Cerebras Systems, Amazon strike deal to offer AI chips on AWS cloud

Amazon.com and Cerebras Systems on Friday said they have reached a deal to combine the two companies' computing chips in a new service aimed at speeding up chatbots, coding tools and other artificial intelligence services.

Valued at $23.1 billion, Cerebras is a chip startup aiming to take on Nvidia by building a fundamentally different kind of AI chip that does not rely on expensive high-bandwidth memory ‌as Nvidia's flagship ⁠chips ⁠do. Earlier this year, Cerebras signed a $10 billion deal to supply chips to ChatGPT creator OpenAI.

Under the deal announced Friday, Cerebras chips will sit inside Amazon Web Services (AWS) data centers and be linked to Amazon's own Trainium3 custom AI chips, connected with custom networking technology from Amazon.


"Every customer large or small is on AWS, from individual developers to the largest banks in the world," Cerebras CEO ​Andrew Feldman told Reuters, saying the deal will "make it easy ⁠as a ‌click to get on Cerebras."

Both companies declined to disclose the size ​of the deal.

Amazon ​and Cerebras will team up to tackle what is known as "inference," ⁠where previously trained AI systems take requests from users and spit ​out answers. The two companies will split up that task ​into two steps, one called "prefill" where the user's request is transformed from human words into the language of "tokens" that AI computers use, and a "decode" stage where the AI computer provides the answer the user is looking for.

Amazon said its Trainium3 chips will handle prefill, while Cerebras chips handle decoding, what Feldman told Reuters is a "divide and conquer strategy."

It is a similar ‌strategy to the one that analysts expect Nvidia to unveil next week, when it details how it plans to combine its own graphics processing ​unit (GPU) chips with ​those from Groq, a startup ⁠it spent $17 billion on in late December. In a statement, Amazon said that it could not yet make a detailed comparison between its offering, which will come online in the ​second half of this year, and Nvidia's as-yet-unrevealed offering, but Amazon expects its service to be a better value.

"The timeline for that (Nvidia-Groq) pairing remains unclear while our Trainium3 program is just months away from running production workloads," Amazon said in response to Reuters questions. "What we can say is that we believe (Trainium3)-and future (Trainium4)-will continue to lead in price-performance versus merchant GPUs."