Cerebras Systems has unveiled the CS-4, a rack-scale computing system built to greatly speed up AI inference, as the chipmaker pushes to compete with NVIDIA and other providers of AI infrastructure.
The CS-4 includes three of Cerebras’ new Wafer Scale Engine 3 Turbo, or WSE-3T, processors. These unusually large chips are designed to keep more computing operations within a single processor, reducing the delays and energy use created when data must travel between separate chips.
Cerebras claims the CS-4 can deliver AI inference performance up to 30 times faster than GPU-based systems and twice the speed of the company’s previous CS-3 system. Fast inference is particularly important for AI agents, which may require multiple model calls to complete a task.
“Rack designs are getting as heterogeneous as the chips inside them, and Cerebras faces the most unique challenges with wafer scale,” Brendan Burke, Research Director at The Futurum Group, told Techstrong Semi. “The Supernova rack puts liquid cooling and power delivery front and center because doubling power per chip is the primary performance unlock.”
Networking is the bottleneck to making disaggregated inference a reliable reality, he added. “Cerebras integrating a new I/O module with Arista Etherlink switches to scale one chip to three per rack shows the industry converging on that view.”
“With 10x manufacturing capacity coming in 2026 and 600 MW of data center capacity in development, this is the rack that has to prove Cerebras can serve trillion parameter models at scale.”
Using a Modular Platform
The new system provides 750 petaflops of AI compute, along with 129.6 petabytes per second of memory bandwidth and 7.2 terabits per second of I/O bandwidth. In a test using a 120-billion-parameter model, Cerebras said the CS-4 generated more than 4,400 tokens per second for each user.
The WSE-3T is manufactured using TSMC’s 5-nanometer process. Its turbo mode offers twice the compute performance, memory fabric capacity and I/O bandwidth of the WSE-3 introduced two years ago.
Cerebras has also redesigned the infrastructure surrounding its wafer-scale processors. The CS-4 is the first system based on the company’s Nexus architecture, a modular platform that separates compute, power and I/O functions. This approach allows Cerebras to upgrade individual parts of the system without requiring a complete redesign.
A new rear-mounted module integrates power conversion, liquid cooling, I/O and control electronics close to the wafer. The CS-4 contains 50% fewer components than the prior generation and uses 60% more automated manufacturing. Cerebras claims these changes can reduce deployment time from days to hours.
The CS-4 also includes two approaches to networking. Customers can connect it to existing infrastructure through Ethernet, while Cerebras’ Direct Wafer Links connects processors across systems without requiring a network switch. Cerebras says these direct connections can reduce wafer-to-wafer latency to as little as two microseconds and support AI models containing more than 50 trillion parameters.
Cerebras expects the first CS-4 systems to ship during the third quarter. The company is developing another generation of its chip and server technology for 2027, with an engineering roadmap that plans a fourfold gain in processing speed and a twentyfold improvement in throughput by that point.
