Cerebras launches CS-4 AI accelerator

Cerebras Systems introduced the CS-4 rack-scale AI accelerator on August 18, 2026, presenting it as its latest system for large-scale inference workloads. The CS-4 is built as a rack solution composed of three Wafer Scale Engine 3 Turbo (WSE-3T) processors, integrating them into a single appliance aimed at high-throughput deployment.

For this three-wafer configuration, Cerebras states that the CS-4 delivers 750 PFLOPS of AI compute. The company positions this as its highest-performing rack system to date, designed to support demanding generative AI and large language model applications that require both speed and scale.

Cerebras publishes detailed performance specifications for the CS-4. The system is specified to provide 129.6 petabytes per second of memory bandwidth and 160.5 petabytes per second of on-chip fabric bandwidth, highlighting the data-movement capacity within and across the wafers.

The CS-4 is also described as offering 7.2 terabits per second of system I/O bandwidth, intended to connect the appliance to external networks and storage. Wafer-to-wafer latency is stated to be as low as two microseconds, which is designed to support scaling to larger clusters and the deployment of very large models.

At the heart of the CS-4 is the WSE-3T processor, which Cerebras characterizes as its latest wafer-scale engine. Each WSE-3T is said to contain four trillion transistors and approximately 900,000 AI-optimized cores, along with 44 GB of on-wafer SRAM.

Cerebras cites the WSE-3T as delivering 250 PFLOPS per wafer, forming the basis for the aggregate 750 PFLOPS figure in the three-wafer CS-4 rack. The dense on-wafer memory and large core count are presented as supporting high parallelism and fast access for AI inference workloads.

In system-level comparisons, Cerebras highlights CS-4 performance on the GPT-OSS-120B model. The company claims the system can deliver more than 4,400 tokens per second per user, a figure aimed at demonstrating throughput for large language model inference.

Cerebras also asserts that, in its head-to-head testing, the CS-4 offers up to 30 times faster inference than GPU-based solutions. These comparisons are used to position the rack system as a high-speed alternative for enterprises and organizations running intensive AI inference workloads.

Relative to its prior-generation CS-3, Cerebras states that the CS-4 delivers up to 10 times more throughput per watt, emphasizing improved energy efficiency. The company also indicates that a single CS-4 can achieve up to twice the speed of a single CS-3 system, underscoring generational performance gains.

Cerebras describes the CS-4 as integrating into a modular Nexus rack architecture, intended to simplify deployment and future upgrades. The company plans to begin first shipments of the CS-4 this quarter, signaling the transition of the platform from announcement to commercial availability.