Cerebras Systems introduced the CS-4, a server rack powered by three of its latest wafer-scale chips designed to accelerate AI chatbot inference tasks such as those used by Anthropic’s Claude models.
The system leverages the company’s Nexus server architecture, which employs pluggable modules to house the processors. Fabricated on TSMC’s 5-nanometer process, the WSE-3 Turbo chip reduces component count by 50% to simplify deployment and shorten data center build times. The CS-4 also incorporates new networking components aimed at improving inter-chip data transfer speeds.
Cerebras plans to begin shipping the CS-4 in the third quarter of this year, with a successor generation slated for 2027. At a media briefing, CEO Andrew Feldman projected the company’s total installed computing power to reach 600 megawatts by the end of 2027, up from current levels.
Feldman also outlined performance targets, stating that the company expects to deliver four times the speed and 20 times the throughput by year-end 2027 compared with its current systems. The announcement was made on August 18, 2026, in San Francisco.
Separately, Cerebras reported an adjusted loss of $6.9 million on revenue of $180.1 million for the week ending August 10, reflecting continued investment in R&D and scaling operations.
