable digital plague?