doubleAI, an AI startup led by Mobileye co-founder Amnon Shashua, has released its first "code execution model," Maxwell, The Wall Street Journal reported. The company positions such models as a new category between code and chips: unlike large language models that help people write code, Maxwell resides directly inside running programs and rewrites code on its own to lower GPU and CPU usage costs; its core breakthrough is using formal mathematical methods to verify the correctness of the rewritten code.
The company says one early customer's compute bill was cut by 67% as a result. Zeev Farbman, co-founder and CEO of AI company LTX, said Maxwell doubled its model inference speed, and that halving GPU time translates directly into improved service costs and profit margins.
doubleAI has raised $220 million in total funding, with investors including Nvidia, Dell, Lightspeed Venture Partners and Bessemer Venture Partners. Shashua said demand for compute is already growing faster than hardware can be built; according to SemiAnalysis data, the hourly rental cost of a one-year contract for Nvidia H100 chips has risen 60% over the past year. The company has used about half of its funding to buy 1,000 Nvidia B300 chips for its own use, and Maxwell charges a percentage of the compute costs it saves customers.