OpenAI’s Cerebras Deal and the Race for Faster AI Inference

OpenAI has announced a partnership with Cerebras to bring more high-speed inference capacity online. I want to unpack why this matters, beginning with Google’s Gemini 3 and the hardware competition around it. Inference is the work a model does when it processes a request and produces a response.

Google released Gemini 3 in November 2025, and I saw it as a strong frontier model on many benchmarks. Google used its own TPUs for training and serving; that illustrates that Nvidia GPUs are not the only hardware capable of supporting major models. It was not, however, the first powerful model trained on non-GPU hardware.

On Christmas Eve, Nvidia and Groq announced a deal involving Groq’s inference technology. Groq described it as a non-exclusive license, said founder Jonathan Ross, president Sunny Madra and other team members would join Nvidia, and said Groq would remain independent. Reports put the deal’s value around $20 billion; my claim about an antitrust motive is an interpretation, not an established explanation.

I see inference as a central source of AI revenue: training involves substantial upfront costs, while a deployed model can serve many requests. Faster, cheaper service could improve margins as demand grows. The economics are more complicated than a model being trained once and served forever: post-training, updates, inference costs and pricing all matter too.

The Wall Street Journal reported that OpenAI’s Cerebras deal was worth more than $10 billion and involved 750 megawatts over three years. The companies announced a multi-year agreement for staged deployment beginning in 2026; a later Cerebras filing says the 750 MW is to be deployed in tranches through 2028, each with a three- or four-year term that OpenAI can extend to five. In June 2026, Cerebras described the agreement as worth more than $20 billion.

I can imagine OpenAI eventually acquiring Cerebras, but that is speculation, not part of the announced deal. Cerebras CEO Andrew Feldman had recently appeared on our live show without mentioning it. I also suspect OpenAI values a diversified supply of compute rather than relying too heavily on one hardware vendor; the companies did not establish that as the sole reason for their choice.

To illustrate speed, I show a chart for OpenAI’s open-weight gpt-oss-120b model at high reasoning effort, with 1,000 input tokens. The chart lists 3,137 output tokens per second for Cerebras and 465 for Groq—roughly 6.7 times in that particular comparison. Provider configuration and measurement conditions matter, and this does not establish that ChatGPT as a whole will be 100 times faster.

The video includes a paid promotion for Zapier Agents, which connects AI models and tools to applications. Its product and integration claims are the sponsor’s promotional claims.

Cerebras also uses a distinctive wafer-scale architecture, combining compute and substantial memory on a large chip. I say everything is on the wafer, but that is too broad: Cerebras documents weight-streaming systems that store model weights in external MemoryX and stream them to the chip. The hardware configuration and workload therefore matter when considering its memory needs.

In a clip from our live show, Feldman says Cerebras does not use the memory whose rising price and limited supply we were discussing. He argues that competing hardware vendors waiting for such components could give Cerebras an advantage. That statement concerns the particular memory market being discussed, not immunity from every supply-chain constraint.

I imagine a much faster ChatGPT and emphasize that speed belongs alongside cost and quality in the AI trade-off. In a poll I ran on X, quality received the most votes, while I recall speed getting around 2%. That informal poll captures my audience at the time, not a representative measure of all users.

For coding, waiting for an agent to finish can interrupt the flow of work. Queued tasks and parallel agents help, but faster responses would allow more iterations; I see similar value in many interactive uses. Deep research is an exception for me: I am comfortable starting it and returning later when the answer is ready.

I expect specialized hardware to play a large part in inference, while general-purpose accelerators continue to play a role in training. TPUs also show that major models can be trained on hardware other than Nvidia GPUs. Actual speed and costs depend on the model and system, so a specialized chip is not automatically faster for every workload.

Additional inference capacity might reduce competition for other computing resources, potentially leaving more capacity for training. I predict that could help OpenAI build better models, but the company’s allocation decisions and the resulting model quality are not established by this deal. Extra capacity is not proof that all existing GPUs will move to training.

At the time of the video I thought the agreement might support Cerebras’s IPO ambitions and encourage other labs to examine its capacity. That was a prediction. As a later development, Cerebras reported in June 2026 that it had completed its IPO.

I had earlier asked OpenAI’s Greg Brockman why frontier labs did not use more services from companies such as Groq and Cerebras, and whether software changes were a barrier. In the interview excerpt, Brockman recalls OpenAI’s excitement about Cerebras in 2017, but says non-GPU architectures proved harder to build than expected.

Brockman says OpenAI spoke with chip companies about the shape of its workloads and found that some did not adopt its perspective on large models. In his view, successful hardware developers paid attention to deep-learning expertise and where workloads were heading. His recollection is an account of those conversations, not proof of any particular company’s internal decisions.

I conclude that more companies are recognizing the value of specialized inference, pointing to Nvidia’s Groq license and OpenAI’s Cerebras partnership. I believe the race for capacity could benefit users, although demand, pricing, access, reliability and implementation will all shape the actual outcome.