Since Cerebras came on the scene in 2019 with its unusual dinner plate-size wafer scale GPU, there was always the potential for a break-out moment when someone or something would elevate the lightning-fast processer from niche AI chip into the mainstream. That moment may have come today.
The Wall Street Journal reports that OpenAI has agreed to a $10 billion deal to purchase up to 750 megawatts of computing power over three years from Cerebras.
This has been a good period for the AI processor group that includes Groq and SambaNova Systems, along with Cerebras. Intel reportedly is in advanced talks to acquire SambaNova for about $1.6 billion, and NVIDIA recently entered a non-exclusive licensing deal with Groq for roughly $20 billion.
As for Cerebras, OpenAI struck the deal to address its ongoing computing deficit.
“OpenAI is racing to secure access to more data-center capacity as it prepares for its next phase of growth,” the Journal stated. “The company counts more than 900 million weekly users, and executives have said repeatedly that they are facing a severe shortage in computing resources.”
OpenAI’s planning includes finding “cheaper and more efficient alternatives to chips designed by Nvidia,” according to the Journal. This strategy has led to a custom chip designed with Broadcom and a deal to use AMD MI450 GPUs.
Cerebras CEO Andrew Feldman told the Journal that the OpenAI deal was signed nearly two months ago. In an interview with the publication, he played a video showing chatbots running on Cerebras chips were faster to handle user queries than those running on other processors.
“What is driving the market right now,” Feldman said, is “this extraordinary demand for fast compute.”
OpenAI infrastructure executive Sachin Katti said the company is particularly interested in using Cerebras to run AI applications for coding.
The development makes Cerebras a hot commodity in the exploding HPC-AI sector. The company is in talks to raise $1 billion at a valuation of $22 billion, according to the Journal and The Information. And research firm PitchBook reports Cerebras has raised funds totaling $1.8 billion thus far.
In retrospect, the OpenAI-Cerebras deal is not a major surprise. Last August, Cerebras announced inference support for gpt-oss-120B, OpenAI’s first open-weight reasoning model, running at inference speeds of 3,000 tokens per second on the Cerebras AI Inference Cloud. Cerebras said this was the first time an OpenAI model leveraged Cerebras’ wafer-scale AI infrastructure to run full-model inference.
The company has been reporting eye-popping processing speeds for years. Last May, Meta said it teamed with Cerebras on AI inference in Meta’s new Llama API, combining Meta’s open-source Llama models with inference technology from Cerebras. Developers building on the Llama 4 Cerebras model in the API saw speeds up to 18 times faster than traditional GPU-based solutions, according to Cerebras.
Here are other reports of Cerebras’ processing prowess.
