Since OpenAI launched ChatGPT in November 2022, top venture capital firms have raced to throw eye-watering sums of money at artificial intelligence companies, driven by the idea that the development of large language models—the technology underpinning generative AI—would require armies of engineers and war chests of chips.
But that logic was flipped on its head last week when a little-known Chinese startup called DeepSeek unveiled an AI model that rivals cutting-edge ones from the U.S. leaders of OpenAI and Anthropic, but which is much less expensive to train and run. DeepSeek has said that its V3 model—which it unveiled in December and is equivalent to OpenAI’s GPT-4o on many benchmarks—had cost it less than $6 million to train. V3 serves as the precursor model to R1.
U.S. AI startups raised a record $97 billion in 2024, according to PitchBook, with $20 billion going to xAI, OpenAI, and Anthropic in the last three months alone.
While DeepSeek’s claims, and origins, are still murky, its novel approach to lowering the computing costs of AI development is irrefutable. Tech markets shuddered at the breakout success: Nvidia, the leading chip manufacturer for AI companies, shed a whopping $600 billion in a single day of trading.
For the venture capital industry, DeepSeek’s emergence could create a paradigm shift in how firms fund AI startups, with the past two years driven by outsize valuations at a time of constriction for the broader tech sector. Investors told Fortune that DeepSeek could cause VC firms to rethink the approach of “closed-source” companies like OpenAI and Anthropic while spurring growth in companies building applied tools, rather than core infrastructure.
Aaron Jacobson, a partner at NEA working on AI investing, said that DeepSeek likely doesn’t represent an “extinction” moment for firms that plowed billions into its competitors, though it could still be a pivot point for the industry. “It’s a gut check relative to valuations,” he told Fortune.
DeepSeek surprise
Until this week, Sam Altman’s OpenAI has been the clear winner of the artificial intelligence arms race, raking in billions of dollars in funding from top venture firms like Joshua Kushner’s Thrive Capital and building what many experts viewed as an insurmountable moat. That lead was only set to increase with January’s announcement of Stargate, a $500 billion project championed by President Trump to build new infrastructure for OpenAI in partnership with SoftBank and Oracle.
OpenAI has taken a closed-source approach to developing its AI models, meaning the source code and “weights,” which determine how models process information, were not available to the public. The other leading U.S. developer, Anthropic, followed the same path.
DeepSeek, in contrast, is open-source. Though it has not released the data to train its R1 model, launched last week, developers can still access its underlying code and model weights and use or modify them. (The company has also revealed far more about how it trained R1 and the kinds of data that it used, compared to what OpenAI has revealed about its reasoning models, o1 and o3.)
Some investors who spoke with Fortune lauded the approach. Jacobson said that OpenAI’s closed-source tack is one of the reasons that NEA decided not to back the startup, predicting that open-source projects would soon outpace Altman’s outfit—and at lower costs, because there would be broader community input. Umesh Padval, managing director at Thomvest Ventures, echoed the sentiment, telling Fortune that OpenAI may see the strain as open-source models take off.
“Companies like OpenAI, which have raised significant capital at high valuations and rely heavily on consumer revenue, may face pressure as open-source models gain traction,” said Padval, an investor in enterprise AI unicorn Cohere.
That doesn’t mean the competition will come from DeepSeek itself. Lingering questions about its actual total development costs, as well as security concerns about its ties to China, mean that DeepSeek’s tech might outlast the company itself. In a rare post on X, Thrive Capital’s Kushner criticized “pro America” technologists for “openly supporting a Chinese model that was trained off of leading U.S. frontier models.”
A representative for Thrive did not respond to a request for comment.
DeepSeek’s biggest impact could be reducing the current froth of AI-related public and private stocks. “Even unrelated to DeepSeek, most U.S. tech companies have raised too much money and could be more efficient,” Asymmetric Capital Partners managing partner Rob Biederman told Fortune.
The big surprise for DeepSeek is not its performance, but its reported cost. After reviewing technical reports, Jenny Xiao, a Leonis Capital partner and former OpenAI researcher, is more inclined than some to believe DeepSeek’s reported sub-$6 million training costs are potentially viable. But regardless of what’s true for DeepSeek, its success demonstrates the different reality for AI development in China.
For American tech giants, “a lot of their GPUs are just sitting around,” Xiao said. “And in the U.S., because these companies are so well funded, they don’t really need to think about GPU utilization.”
Samir Kumar, a Touring Capital general partner, argued that export restrictions on China require developers like DeepSeek to squeeze performance out of older GPUs. “This should catalyze the whole ecosystem to focus on the efficiency of AI,” he told Fortune.
The application layer
Despite Monday’s tech stock selloff, many investors expressed optimism that DeepSeek’s new technology could further development for AI startups, with reduced costs for running LLMs for less-funded outfits.
Much of the past two years’ glut in funding has targeted infrastructure companies like OpenAI, while “application layer” startups—building tools that utilize AI—have yet to reach their own breakthrough moment. “In the application layer, DeepSeek’s breakthroughs will only accelerate the rate at which AI will get deployed into consumer and enterprise products,” said Michael Mignano, a partner at Lightspeed, which has invested in Anthropic, though Mignano declined to discuss his firm’s involvement.
Dzung Nguyen, a managing director of technology banking at Wells Fargo, told Fortune that DeepSeek’s launch has driven an influx of calls from startups asking about the impact of the new model, especially for early-stage companies. “How can we leverage that to scale ourselves?” she said. “We’re putting our ears to the ground.”
Others have cited a phenomenon known as the Jevons Paradox, which states that as a resource becomes more efficient, overall consumption tends to increase. Pradeep Tagare, the head of investments at the utility company National Grid’s corporate venture arm, told Fortune that he was “frankly encouraged” by DeepSeek’s claims of performing AI workloads using a fraction of computing resources. He added this could help address concerns over the electricity demand spurred by AI usage, with customers now able to run models on their own infrastructure. Gabriel Kra, managing director at Prelude Ventures, agreed, telling Fortune that if models take substantially less electricity to train, that could have a “dramatic impact” moving forward.
NEA’s Jacobson said that DeepSeek has created a “checkpoint” for AI, likely advancing the open-source community and throwing OpenAI’s lead into doubt. “It’s never been more exciting to be an investor,” he told Fortune.
