Customer concentration and manufacturing capacity
Q: Timothy Arcuri of UBS asked how the company views customer concentration for next year given the OpenAI ramp and the potential for AWS to be a large contributor.
A: Andrew Feldman stated that OpenAI will remain a meaningful portion of revenue next year, but AWS and other customers in coding and security will represent a larger share as the business diversifies.
Q: Timothy Arcuri of UBS asked for the expected growth in manufacturing capacity for next year.
A: Bob Komin noted that manufacturing capacity will end 2026 at over 10x the previous level, and the company has already contracted for 3x to 4x more growth for 2027.
AWS and AMD partnerships
Q: Joshua Buchalter of TD Cowen asked about the economics of the Amazon deal and its availability.
A: Andrew Feldman explained that the service will be delivered through AWS Bedrock, and the company is currently organizing deployments with an expected Q1 2027 launch.
Q: Joshua Buchalter of TD Cowen asked about the go-to-market strategy for the AMD Helios rack partnership and how AMD's recent acquisition fits in.
A: Andrew Feldman stated the joint solution uses Helios racks for prefill and Cerebras systems for decode, which is a compelling offering with existing buyers, while noting that AMD's recent acquisition is not immediately applicable to data center inference.
Q: Kyle Bleustein of Barclays asked if revenue from the disaggregated inference solution using AMD Helios racks goes to Cerebras.
A: Andrew Feldman confirmed that it does.
Q: Kyle Bleustein of Barclays asked if there is a path to hosting Trainium and CS3 together in the company's cloud or other hyperscale clouds.
A: Andrew Feldman confirmed the company is interested in deploying hyperscaler hardware outside of their own data centers, which is on the table for the future.
Disaggregated inference and product roadmap
Q: Quinn Bolton of Needham & Co. asked if the AMD deal represents an additional revenue opportunity if the service is delivered to OpenAI.
A: Andrew Feldman explained that increasing throughput while maintaining speed increases revenue per system, improves margins, and makes data center investments more valuable.
Q: Quinn Bolton of Needham & Co. asked if the planned 20x increase in throughput by the end of 2027 removes the need for disaggregated compute.
A: Andrew Feldman stated the company is exploring various ways to increase throughput, including disaggregated solutions and new technologies, as part of its roadmap.
Q: Joseph Moore of Morgan Stanley asked about the status of commercializing disaggregated decode.
A: Andrew Feldman stated that disaggregated inference with GPUs is currently running in labs and will be deployed and available in Q4.
Q: Joseph Moore of Morgan Stanley asked if working with Amazon and AMD hardware will be better than using existing NVIDIA GPUs.
A: Andrew Feldman noted that newer hardware generations will provide better solutions, but disaggregated solutions will be vastly better than non-disaggregated ones regardless of the GPU generation used.
Data center capacity and market expansion
Q: Vijay Rakesh of Mizuho asked if the 600-megawatt capacity and 10x manufacturing increase will help accelerate 2027 ramps.
A: Andrew Feldman confirmed that the company is actively pursuing data center capacity globally, and faster deployment directly correlates to faster revenue growth for both cloud and on-prem customers.
Q: Vijay Rakesh of Mizuho asked how the pipeline is developing for neoclouds.
A: Andrew Feldman stated that neoclouds are increasingly diversifying away from single hardware vendors, creating a large opportunity for Cerebras in 2027.
