This article first appeared in Sifted’s Up Round newsletter, sign up here.
DeepSeek, the Chinese AI startup that launched an AI model comparable to ChatGPT for allegedly far less money, has shaken the tech world — and led to chipmaker Nvidia’s market cap being slashed by nearly $600bn earlier this week. It’s got me (and seemingly everyone else) wondering: has the AI game changed for VCs?
The answer, annoyingly, seems to be yes and no.
What has changed is that the throw-money-at-the-AI-model approach looks to be on its way out. “I think capital was seen as a barrier to entry and a competitive advantage somehow; people were like, ‘Oh, we need to do these big checks… if you cannot raise money, you will need to use other people’s models and this sort of weakens your proposition.’ This seems to be wrong now,” Francesco Ricciuti, an associate at deeptech and AI VC Runa Capital, tells me. “Probably, if these people look back at their investment memo, they will see that some of their hot takes were wrong.”
Some are suggesting that this will be an “extinction-level event” for VCs — and while the jury is still out on that front, those who may be nervous include investors in French OpenAI challenger and LLM maker Mistral, which raised a whopping €600m at a €5.8bn valuation last June. Some of its backers are privately less bullish, according to a new FT report. (I reached out to a couple Mistral investors who either declined or didn’t respond in time to comment.)
I wager that the DeepSeek news is making lots of VCs rethink the price tags they paid for getting an “AI company” in their portfolio — and may prompt more conversations about the traction these startups are actually getting.
“It is for sure an area where a bit of money will be lost,” Oliver Schoppe, a principal at Munich-based UVC Partners and an investor in German AI company Aleph Alpha, tells me of the LLM space in particular. “There are VCs out there — I don’t want to go into naming and shaming — that have put a lot of money into this stuff with a return expectation from building really expensive foundational models. This may be questionable.” (Aleph Alpha, UVC’s portfolio company, pivoted away from developing models and into building more of an operating system last year.)
But investors clearly aren’t throwing in the towel on big LLMs just yet. Japanese investment titan SoftBank is this week reportedly in talks with OpenAI to invest up to $25bn into the company, according to the FT. (Woof.)
So, what hasn’t changed?
The industry had already largely moved away from foundational models; the potential winners have likely been established, many believe, and VCs have instead been turning to more vertical AI applications or agents.
Schoppe goes a step further in his predictions: DeepSeek’s open-source model “enables so many companies right now to start or to do something different; it becomes more accessible. So I would rather expect another small wave of more innovation and more startups around that.”
He also doesn’t expect valuations to suddenly plummet just because of DeepSeek: for early-stage AI startups, at least, high valuations are more a matter of supply and demand, he says. He also points out the distinction between the Mistrals and OpenAIs of the world and startups building on the AI application layer.
Valuations are a mixed bag at the moment, as it’s hard to predict how the funding needs and competitive dynamics of different AI startups — like agents or application-layer startups — will develop, says Andreas Goeldi, a partner at B2venture, which counts AI language platform DeepL among its portfolio. “I think everybody is currently operating from their own assumptions that, of course, keep changing.”
Two schools of thought
Goeldi believes there are two conflicting schools of thought with VCs right now: one group thinks DeepSeek is “just a minor bump in the road” amid the unprecedented tech shift happening with AI, and that the best-funded labs and companies are still “best positioned to harvest most of the value”. “I’m not sure that group is going to change its stance anytime soon unless something much more dramatic than DeepSeek happens,” he says.
Another cohort believes current AI valuations are far too high and need to better reflect that AI development and deployment will “take much longer than we think,” Goeldi says. “I think this group feels vindicated by DeepSeek because it shows that even the best-financed companies don’t have much of a long-term moat,” he says.
Several VCs believe DeepSeek will prove a good thing — for certain companies. Schoppe says that since running AI is so expensive, application-layer startups’ gross margins are closer to 70% compared to the traditional gross margins that VCs look for — of 90% — in software companies. DeepSeek’s cheaper model “has huge implications for the capital intensity” and valuations of those startups. “I would argue that for many application layer companies, unit economics only now start to make sense,” he says. “It’s actually a good thing from an investor’s perspective.”
Whether or not the latest AI drama is good or bad for their individual portfolios, the DeepSeek episode is a “sobering” moment for VCs.
“You can’t ignore it anymore: the development cycles are so fast that whatever is leading right now is a commodity in a year or a couple months,” Schoppe says. Differentiation, then, may be through more commercial advantages, he believes, like how well startups understand their customers, their access and distribution or network effects.
VCs, what do you make of this? Are you revisiting your AI theses? Do you think AI valuations will start cooling off a bit? Where do you see opportunities in the market now? Are you worried? Send me a line.
Read the orginal article: https://sifted.eu/articles/what-deepseek-means-vcs/
