Cerebras’ $10 billion deal with OpenAI positions the startup and its wafer-scale engine as a challenger to Nvidia in the AI chip market, while helping OpenAI try to accelerate the performance of its large AI models.
The multiyear deal, revealed on Jan. 14, requires Cerebras to deliver 750 megawatts of wafer-scale systems to OpenAI starting later this year. OpenAI will use the wafer-scale engine to deliver near-real-time responses for tasks such as coding, inference, image generation, and complex reasoning. The Cerebras version of an AI chip is larger and, the vendor says, faster than Nvidia GPUs.
The agreement gives Cerebras, which has struggled to expand its customer base since it was founded in 2015 beyond Abu Dhabi-based AI and tech holding company G42, a foothold in the AI chip market as one of several chipmakers trying to rival Nvidia. It also addresses a key challenge for enterprises finding that massive generative AI models are often too slow or costly for real-time use.
A Stage for Cerebras
The deal gives Cerebras an opportunity to prove the capabilities of its wafer-scale engine.
While some have previously seen the AI vendor as a somewhat experimental company focused on science applications, this partnership is the “ultimate stamp of legitimacy,” said Mike Leone, an analyst at Omdia, a division of Informa TechTarget.
“It transforms them overnight from a niche alternative into a serious contender that every other AI lab now has to pay attention to,” Leone said.
Potential Benefits for Enterprises and OpenAI
Competition from chipmakers, including longstanding semiconductor companies Broadcom and AMD, along with Cerebras, could benefit enterprises by potentially driving down AI service prices over time.
“Overall, an enterprise customer is going to have more choices when it comes to how they get their AI stuff that they want,” said David Nicholson, an analyst at Futurum Group.
The deal also tackles the paramount market issue: AI model speed.
“The industry is grappling with a difficult trade-off right now where smarter models are becoming much heavier and slower to run,” Leone said. “It appears this partnership is trying to solve that specific friction point. By focusing on inference speed, they seem to be trying to ensure that future AI agents can handle complex tasks without the lag that currently frustrates users.”
For OpenAI, which has seen competition from Google and others intensify amid questions about OpenAI’s long-term financial viability, the agreement with Cerebras, is a win, particularly after Google’s major agreement with Apple earlier this week to power Apple’s AI initiatives with the Google Gemini foundation model. This deal shows OpenAI is expanding its infrastructure for a future when AI does the work for users, not just converses with them, Leone said.
“There is a deeper story here about the shift from chatbots to actual agents,” he said. “When you’re just chatting, a few seconds of delay is fine. But when you have an AI agent trying to solve a complex problem that requires twenty steps of ‘thinking’ in the background — that requires a level of speed that standard hardware struggles to deliver efficiently.”
Chipping Away at Nvidia
If Cerebras proves its hardware accomplishes that, it could lead to erosion of market share for Nvidia.
“Cerebras, in my opinion, is the biggest single threat, from a hardware company perspective, to Nvidia,” Nicholson said. “This negatively affects Nvidia in the long term, if OpenAI proves out that a data center built on Cerberus technology is superior.”
Nicholson noted that some observers see Nvidia as being at a relative disadvantage because it lacks a wafer-scale product. Nvidia’s approach involves producing many separate chips from a wafer by cutting them out and assembling them, which can result in mechanical and electrical errors due to discarded, faulty chips. In contrast, Cerebras keeps the entire wafer intact as a single, large chip, connecting only the working cells. This approach may reduce system complexity and potential error points.
“At its surface, it’s so obvious that what Cerebras does is technically superior,” Nicholson added. “But Nvidia has been able to get away with its inferior technology because it has the full stack and it has the industry connections and momentum.”
However, Cerebras faces the challenge of delivering at a massive scale almost immediately, Nicholson said.
Also, integration is a challenge. Enterprises interested in Cerebras might find its technology too complicated to integrate with their systems and find they need more talent to them it work better. This could make it less appealing to non-hyperscalers.
“It’s a lot easier to go with the ready-made solution from Nvidia,” Nicholson said.
Meanwhile, Nvidia forged a $20 billion licensing agreement with Groq, another AI chip startup, last month. And Nvidia has been in talks with OpenAI to sell the generative AI vendor Nvidia chips, representing 10 gigawatts.
Cerebras says it has also entered into deals with IBM and Meta. OpenAI and Cerebras had a pre-existing relationship, with OpenAI at one point exploring an acquisition of the chipmaker, the Wall Street Journal reported. OpenAI CEO Sam Altman is a personal investor in Cerebras.
