DeepSeek’s new open source AI reasoning model, R1, sparked a sell-off of Nvidia’s stock and caused its consumer app to soar to the top of the app stores.
Last month DeepSeek said it trained a model using a data center of some 2,000 of Nvidia’s H800 GPUs in just about two months at a cost of around $5.5 million. Last week, it published a paper showing that its latest model’s performance matched the most advanced reasoning models in the world. These models are being trained in data centers that are spending billions on Nvidia’s faster, very pricey AI chips.
The reaction across the tech industry to DeepSeek’s high-performance, lower-cost model has been wild. Pat Gelsinger, for instance, took to X with glee, posting, “Thank you DeepSeek team.”
Gelsinger is, of course, the recently former CEO of Intel, a hardware engineer, and current chairman of his own IPO-bound startup, Gloo, a messaging and engagement platform for churches. He left Intel in December after four years and an attempt at chasing Nvidia with Intel’s alternative AI GPUs, the Gaudi 3 AI.
Gelsinger wrote that DeepSeek should remind the tech industry of its three most important lessons: lower costs mean wider-spread adoption; ingenuity flourishes under constraints; and “open wins. DeepSeek will help reset the increasingly closed world of foundational AI model work,” he wrote. OpenAI and Anthropic are both closed source.
Gelsinger told TechCrunch that R1 is so impressive, Gloo has already decided not to adopt and pay for OpenAI. Gloo is building an AI service called Kallm, which will offer a chatbot and other services.
“My glue engineers are running R1 today,” he said. “They could’ve run o1 — well, they can only access o1, through the APIs.”
Instead, in two weeks, Gloo expects to have rebuilt Kallm from scratch “with our own foundational model that’s all open source,” he said. “That’s exciting.”
He said he thinks DeepSeek will make AI so affordable, AI won’t just be everywhere. Good AI will be everywhere. “I want better AI in my Oura Ring. I want better AI in my hearing aid. I want more AI in my phone. I want better AI in my embedded devices, like the voice recognition in my EV,” he says.
Gelsinger’s happy reaction was perhaps at odds with others who were less thrilled that reasoning foundational models now have a higher-performing and far more affordable challenger. AI has been growing more expensive, not less.
Other reacted by implying DeepSeek must have fudged its numbers somehow and training must have been more costly. Some thought it couldn’t say it used higher-end chips because of U.S. AI chip export restrictions to China. Others were poking holes in its performance, finding spots where other models did better. Still others believe that OpenAI’s next model, o3, will so outpace R1 when it is released that the status quo will be repaired.
Gelsinger shrugs all of that off. “You will never have full transparency, given most of the work was done in China,” he said. “But still, all evidence is that it’s 10-50x cheaper in their training than o1.”
DeepSeek proves that AI can be moved forward “by engineering creativity, not throwing more hardware power and compute resources at the problem. So that’s thrilling,” he said.
As for this being a Chinese developer with all that implies, like concerns over privacy and censorship, Gelsinger metaphorically shakes his head.
“Having the Chinese remind us of the power of open ecosystems is maybe a touch embarrassing for our community, for the Western world,” he said.
TechCrunch has an AI-focused newsletter! Sign up here to get it in your inbox every Wednesday.
