Cerebras WSE-3 ☆

Cerebras builds the largest chip ever made — an entire 300 mm wafer as one device. WSE-3 packs 4 trillion transistors, 900,000 cores and 44 GB of on-chip SRAM with 21 PB/s of internal bandwidth, sidestepping the HBM bottleneck entirely. The trade: models must be sharded across machines for weights, which Cerebras turned into an inference-speed story — serving Llama-class models at thousands of tokens per second. G42 of the UAE is its anchor customer; the company filed to IPO in 2024.