OpenAI’s big Cerebras bet

OpenAI has agreed to spend more than $20 billion over three years on servers built with Cerebras chips, according to Reuters and The Information. (reuters.com) The deal could also give OpenAI an equity stake in Cerebras, the Sunnyvale, California chipmaker that has been trying to challenge Nvidia in artificial intelligence hardware. Reuters reported the agreement on April 16, citing The Information. (reuters.com) Cerebras said on April 17 that it had filed an S-1 with the U.S. Securities and Exchange Commission for a proposed initial public offering. Reuters said the company had withdrawn an earlier IPO filing in October 2025 after a funding round valued it at about $8 billion. (cerebras.ai) (reuters.com) The purchase points to the bottleneck that has defined the artificial intelligence boom: computing power. Training and running large models requires vast clusters of chips, and OpenAI has been lining up more suppliers and more data center capacity as demand for ChatGPT and related products grows. (reuters.com) (openai.com) OpenAI, SoftBank, Oracle and MGX announced the Stargate project in January 2025, saying it intended to invest $500 billion in United States artificial intelligence infrastructure over four years, with $100 billion to be deployed immediately. OpenAI said it would have operational responsibility, while SoftBank would lead financially. (openai.com) (group.softbank) Cerebras sells a different kind of processor from the graphics processing units, or GPUs, that dominate the market. Its flagship Wafer-Scale Engine is built on a single giant silicon wafer, and Cerebras says the current WSE-3 packs more than 4 trillion transistors and powers its CS-3 systems. (cerebras.ai 1) (cerebras.ai 2) The company has been pitching those systems as easier to scale for very large models because more of the computing happens on one chip instead of being split across many smaller ones. Cerebras says a single CS-3 can scale to 24 trillion-parameter models as one logical device. (cerebras.ai 1) (cerebras.ai 2) Cerebras already has experience building large artificial intelligence clusters with G42, the Abu Dhabi technology group. In 2023, the two companies announced Condor Galaxy, a planned network of nine interconnected AI supercomputers, and in March 2024 they said Condor Galaxy 3 would use 64 CS-3 systems and deliver 8 exaFLOPs of AI performance. (g42.ai) (cerebras.ai) For OpenAI, the reported Cerebras deal would add another path to secure scarce compute without relying entirely on Nvidia-based supply. For Cerebras, landing OpenAI as a customer and possible shareholder would strengthen its case to public-market investors days after reviving its IPO plans. (reuters.com 1) (reuters.com 2)