# Cerebras An AI chip startup that builds the world's largest processor — the Wafer-Scale Engine (WSE), which fabricates an entire silicon wafer as a single chip rather than dicing it into hundreds of smaller dies. Founded in 2016 by Andrew Feldman and headquartered in Sunnyvale, CA, Cerebras offers a fundamentally different architecture from [[NVIDIA]]'s GPU-cluster approach: one massive chip with enormous on-chip memory bandwidth, eliminating the inter-chip communication bottleneck that limits conventional AI accelerators. The third-generation WSE-3 is fabricated by [[TSMC]] on a 5 nm process and contains 4 trillion transistors, 900,000 AI-optimized compute cores, 44 GB of on-chip SRAM, and delivers 125 petaflops of peak AI performance across a 46,225 mm die (roughly 21.5 cm on a side). The WSE-3 powers the CS-3, a water-cooled appliance about the size of a mini-fridge that pairs the chip with up to 1.2 petabytes of external memory. ## Inference Play Cerebras has positioned itself primarily around AI inference rather than training — the reverse of most chip startups. On large language models the WSE architecture delivers responses up to 15x faster than GPU-based systems. Independent benchmarks showed Cerebras running DeepSeek R1 (Llama 70B) at over 1,500 tokens per second, roughly 57x faster than GPUs, and Llama 4 Maverick (400B parameters) at 2,500 tokens per second per user — more than double [[NVIDIA]]'s DGX B200 Blackwell system. In early 2026 [[OpenAI]] signed a multi-year agreement to deploy 750 MW of Cerebras wafer-scale systems for latency-sensitive workloads, including agentic AI applications. Deployment began in early 2026 and will continue in phases through 2028. ## IPO and G42 Controversy Cerebras filed its first S-1 in late 2024 but withdrew it after [[CFIUS]] opened a national security review of a $40 million investment by [[G42]], an Abu Dhabi-based AI company with ties to the UAE royal family. Roughly 86% of Cerebras's 2025 revenue (~$510M total) came from just two UAE-linked customers, raising concerns that the company could become a conduit for advanced American AI technology reaching China. [[G42]] divested its roughly 1% ownership stake by December 2025, and [[CFIUS]] cleared the deal in early 2025. Cerebras refiled its S-1 in April 2026 and completed the largest IPO of 2026 on May 14, listing on Nasdaq under ticker CBRS. The offering raised $5.55 billion at an implied valuation of approximately $56 billion. [[Sam Altman]] was an early investor in Cerebras through Y Combinator's continuity fund. ## Stargate Connection Cerebras was selected to build inference systems for the [[Stargate Project]], the $500B AI infrastructure initiative. Separately, Cerebras is deploying infrastructure for a UAE Stargate campus in Abu Dhabi — a planned 5 GW AI campus whose first 200 MW phase was expected online by 2026. ## KB Relevance Cerebras sits at the nexus of several vault themes: the AI chip supply chain beyond [[NVIDIA]], the national security implications of Gulf-state AI investment (the [[G42]]/[[CFIUS]] saga), the [[Stargate Project]] buildout, and [[The Quest for 90 Gigawatts]] driving AI infrastructure. Its IPO trajectory — delayed by geopolitical scrutiny, then completed at a massive valuation — illustrates how AI chip companies navigate the tension between capital needs and export-control politics. ## See Also - [[NVIDIA]] - [[OpenAI]] - [[G42]] - [[TSMC]] - [[CFIUS]] - [[Stargate Project]] - [[Sam Altman]] - [[Artificial Intelligence]] - [[The Quest for 90 Gigawatts]]