Last month’s news that a Chinese startup called DeepSeek had developed a large language model for a tiny fraction of the cost of those built by U.S. companies triggered a nearly $1 trillion selloff in the stock market and sent investors scurrying for their calculators.
Did DeepSeek’s innovations permanently alter the economics of artificial intelligence model training? And if so, what impact would that have on venture capital investors who threw more than $100 billion at AI startups in 2024 alone?
As it turns out, the impact may not be so great. Interviews with eight venture capitalists who have stakes in AI companies found that most shrugged off the DeepSeek news as inevitable and see only an upside for startups as training costs decline.
Most believe improving economics will shift action in the venture industry away from building costly models to smaller models deployed for narrow use cases. Some said DeepSeek’s success in squeezing models into smaller space could even ignite new activity in the network’s edge.
“This should have been expected and we shouldn’t have been surprised by it,” said Jim Curry, co-founder and CEO of BuildGroup LLC, which invests in AI-enabled workflow companies. “I expect more surprises.”
Although recent allegations that the Chinese firm fudged the numbers and may have purloined training data have clouded claims that its innovations are a breakthrough, investor assumptions may now change about the true cost of AI model-building.
Those perceptions have been shaped by the pioneers who poured hundreds of millions or even billions of dollars into development. Large models may still require nine-figure investments, but the bigger trend is toward lower-cost use cases involving smaller and open-source language models. That will drive overall growth in AI development and more venture funding.
“There’s a lot of room for efficiency improvements, and we’re going to keep seeing costs falling,” said Javier Rojas, co-founder and managing partner of Savant Growth LLC, a private equity firm that invests in AI-driven software-as-a-service companies. “Lower costs increase return on investment.”
Dan Engel, founder and managing partner of Santa Barbara Venture Partners, sees the overall AI market growing as barriers to entry fall. “When you have such a shock to the system it’s a whole new ballgame for what’s going to be possible,” he said. “The benefit should be massive in terms of what this opens up.”
Race to the bottom
Kevin Surace, a venture capital investor for more than 30 years, thinks the investment dynamics are shifting. “Foundational models are going to be commodities; we’ll see a plethora of them,” he said. He cited the example of OpenAI LLC’s recent entry into the agentic AI market as evidence that even the largest players are anticipating diminishing returns.
Karthee Madasamy, managing partner at MFV Partners, a VC firm focused on AI chips, quantum computing, and robotics, expects to see “more of the foundational models companies offer higher-layer solutions. I think there will now be more demand from VCs in investing in AI startups at the application layer,” he said.
Some venture capitalists drew an analogy to the internet boom of the late 1990s that saw startups with two-page business models raising tens of millions of dollars in funding. Few survived the shakeout.
Steve Brotman, founder and managing partner of Alpha Partners, likens the situation to the massive investments that went into laying fiber-optic cable in the 1990s in anticipation of an internet boom. “Things ended badly where they produced too much fiber and drove the cost to zero,” he said. “Then Google, YouTube and Facebook came along. That’s when consumers won.”
Venture capitalists are already seeing signs of increased startup activity. “I think this makes the investment climate better,” Surace said. “If you start an AI company and the first thing you need is $100 million, that’s crazy. It’s not a model that hunts in the long run.”
Jeep Kline, founder and managing partner of Raisewell Ventures LLC, said aspiring venture capitalists could benefit as well. “This opens up a whole new sector for early-stage VC funds that don’t have $500 million to invest,” she said.
Return to basics
The DeepSeek shocker could reset investors’ expectations for the better. “Scarcity of resources leads to great innovation,” Curry said. “Silicon Valley needs to step back from the firehose of capital and focus on how to build sustainable businesses.”
One thing everyone agrees on is that lower costs will spur competition with dividends for nearly everyone. “The more accessible the models become, the easier it will be to build models and that increases competition in the market,” said Kevin Stevens, a partner at Energize Capital LLC, which invests primarily in sustainable-energy businesses. “We think pace of innovation will accelerate.”
The relatively compact size of DeepSeek’s model could also open opportunities at the edge and in robotics. DeepSeek’s use of multi-head latent attention, a technique for improving efficiency and performance by focusing on the most relevant input features to reduce memory overhead, could be a breakthrough in edge inferencing, Madasamy said.
“This enables data-gathering from edge devices to provide actionable intelligence rather than relying solely on cloud-based servers or data centers,” he said.
The large models “were just way too expensive and their footprint was too big,” said Alpha Partners’ Brotman. “You couldn’t have the latency of going back to the cloud for everything.”
More dollars are also likely to flow to companies building small language models targeted at specific applications. “I think you’ll see a significant rise in value for SaaS [software-as-a-service] companies that have data and workflows that can be optimized,” Rojas said.
He cited the example of Savant Growth’s recent investment in QountHQ Inc., a maker of practice management software for practice management firms. Such domain specific applications require smaller models and less infrastructure to solve problems that businesses are willing to pay for. “We think those companies will explode in value,” he said.
DeepSeek’s decision to release part of its model under an open-source license may ultimately be more disruptive than the model itself, some investors said. “It was a judo move to open-source it,” Curry said. “A lot of people are now playing around with DeepSeek and open-source alternatives. The costs will be cheaper, although the underlying infrastructure will still be expensive.”
How expensive is still to be determined. Despite the excitement about a potential breakthrough in training costs, Raisewell’s Kline pointed out that much is still unknown about the secretive Chinese company. “They’ve shown they can do it faster and cheaper, but there are a lot of problems in the AI model with control, infrastructure and content being filtered,” she said. “I think quality will still require a lot of investment.”
The challenge from across the Pacific should be a rallying cry for U.S. firms, Engel said. “We overlook the potential of other people being able to do the same things we do,” he said. “Apple’s made a business out of not being first but being best.” In the end, quality counts.
