Cerebras Systems (CBRS) investor relations material

Key announcements and product roadmap

• Announced CS-4, a next-generation system delivering up to 2x faster tokens and 10x more tokens per watt, with a modular Nexus rack-scale platform for rapid deployment and future scalability.

• Roadmap targets doubling speed every year and achieving 20x more throughput by 2027, with CS-5 and CS-6 planned on the same modular platform.

• CS-4 features a new Wafer-Scale Engine (WSE-3 Turbo), three wafers per system, and 43 PBps memory bandwidth per wafer, enabling up to 30x faster inference than GPUs.

• Disaggregated inference partnerships with AMD and AWS allow optimal hardware use for prefill and decode, maximizing speed and throughput.

• Early access for CS-4 is available now, with general availability later this quarter.

Industry partnerships and customer impact

• Collaboration with OpenAI enables GPT-5.6 Sol Ultrafast mode, delivering up to 14x faster inference and removing the speed-intelligence trade-off.

• Partnerships with Figma and Cognition showcase agentic applications in design and coding, leveraging Cerebras hardware for real-time, high-quality results.

• CrowdStrike partnership highlights the critical role of inference speed in cybersecurity, enabling faster, more effective threat detection and response.

• Arista partnership focuses on networking innovations to support hyperscale AI clusters, emphasizing the need for high bandwidth, low latency, and robust recovery.

• AMD partnership enables disaggregated solutions, combining GPU throughput with Cerebras decode speed for optimal economics and user experience.

Technology and performance insights

• Wafer-Scale Engine architecture eliminates memory bandwidth bottlenecks, allowing all model weights to reside on-chip and enabling ultra-fast inference.

• CS-4’s modular design supports rapid manufacturing, deployment, and future upgrades in power, compute, and I/O.

• New wafer I/O module delivers 2x bandwidth, 2x lower latency, and direct wafer-to-wafer links, supporting models up to 10 trillion parameters at 1,000 tokens/sec.

• Disaggregated inference splits prefill (GPU) and decode (Cerebras), reducing latency and maximizing throughput for large models.

• Real-world benchmarks show CS-4 achieving up to 30x faster inference than GPUs across a range of models, transforming user experience and agentic workflows.