Cerebras claims its CS-4 server delivers up to 30x faster performance than Nvidia GPUs on AI chatbot query workloads.
Cerebras Systems (CBRS) unveiled its next-generation CS-4 server rack on Tuesday, claiming the new hardware delivers up to 30 times more tokens per second per user than graphics processing units when running large language models. The announcement marks an aggressive push into the AI inference market, where the Sunnyvale, California-based company is positioning its wafer-scale architecture as a faster, more efficient alternative to the GPU clusters that dominate data centers today.
The CS-4 is a rack-scale system powered by three WSE-3 Turbo processors, chips the company describes as the largest AI semiconductors ever built, each packing 4 trillion transistors. Cerebras said the system can produce more than 4,400 tokens per second per user on the GPT-OSS-120B model, roughly double the speed of its CS-3 predecessor while delivering up to 10 times higher throughput per watt. Initial CS-4 shipments are scheduled to begin during the current quarter, with the chips fabricated using TSMC's 5-nanometer manufacturing process.
The system is built around the company's Nexus server architecture, which uses pluggable modules to house the chips. Cerebras said the design requires 50 percent fewer components than previous generations, a change Chief Technology Officer Sean Lie said would accelerate data center construction timelines. The rack also includes new networking components aimed at speeding data movement between chips.
"Historically, fast inference meant using smaller and less capable models. Cerebras CS-4 delivers industry-leading speeds on the largest frontier models, fundamentally changing the paradigm," Chief Executive Officer Andrew Feldman said in a statement. "Every aspect of the design has been optimized to deliver the highest speeds with massive throughput. With the CS-4, AI is so fast that it fundamentally reshapes product experiences."
Feldman told reporters at a San Francisco briefing that engineering efforts are focused on boosting data throughput. "We're going to get four times as fast between now and the end of 2027, and we're going to get 20 times more throughput," he said. The company expects to deliver 600 megawatts' worth of computing power by the end of 2027 and plans another generation of chip and server hardware that year.
The performance claims hinge on a fundamental architectural difference. Cerebras processors use static random-access memory, or SRAM, rather than the dynamic random-access memory found in conventional GPU systems. SRAM is significantly faster but also more complex and expensive, making it impractical for standard-sized chips. Cerebras argues its dinner-plate-sized wafers provide enough physical space to leverage SRAM effectively. The single-chip design also reduces the distance data must travel compared with multi-chip GPU clusters from Nvidia (NVDA) or AMD, which must shuttle information between discrete processors.
Cerebras said this advantage is particularly relevant for inference workloads — the computing process of generating responses in AI chatbots such as Anthropic's Claude. In addition to selling hardware, the company operates its own AI cloud services, renting out access to systems running on its chips. The broader competitive landscape remains challenging. Nvidia continues to dominate the AI accelerator market with its H100 and Blackwell product lines, and major cloud providers have built extensive infrastructure around GPU-based systems. Cerebras is betting that inference — which accounts for the day-to-day cost of running deployed AI models — represents an opening for its wafer-scale approach.
The product launch comes during a volatile period for Cerebras shares. The company went public in May at $185 per share and began trading at $350, but the stock has since fallen more than 35 percent, trading around $218 as of midday Tuesday. Last week, Cerebras reported an adjusted loss of $6.9 million on sales of $180.1 million for the second quarter. The company posted a loss per share of $2.98, compared with a profit of $1.91 in the same period a year earlier. Despite offering better-than-expected third-quarter guidance, the results failed to satisfy Wall Street. Shares were little changed in premarket trading following the CS-4 announcement, suggesting investors remain focused on whether the performance claims translate into commercial adoption.
Whether that bet pays off will depend on the company's ability to convert technical benchmarks into sustained customer demand, particularly among enterprises weighing power consumption and infrastructure costs as AI deployments scale.