Cerebras Announces CS-4 AI Chip with Double Performance

Cerebras announced CS-4 AI accelerator with double CS-3 performance on same chip, enabling faster AI applications in healthcare and resource optimization.

Cerebras announced its CS-4 AI accelerator on August 24, 2026. The chip delivers double the performance of its CS-3 predecessor on the same 5nm WSE-3 architecture, achieving 4,400 tokens per second per user and up to 30 times faster than Nvidia GPU setups. It maintains 44 GB memory per wafer and uses a new modular 'Backpack' design for disaggregated inference with AMD and AWS Trainium. OpenAI has deployed CS-4 for Codex Spark, and the product scales as a single data center cabinet unit. This advancement could lower costs for AI applications in healthcare, climate modeling, and resource optimization. However, SemiAnalysis analysts note networking gains are modest, meaning real-world cost reductions depend on implementation efficiency rather than raw speed alone.

Source: The Decoder