Ex-Intel CEO Pat Gelsinger's startup chose China's DeepSeek instead of OpenAI

DeepSeek R1, a Chinese open-source LLM, is making waves in the tech landscape for its superior performance compared to OpenAI's best while requiring substantially fewer computing and training resources. Former Intel CEO Pat Gelsinger took to X and praised the model with compliments. In a chat with TechCrunch, he revealed that his startup will use DeepSeek over OpenAI.

If you haven't been keeping up with industry news, Pat Gelsinger stepped down as CEO of Intel last month and was replaced by two interim CEOs. Speculation suggests that Intel's disappointing stock market results contributed to Gelsinger's dismissal or ousting by the board. Nonetheless, Gelsinger now serves as chairman of his new startup, Gloo, which is reportedly a messaging platform for Churches.

Gloo is developing an AI service dubbed "Kallm," described briefly as an AI-powered chatbot. After comparing DeepSeek's R1 vs OpenAI's o1 model, the former was a better fit for Gloo, being open-source and likely easier to integrate. "My Gloo engineers are running R1 today, they could’ve run o1 — well, they can only access o1, through the APIs," said Gelsinger. Expanding the tweet below, you'll see that Gelsinger also shared some praise at X, thanking DeepSeek for bringing affordable AI and driving competition.

DeepSeek used Nvidia's H800 GPUs to train its R1 model but sticks with (mostly) homegrown Huawei Ascend AI accelerators (likely the Huawei 910C) for inferencing to save on costs and reduce dependence on Western hardware. Sam Atlman famously declared that AI startups with $10 million are "totally hopeless", however, DeepSeek claims its total training expenditure was just $5.6 million while requiring 11x less compute than Meta's Llama 3 405b model.

Industry experts have contested these numbers, but Gelsinger asserts, "You will never have full transparency, given most of the work was done in China. But still, all evidence is that it’s 10-50x cheaper in their training than o1." The focal point remained on how DeepSeek is pushing the industry to go open-source and find creative solutions rather than throwing hardware at the problem.

DeepSeek is under investigation by Microsoft and OpenAI for allegedly using data from ChatGPT illicitly in distillation. Likewise, DeepSeek collects hordes of user data and stores it in servers based in China, though at least they are transparent about their practices.