China’s AI star shocks US tech sector

The heart of the American AI boom is full of hipster chic. Hip-hop blares across the central park and sidewalks of San Francisco’s Hayes Valley, where men and women are lifting weights in an open-air gym. Visitors to a juice bar next door sip celery juice and algae concentrate, while luxury boutiques on the other side of the street sell baggy sweaters for $170.

For the last few years, this neighborhood has been nicknamed «Cerebral Valley.» Today, artificial intelligence is omnipresent in America’s technology mecca, but Hayes Valley in particular has become a kind of capital of the AI revolution. New AI startups do their best to establish an address here. Their founders live in «hacker houses,» and community members gather every evening at AI meetups, hackathons and demo days.

Certainties are crumbling in Silicon Valley

But America’s tech sector has just been shown up by an unexpected rival. The latest and greatest chatbot no longer comes from California. Surprisingly for many, a little-known startup from China has unveiled a model that costs a fraction of what OpenAI, Anthropic and other Silicon Valley AI companies are spending to develop their models. Overnight, DeepSeek has made its way onto millions of smartphones worldwide – a meteoric rise of the kind previously shown only by ChatGPT.

The sector’s core beliefs are crumbling: for example, that larger AI models are fundamentally better. That the development of such a model necessarily costs between $100 million and $1 billion. That developers absolutely need the latest high-performance chips from Nvidia. And perhaps most painfully, that America’s tech companies are leading the AI revolution.

What do people here in the «Cerebral Valley» think about this? «DeepSeek’s new model is really good,» says a young man who is working in a coffee shop opposite the neighborhood’s central park – along with about two dozen other millennials this afternoon. The 36-year-old declines to give his name, as he says he works as an app developer for an AI startup. His laptop computer is decorated with a sticker from Anthropic, developer of the Claude chatbot.

The developer says that DeepSeek had been on his radar for the last three to five months. «Everyone around here was aware of it,» he notes. Nevertheless, he says he was surprised at just how good DeepSeek’s latest model has turned out to be. The quality of the results and the low development costs are remarkable, he adds. Will that change the work being done here? He nods. «The pressure on companies will increase enormously. But so will the level of government funding,» he predicts. The global competition for AI leadership between China and the United States will now take off in earnest, he adds.

Outwardly calm, but setting up «war rooms»

Big Tech companies in particular must now ask themselves a series of difficult questions. The Chinese firm’s success has exposed them to the charge that they are acting inefficiently and wasting capital. The quarterly figures reported on Wednesday by Meta and Microsoft were overshadowed by the DeepSeek news.

DeepSeek’s advance seems particularly bizarre at a time when America’s tech companies have just announced billions of dollars of investment in AI. For example, OpenAI, Oracle and SoftBank have said they intend to spend half a trillion dollars on AI infrastructure as part of the recently unveiled Stargate project.

Meta has also just announced that it intends to spend $65 billion on AI this year alone. The company is building a new data center «so large it would cover a significant part of Manhattan,» said CEO Mark Zuckerberg last Friday. Investors may now be wondering whether it might not instead be enough to rent a garage in Manhattan to build the world’s best AI model.

Panic now seems to have broken out at Meta. According to The Information, a tech industry publication, the company has set up four special «war rooms» for AI researchers to dissect DeepSeek’s algorithms. The company is scrambling to figure out how it might make its own Llama model cheaper and more efficient, the report said. The next version of Llama is due to be released this quarter.

Microsoft, which has invested around $14 billion in OpenAI, is making an effort to remain demonstratively calm. «Jevons paradox strikes again!» wrote CEO Satya Nadella on X, the platform formerly known as Twitter – that is, if a resource can be produced more efficiently, demand for it will rise still further over the long term. The fact that Nadella made his post late on a Sunday evening might indicate that he is not entirely relaxed about the new development. Given the panicked selling by stock-market investors, it was exactly the answer that might be expected from Microsoft’s CEO, opined one New York Times commentator.

Startups already switching to DeepSeek

Many startups, on the other hand, are acting pragmatically. One fintech founder told the Wall Street Journal that he had already switched from Anthropic to DeepSeek last fall. Tests had shown that the results of the two models were comparably good, but the costs of using DeepSeek were only a quarter of those associated with Anthropic, the founder said.

Founder Richard Socher has also quickly integrated DeepSeek into his California-based startup You.com, which offers AI assistants based on several models. «It’s a better approach to remain flexible anyway,» the native of Germany told the NZZ.

Socher, who is himself a leading AI researcher, is critical of DeepSeek’s success, however. «China has been taking ideas from the West for years, and scaling them up at a lower cost,» he says. Other American experts echo this point. DeepSeek’s success in this form was possible only because the startup was able to build on research in which U.S. firms had invested hundreds of billions of dollars, they say,

Socher is also critical of the claim that the model cost only $5.6 million to develop. This figure does not include the acquisition costs for the graphics processing units, or GPUs, required to carry out complex AI tasks, he says. Nor does it incorporate spending for experiments with hyperparameters. These are parameters that are used to control the training algorithm. Unlike other parameters, their value must be set before the model is actually trained.

However, the DeepSeek case does show that progress in algorithms will continue to be important, Socher adds. «The largest GPU cluster won’t necessarily win,» he says.

Venture capitalists are enthusiastic

Some venture capitalists also seem to share this view. Praising the model extensively on X, prominent Silicon Valley investor Marc Andreessen said the release of DeepSeek’s model was a «sputnik moment.»

After the Soviet Union’s launch of the Sputnik 1 satellite in the Cold War, beating the U.S. into space, investment in space travel took off. And in fact, DeepSeek could be good news for the startups in which venture capitalists like Andreessen invest. Thanks to efficiency gains, even small companies with limited budgets could become the next shooting stars of the AI revolution, says Philipp Stauffer. The Switzerland native heads the venture capital firm Fyrfly Venture Partners in Silicon Valley. «That is precisely the promise of artificial intelligence,» he adds.

Zack Kass, a U.S. AI expert and former top executive at OpenAI, made a similar point to the Wall Street Journal, arguing that DeepSeek shows that scarce resources can often inspire creativity.

Pampered founders in «hacker houses»

Kass’ statement is particularly thought-provoking when taking a stroll from the central park area in Hayes Valley across «Cerebral Valley» to the nearby hill of Alamo Square Park. From here, San Francisco looks like something out of a picture book. It was once the setting for the cult TV series «Full House.» Today, many of these Victorian mansions have been converted into hacker houses – hip communal living spaces for AI founders.

One of the best known of these is HF0, the house used for the Hacker Fellowship Zero incubator program. A hundred years ago, the archbishop of San Francisco lived in this magnificent villa. Today, fellowship winners are hosted here for 12 weeks at a time. They are given living space and meals, laundry service, and $250,000 in startup funding. Founders should not have to worry about anything other their work, is the promise to the fellows. In earlier days, some of the best startups used to be founded in garages. The question hanging over the industry now is whether these garages are still in Silicon Valley – or increasingly in China.

Latest articles

Global reporting. Swiss-quality journalism.

In today’s increasingly polarized media market, the Switzerland-based NZZ offers a critical and fact-based outside view. We are not in the breaking-news business. We offer thoughtful, well-researched stories and analyses that go behind the headlines to explain relevant events in the U.S., in Europe and worldwide. To produce this work, the NZZ maintains an industry-leading network of expert reporters around the globe who work closely with our main newsroom in Zurich.

Sign up for our free newsletter or follow us on Twitter, Facebook or WhatsApp.