The Fallout From DeepSeek

Hello, and happy Sunday! After a week in which DeepSeek’s R1 model shocked the tech world, Alex Duffy covers the response from U.S. incumbents and startups, and Behance founder Scott Belsky—who’s leaving tech for Hollywood—updates us about how he uses AI. Scroll down to read that and everything we published last week.

ICYMI: More than 70 students—including vice presidents at public tech companies, marketing heads at asset managers, venture capitalists, and engineers—have already enrolled for our class How to Write With AI. Taught by Evan Armstrong, the next cohort starts on February 13. Learn more about the course and reserve your spot:

Finally, we’re thrilled to welcome Michael Reilly to Every as our new managing editor. Michael joins us from, most recently, The Markup, and previously was at Protocol, MIT Technology Review, and New Scientist.—Kate Lee

Release notes 

DeepSeek's big week: A wake-up call to stop waiting for OpenAI

A couple of months ago, almost nobody outside of some machine learning researchers had heard of DeepSeek. This past week, its app surged to the number-one spot in the App Store, headlines declared the startup was responsible for wiping out over a $1 trillion in stock market value, big tech was in a panic, and many—including OpenAI CEO Sam Altman and even President Donald Trump felt obliged to respond. 

But the story behind the hyperventilating is more nuanced—and more interesting. Yes, DeepSeek’s R1 model is impressively cost-effective and almost on par with some of the best large language models around. Yes, markets reacted, with Nvidia’s stock diving 17 percent at one point. But many of the most educated voices were quick to point out that it's unlikely the demand for Nvidia chips will decline any time soon, and the chip maker’s price has since recovered somewhat. There was also a more profound takeaway: R1 is a wake-up call. It proves that advanced AI needn’t only come from the biggest, most well-funded companies, and that smaller teams can push the envelope instead of waiting around for GPT-5.

Let’s break down what happened, why so many reacted the way they did, and how major players like Microsoft have seized the moment to push forward their own agendas.

The frenzy: Stocks, narratives, and why everyone freaked out

DeepSeek’s R1 was released on January 20 to the excitement of researchers in the machine learning community. It did not come as a surprise  as DeepSeek has been openly putting out superior models and research for most of the past year, but this time there were a few key differences. The model was much better in practice, significantly cheaper, and had no rate limits— developers could make requests to R1 as often as they liked with no restrictions (OpenAI and Anthropic, meanwhile, have been struggling to meet high demands). The release also coincided with inauguration day in the U.S., which might explain why it otherwise flew under the radar until the New York Times published an article three days later. Other mainstream U.S. media outlets soon followed, largely latching onto a single storyline about the threat to U.S. dominance.

Yet, as we’ve seen repeatedly in AI, big claims about “killing GPU demand” rarely hold up. Both Andrej Karpathy and Yann LeCun, two of the most influential AI researchers in the world, argued that massive compute is still essential. Once we have incredible AI, we'll need to serve it to billions of people daily. Hence, more compute spending, not less.

Meanwhile, as news of R1’s impressive performance and price point (about 96 percent cheaper than OpenAI’s o1 model) spread, AI leaders were compelled to respond.

The CEO perspective

OpenAI’s Altman rarely comments directly on competing models, so it was noteworthy that he weighed in. He called R1 “impressive for its price,” gave "credit to R1" for updating OpenAI’s views on thinking tokens, discussed open-source strategy, and promised that OpenAI’s next releases would be “pulled up” (i.e., done sooner) to show just how crucial bigger budgets and “more compute” still are. As if on cue, OpenAI announced the release of its new model, o3-mini, Friday afternoon—a cheaper, better reasoning model positioned to directly compete with, and even outperform, R1. However, o3-mini does cost almost twice as much as R1 per word generated. 

Open AI also accused DeepSeek of improperly harvesting its data, which was met by a large chorus of published authors, internet creators, and social media users reminding the company that it did the same thing.

Anthropic CEO Dario Amodei went further. He published a 2,700-word blog post arguing that while R1’s cost savings fit into a known trend and shouldn’t come as a surprise, it may yet be a matter of national security:

"If the U.S. and China were at parity in AI systems, it seems likely that China could direct more talent, capital, and focus to military applications of the technology. Combined with its large industrial base and military-strategic advantages, this could help China take a commanding lead on the global stage, not just for AI but for everything."

He emphasized the importance of export controls, saying that if China can’t secure millions of high-end chips under new U.S. rules, it may slow the country’s progress in the race to build truly transformative AI. While Amodei’s argument makes sense, one reason he may have written such a strong reaction is that R1 poses direct competition for Anthropic. The company hasn’t built many consumer products on top of its homegrown AI model, Claude, and instead relies primarily on selling direct access to its model via API for other businesses to build with. (Claude has a similar performance to R1, but is much more expensive to run.)

However, there was one notable large language model provider that was clearly prepared.

Google’s quiet 'state-of-the-art' drop