Cerebras unveiled its fourth-generation CS-4 AI system, claiming inference speeds exceeding 4,400 tokens per second per user on GPT-OSS-120B — up to 30 times faster than GPU-based solutions. The CS-4 delivers 750 PFLOPS of AI compute and 7.2 Tbps of I/O, with 10 times more throughput per watt than the previous CS-3. The company reported second-quarter 2026 core cloud and services revenues of $127.7 million, up 287% year over year. Total core revenues rose 103% to $209.9 million. Cerebras disclosed $25.4 billion in remaining performance obligations and over 600 megawatts of data-centre capacity under contract for delivery by end of 2027. The company faces competition from NVIDIA, which reported $75.2 billion in first-quarter fiscal 2027 data-centre revenues, and AMD's Helios platform.

Cerebras Systems unveiled its CS-4 server rack for AI inference on 18 August in San Francisco. The Sunnyvale, California-based chipmaker said the system uses three WSE-3 Turbo chips and new networking components. CS-4 is the first product based on Cerebras' Nexus architecture, a modular design for compute, power, and input/output. The system features programmable input/output and direct wafer links to connect wafers within and across racks with lower latency. The chips are manufactured using TSMC's 5-nanometre process. The launch follows Cerebras reporting an adjusted loss of $6.9 million on sales of $180.1 million last week. The system will be available in the third quarter. Cerebras competes with Nvidia in inference hardware for workloads such as chatbot response generation.

Cerebras Systems has launched its CS-4 AI system, promising up to twice the token-generation speed of its predecessor, six times higher system-level performance, and up to 10 times more tokens per watt in certain applications. The company announced partnerships with OpenAI and AMD to enhance inference capabilities. The AMD collaboration features a split architecture where GPUs handle model prefill whilst Cerebras manages token decoding. Cerebras is expanding its data-centre capacity, with 600 megawatts of power expected online or under contract by the end of next year. The company is targeting AI-agent, design, coding, and cybersecurity applications that benefit from lower latency. CEO Andrew Feldman emphasised that inference speed has become a product-level consideration. Cerebras hardware powers OpenAI's GPT-5.6 Sol Ultrafast mode, which runs models at up to 14 times standard speed.