Cerebras makes and sells AI computing infrastructure built around a single, distinctive piece of hardware: the Wafer-Scale Engine (WSE), which uses an entire silicon wafer as one processor rather than cutting it into smaller chips. The current WSE-3 is 58 times larger than NVIDIA's B200 chip and contains 900,000 compute cores. Cerebras's core argument is that AI inference is a memory bandwidth problem — large language models must move model weights from memory to compute for every token generated, and GPU-based systems bottleneck on off-chip data transfers. By keeping compute and memory on one massive chip, Cerebras claims inference speeds up to 15 times faster than leading GPU-based solutions. Cerebras generates revenue through two channels: hardware sales (~70% of FY25 revenue), where Cerebras sells on-premises AI supercomputer systems directly to customers, and cloud services (~30%), where customers access Cerebras compute on either take-or-pay dedicated contracts or consumption-based on-demand pricing. Cerebras operates a fabless model, outsourcing all wafer fabrication to TSMC. The company sells to a concentrated customer base — MBZUAI and G42 together accounted for 86% of FY25 revenue. The business is transitioning toward a cloud-first model, anchored by a $20B+ multi-year agreement signed in December 2025 with OpenAI to purchase 750 megawatts of inference compute capacity over 2026-2028. Cerebras is also expanding into Sovereign AI — national AI programs, particularly in the Middle East, seeking domestic infrastructure with data sovereignty.