Cerebras unveiled its CS-4 rack-scale system Friday: three wafer-scale engines per rack — processors etched across an entire silicon wafer rather than diced into chips — with power and compute separated into modular units. The company claims up to 30 times the inference speed of the nearest GPU competitor and ten times the throughput of its own CS-3, pitching latency as the thing users actually feel. One wafer, Cerebras says, replaces hundreds of GPUs. [1][3]

The claim is Cerebras's own; independent benchmarks do not exist yet. But the timing is the message. Six days from now Nvidia reports quarterly results, with consensus near $91.9 billion in revenue — a number that would have been unthinkable for a chip company a few years ago. [2] Nvidia spent the week denying reports of a China-specific processor and confirming early talks with Korean startup Rebellions that could end in partnership, investment, or acquisition. [1]

The engineering logic is almost austere: make the chip the size of the wafer, keep the data on one piece of silicon, and sell speed as the product. Whether 30x survives contact with customers is exactly what the next week tests — first with benchmarks, then with an earnings call carrying the whole market's assumptions about AI economics on its back.

-- KENJI NAKAMURA, Tokyo