DeepSeek sparks fears that SF’s AI boom is over before it started 

San Francisco’s tech community was forced to contemplate something shocking over the weekend: that this self-obsessed West Coast city might not be the undisputed center of the AI boom.

The freakout came as a chatbot called DeepSeek, the side project of a Chinese hedge fund, on Sunday climbed to the top of the U.S. Apple Store’s ranking of most downloaded free apps. 

DeepSeek is now the talk of Silicon Valley. Venture capitalist Marc Andreessen evoked the USSR blindsiding the West when it launched the first satellite into space in 1957, calling the Chinese model’s release “AI’s Sputnik moment.” Meanwhile, Meta researchers panicked that their next AI model won’t perform as well as DeepSeek and reportedly set up “war rooms” to dissect the model’s code.

The company behind DeepSeek, Hangzhou-based High-Flyer Capital Management, was founded in 2015 — the same year as OpenAI. But the trading firm established its exploratory project in 2023, a year after ChatGPT kicked off the consumer AI boom. 

In less than two years, High-Flyer has debunked the conventional Silicon Valley wisdom that no company can compete with U.S. firms without spending hundreds of millions of dollars on specialized chips. 

It has also thrown into question whether the industry hype wave of San Francisco’s economy as the “AI capital of the world” has legs. The billions in funding that have gone to support homegrown companies like OpenAI and Anthropic have helped support local businesses and uplifted the flagging commercial estate market, functioning as a bright spot for a city with a dearth of good news. 

DeepSeek this month released a version that rivals OpenAI’s flagship “reasoning” model, trained to answer complex questions faster than a human can. The model can be used as an AI assistant, just like ChatGPT. 

And if that isn’t enough to raise a techie’s blood pressure, DeepSeek’s model cost less than $6 million to develop — far less than many Silicon Valley executives make in a year — and was trained on 2,000 Nvidia chips with inferior capabilities to the tens of thousands of cutting-edge chips used by U.S. firms. As a result, its model is cheaper for users, with its most powerful version going for 95% less than OpenAI. 

DeepSeek was effectively forced to do more with less as a result of the U.S. limiting chip exports to China over the past three years. This raises concerns that measures meant to throttle China’s advancements in AI are having the opposite effect — driving technological innovation and efficiency — while U.S. firms fall behind despite billions of dollars in investment. 

The realization has caused a panic that the AI bubble is on the verge of bursting amid a global tech stock sell-off. The Nasdaq fell more than 3% Monday; Nvidia shares plummeted more than 15%, losing more than $500 billion in value, in a record-breaking drop. Shares of OpenAI backer Microsoft fell more than 3%. 

Deedy Das, a former software engineer and AI investor at Menlo Ventures, considers DeepSeek’s achievement a significant breakthrough. 

“What’s more is that it’s completely open-source,” Das said, referring to anyone being able to see the source code. “They were extremely transparent about their technique. They’re giving it away for free, and you can tune it to not abide by China policies … which is really hard to criticize.”

Das said that for Silicon Valley, the biggest impact of DeepSeek’s breakthrough is the realization that “Whoa, the U.S. may not be the epicenter of cutting-edge tech when it comes to training models,” adding, “That hurts the American psyche in a lot of ways.” 

Das said he was impressed by DeepSeek’s research paper explaining the methodology, which included “multiple entirely new paradigms” of doing research, such as performing novel optimizations to Nvidia H800 chips. 

But while it’s been a rattling moment for Silicon Valley, Das said, it’s unclear how significant this will be in the battle between the U.S. and China for AI supremacy. 

“We’re still very much in the thick of the AI race, and things could turn easily,” he noted.

Chris Saad, a serial entrepreneur who runs a business coaching founders on how to adopt “Silicon Valley thinking,” said Chinese firms are good at copying and stealing U.S. intellectual property. 

“However, to stay ahead of the curve and invent real AGI and then superintelligence, they’re gonna need to do a lot better than that,” he said, adding that OpenAI and others are going to have to double down on protecting their intellectual property.

For the growing chorus of people concerned with the environmental impact of generative AI — one ChatGPT query requires nearly 10 times as much energy as a Google search — the fact that DeepSeek’s breakthrough uses significantly less computing power than U.S.-created options is a welcome development. 

Saikat Chaudhuri, a business and engineering professor at UC Berkeley, said the vast amounts of energy and water used for cooling generative AI hardware “will have to be dramatically reduced if we want to have AI deployed everywhere in the revolutionary and ubiquitous ways that the sector envisions.” 

In other words, DeepSeek’s coming out is potentially good news for the tech world — even if it’s bad news for San Francisco’s status at the center of it.