Cerebras
AI chip company that builds the Wafer Scale Engine — literally a chip the size of an entire 300mm wafer. WSE-3 (2024): 4 trillion transistors, 900,000 cores, 44 GB on-chip SRAM. Scale via monster chip instead of multi-chip distribution.
Cerebras Inference: 1800+ tokens/sec on Llama-70B (2024). Took over Groq's position as "fastest hosted LLM inference". Training: large training loads on a single WSE skip all multi-GPU communication overhead. Price/wafer ~$3-5M. Use: pharmaceutical AI (Argonne National Lab), defense, oil & gas. Distinct from Nvidia: extreme simplification (one unit) instead of flexible multi-GPU. SambaNova is competitor.