DeepSeek, a Chinese upstart competitor to foundational model companies like OpenAI and Anthropic, has made waves across the AI landscape with its new open-source chatbot.
The speed and sophistication DeepSeek achieved with its model was done reportedly at a fraction of rivals’ costs, raising questions about the high valuations and capital requirements of these companies.
Caught off-guard, several investors with big positions in large language model startups are expressing concern. Thrive Capital‘s Josh Kushner posted on X, calling DeepSeek “a Chinese model that was trained off of leading US frontier models,” while Coatue founder Philippe Laffont questioned whether AI models should even be allowed to be open source.
They have good reason to be worried. VCs have spent the last two years plowing piles of cash into large language models. OpenAI’s most recent $6.6 billion financing brought its valuation beyond that of any VC-backed tech company in history at their time of IPOs. In 2024, VC funding for foundational model startups doubled over the previous year with a record $40 billion invested.
Meanwhile, DeepSeek’s advanced reasoning model, R1, catapulted it to the top of Apple’s app store and wiped more than $600 billion off Nvidia‘s market cap Monday.
“This is going to put a massive amount of pressure on the pricing of OpenAI,” said Sierra Ventures managing partner Tim Guleri.
DeepSeek’s success could significantly undermine the assumptions of OpenAI’s current valuation—along with those of competitors like Anthropic, Cohere and Mistral—if their technological moat erodes.
“When you have a $160 billion valuation company like OpenAI losing $5 billion at $3.7 billion in revenue, how do you make money? I can’t see a path,” says Umesh Padval, managing director at Cohere-backer Thomvest Ventures. “This is all a wake-up call to these astronomical term sheets of companies that don’t make sense.”
A notable aspect of DeepSeek’s rise is how little the company spent to train its models, challenging the assumption that large language models require millions upon billions to develop. The company has said its latest model cost $6 million to train—just a fraction of the $100 million OpenAI spent to train GPT-4.
Jesse Middleton, general partner at Flybridge, which includes small language model developer Arcee.ai in its portfolio, notes that DeepSeek’s rise shows there is still opportunity in the foundational layer segment.
“Just 12 to 18 months ago, it seemed that creating any new ‘foundational’ model would require billions of dollars,” he said. “It certainly changes the idea that, without billions of dollars, no one can compete in this market.”
An important question remains: How does the company make inroads with enterprise customers? Tribe Capital managing partner Boris Revsin said it’s too early to ascertain how DeepSeek will be able to compete in the space.
“OpenAI, xA and their competitors still have the enterprise sales team; they still have the enterprise connections,” he said. “I’m bullish.”
Jordan Rubio contributed to this article.
Featured image by VCG/Getty Images
Learn more about our editorial standards.
