Cerebras is announcing CS-4, a wafer-scale inference accelerator system using three WSE-3 Turbo processors per rack. The core claims: 30x faster inference than GPUs on production workloads, up to 10x better throughput-per-watt versus their prior CS-3 generation, and ability to sustain 1,000+ tokens/second on 10+ trillion parameter models with 2-microsecond wafer-to-wafer latency. The engineering specifics worth noting for deployment: modular "backpack" compute design that lets you stage power/cooling infrastructure separately (reducing install from days to hours), power delivery at 0.5mm from chip versus ~50mm on GPUs, and a programmable I/O subsystem that eliminates need for a separate switch fabric between wafers. First shipments start this quarter. The source doesn't provide independent benchmarks, pricing, or power envelope specs—all numbers are vendor claims. One practical question: how does the claimed 30x speedup map to real model sizes and batch configurations practitioners actually run, versus the favorable cases Cerebras likely tested?