Market Drama Sparked by DeepSeek Rattles Nerves Even as it Shows AI's Potential
Plus, Series A funding gap looking like a chasm
The Main Item
DeepSeek Spurs Lively Debate About AI Innovation, ‘Distillation,’ VC Valuations & China
This week marked a coming-of-age moment for the AI industry: its very own market panic.
Nvidia on Monday saw its value drop by nearly $600 billion on the news that China’s DeepSeek had apparently built a competitive LLM without giant data centers full of cutting-edge GPUs.
That’s more than what the biggest of big-tech companies were worth less than a decade ago. Broadcom and AMD plummeted too, along with energy and infrastructure companies.
Suddenly, it seemed that VCs who’d pumped billions into Open AI and the other foundation model companies — we’re looking at you, Thrive, Sequoia, and a16z — might have made an historically bad bet.
The panic eased as the week went on, and then Thursday came a Wall Street Journal report that OpenAI was raising new funds at almost double its current valuation. The news was perfectly timed to blunt the doomsday predictions about the big American LLM companies. Whether the deal closes at that price will be a litmus test of its own.
Meanwhile, though, Microsoft and Meta, in reporting earnings this week, showed no hints of a pullback in AI spending. By midday Friday, Nvidia and the Nasdaq had regained all their losses and then some.
There’s still a lot of disagreement as to whether DeepSeek, allegedly trained for a fraction of what U.S. companies are spending, is a game-changer or more of an incremental advance, though rivals acknowledged that the achievement is real.
DeepSeek shows that generative AI still has a lot of juice, despite the chatter about “scaling walls” and worries about spending and margins. Models are still improving. A bounty of agents and applications are in the pipeline.
Benchmark’s Eric Vishria called foundation models “the fastest depreciating asset in history,” and DeepSeek does put a finger on the question of how defensible a better proprietary LLM is on its own.
“That market is already fiercely competitive and slowing down or missing a step already means they are in trouble,” Madrona’s Jon Turow told us over text. Even if LLMs don’t end up being profit machines, though, there are still massive businesses to be built.
The “Jevons Paradox,” where the declining cost of a commodity creates a huge surge of new demand, is having a moment in the sun.
The DeepSeek surprise is also a reminder that China is a serious global competitor. Yet its tech community — as opposed to the CCP — remains a friend of Silicon Valley in many ways.
We liked this Bill Gurley line about the DeepSeek team: “Why are they ‘enemies?’ No one that works at DeepSeek is an enemy of mine. The fact that they love open-source makes me think I would gel with them quite well.”
Others were more hostile to the Chinese startup’s triumph. VC David Sacks in his new capacity as President Trump’s AI czar said “they cheated.”
Sacks and others speculate that DeepSeek may have used a process known as distillation, or learning from a different LLM’s answers, to develop its tech so quickly. Joshua Kushner, a major OpenAI shareholder, also accused DeepSeek of training off of leading US models.
There’s also a lot of skepticism of the claim that the model was trained for just $6 million on a small cluster of Nvidia’s less-advanced chips. Scale AI’s Alexandr Wang and others say DeepSeek in fact had access to a cluster of 50,000 Nvidia H100s, raising the issue of whether and how DeepSeek had dodged export controls that bar Chinese companies from acquiring the most advanced chips.
The analysts at SemiAnalysis estimate that DeepSeek’s full hardware spend is much higher than $500 million, and that its hedge fund owner High-Flyer started building out its chip cluster in 2021, well before any export restrictions were in place.
Bloomberg reported Thursday that White House officials are probing whether the company might have obtained Nvidia chips through middlemen in Singapore.
A second category of reactions are from those who say it’s all in a day’s work for the tech industry. Even President Trump himself suggested the advance should be a good motivator for American companies to stay ahead. Marc Andreessen called it a “new Sputnik moment” for U.S. researchers, a panic that can also spark innovation.
Anthropic’s Dario Amodei offered a detailed argument on why DeepSeek, though important, wasn’t quite what it seemed. It’s unsurprising that he’d downplay the threat, but his logic is worth a look. “DeepSeek-V3 is not a unique breakthrough or something that fundamentally changes the economics of LLMs; it’s an expected point on an ongoing cost reduction curve,” Amodei wrote.
DeepSeek underscores that technical leadership in LLMs is only one piece of the puzzle when it comes to value-creation in AI.
OpenAI has what most of its rivals — and upstarts like DeepSeek — do not: lots and lots of customers. ChatGPT hit 300 million weekly users in December, Sam Altman said onstage at the DealBook summit. Anthropic is nowhere near that number, but it’s been popular with business customers, and expects its “Computer Use” agent to provide another boost.
That said, DeepSeek is currently number one in the Apple app store.
Cheaper-than-expected LLMs could shift some of the overall competitive dynamics. Basis Set’s Lan Xuezhao said startups building in the inference layer will be the quickest winners. Industries with specialized data needs — think healthtech and biotech startups using AI for drug discovery, or developers of foundation models for robotics, for example — will have a big appetite for AI products, he added.
“All SaaS apps based on LLMs will benefit,” Gradient Ventures’ Darian Shirazi texted me. He argued that the companies with the most to gain are those that help model producers find compute efficiency through profiling and optimization, like CentML or Unsloth.
“People make this too much of a big deal on US and China,” said Xuezhao. “They told you the recipe, just take it and make your own dish.”
Podcast Highlight
DeepSeek Proves the Power of Open Source
In our latest episode, we make the case that DeepSeek is a net positive for the AI community at home — precisely because of its open source nature.
We’re cosigning Bill Gurley’s take that Silicon Valley and China’s tech communities are actually pretty in-sync here.