OpenAI Ditches Nvidia Chips Over Speed Concerns, Signs Ten Billion Dollar Deal with Cerebras

The Chip Revolt at OpenAI

OpenAI has been exploring alternatives to Nvidia chips since last year, driven by dissatisfaction with inference speeds for specific applications, according to a report citing eight sources familiar with the matter. The development marks a significant shift in the AI hardware landscape as the industry's focus moves from model training to real-time inference performance.

The ChatGPT developer has held discussions with startups including Cerebras and Groq about chips featuring on-chip SRAM memory, which offers faster data retrieval than the external memory used in Nvidia's GPUs. OpenAI is seeking hardware that could eventually account for about ten percent of its inference computing needs.

Speed Problems with Code Generation

Inside OpenAI, the speed issue became particularly visible with Codex, the company's code-generation tool. Staff attributed some of Codex's limitations to Nvidia's GPU-based hardware. During a recent call with reporters, CEO Sam Altman acknowledged that customers using OpenAI's coding models put a big premium on speed for coding work.

OpenAI has responded by signing a ten billion dollar deal with Cerebras. The agreement will deploy seven hundred and fifty megawatts of Cerebras wafer-scale systems through twenty twenty-eight, making it the largest high-speed AI inference deployment in the world. Benchmarks show Cerebras achieving output speeds of over three thousand one hundred tokens per second, dwarfing rivals that often hover below five hundred.

Nvidia Strikes Back

As OpenAI signalled concerns about Nvidia's technology, the chipmaker moved to shore up its position. In December, Nvidia struck a twenty billion dollar licensing deal with Groq, acquiring key technology and hiring away chip designers including founder Jonathan Ross. The deal effectively halted OpenAI's discussions with Groq.

Nvidia CEO Jensen Huang dismissed reports of friction with OpenAI as nonsense during remarks in Taipei, stating that Nvidia planned a huge investment in OpenAI. However, Huang clarified that a previously reported one hundred billion dollar investment proposal was never a commitment.

Market Implications

Nvidia shares fell following the news, extending earlier declines triggered by reports that the one hundred billion dollar investment talks had stalled. An OpenAI spokesperson stated that the company relies on Nvidia to power the vast majority of its inference fleet and that Nvidia delivers the best performance per dollar. However, competing products such as Anthropic's Claude and Google's Gemini benefit from deployments relying more heavily on proprietary chips optimised for inference calculations.

The Groq deal has bolstered the standing of other startups building their own AI chips, including D-Matrix and SambaNova, strengthening their valuations and making them more attractive acquisition targets in twenty twenty-six.