Cerebras

Cerebras Systems, founded in 2015 and based in Sunnyvale, California, builds specialized hardware for AI computation, centered on its Wafer-Scale Engine (CS-3) chip. The company describes the chip as dramatically larger than a standard GPU die, a design choice that lets it hold entire model weights on-chip in SRAM rather than shuttling them across networked GPUs, which the company says translates into markedly faster inference. On top of that hardware, Cerebras runs an Inference Cloud service offering hosted access to open models such as Llama, Qwen, and GLM through an API. The API is built to be a drop-in replacement for the OpenAI Chat Completions API, so existing applications can switch providers with minimal code changes. Cerebras reports throughput figures in the thousands of tokens per second on supported models, and offers three deployment modes: a hosted cloud, dedicated infrastructure for a single customer, and fully on-premise systems for organizations with data residency requirements. The company has attracted partnerships with major AI players, including OpenAI, AWS, and Meta, and recently announced work with AMD combining Cerebras's fast token generation with AMD's prefill capacity. Cerebras targets developers and enterprises that need low-latency inference at scale rather than training-only workloads.