Cerebras Systems on Tuesday unveiled its CS-4 AI accelerator, positioning the wafer-scale system as the industry's fastest inference platform amid a broader push to expand data center capacity to 600 megawatts by the end of 2025.
The CS-4, currently in early access, delivers up to 30 times faster inference than GPU-based solutions on frontier models and doubles tokens-per-second throughput compared with its predecessor, the CS-3. Key performance metrics include 43 petabytes per second of memory bandwidth per chip—2,000 times greater than NVIDIA's next-generation GPU measure—along with wafer-to-wafer input/output latency reduced to 2 microseconds, a 2.5 times improvement. Network latency has also been cut to 3 microseconds from 5 microseconds previously.
The system leverages Cerebras's Wafer-Scale Engine (WSE-3 Turbo) and Nexus rack-scale platform, featuring 50% fewer components and 60% more manufacturing automation. Deployment times have been reduced to hours from days, while the backpack design offers twice the power and cooling capacity. On-chip communication provides 200 times more bandwidth than off-chip cables.
Cerebras also outlined plans to double raw performance annually through 2027, targeting 20 times higher throughput by that year. The CS-5 system, slated for 2027, will serve as the next-generation wafer-scale engine. The company has 600 MW of data center capacity either online or under contract for delivery by the end of 2025, with locations spanning Santa Clara, Toronto, Dallas, Minneapolis, Montreal, Oklahoma City, Lyon, Mikkeli, and additional sites in Norway.
Revenue reached $680.67 million in the last twelve months as of Q2 2026, up 91% year-over-year. Cerebras's shares closed at $219.92 on Tuesday, down 12.73% from the previous session's $251.98 close, though after-hours trading showed a modest gain to $220.39. The stock has traded between $160.81 and $386.34 over the past 52 weeks.
During the Supernova 2026 conference, CEO Andrew Feldman highlighted the CS-4's performance on OpenAI's GPT-5.6 Sol Ultrafast mode, which runs frontier models up to 14 times faster than standard tiers. Thibault Imbert, OpenAI's head of core products and platforms, noted that the ultrafast mode eliminates trade-offs between speed and intelligence while reducing cognitive load for end users.
Cerebras also demonstrated real-world applications, including a Figma design agent fine-tuned on billions of designs and Cognition's Sui 1.7 model, which achieves GPT-5.5 Opus-equivalent quality at nearly 1,000 tokens per second. The company's SuiGrab and SuiGrab Mini variants operate at 680 and 2,800 tokens per second, respectively, enabling simultaneous tool calls that improve end-to-end task completion by four times compared with sequential methods.
In cybersecurity, Cerebras reported that faster inference allows five to ten times more inspection within acceptable time windows, with Armis completing code security scanning in one-third the time of GPU alternatives. The company also highlighted a benchmark where its system completed a 2,500-question exam in 11 hours, 11 minutes, and 26 seconds—less than half a day—compared with competitors taking over three days.
