Chinese AI startup DeepSeek stunned investors when it released a high-performing chatbot that can rival those built by some of the world’s biggest tech companies –– seemingly at a fraction of the cost.
The Hangzhou-based company said its R1 large language model (LLM) cost less than $6m to develop, a tiny sum compared to the hundreds of millions spent by the likes of OpenAI and Google.
Speculation over the AI bubble bursting saw panicked investors selling stock at pace on Monday. By the end of the day, shares in computer chip giant Nvidia had plunged 18% –– wiping out almost $600bn in market value –– while Amazon and Microsoft shares had slipped around 5%.
But where the value of public companies rises and falls for everyone to see, VCs invested in younger startups have the luxury of reassessing the value of their investment behind closed doors –– at least for a while.
Application pending
While those running AI startups in Europe say they remain optimistic, some investors are clear who the winners and losers of DeepSeek’s debut are.
“Founders building at the application level have just been handed a way to achieve good performance at a significantly lower cost,” says Nathan Benaich, founder and general partner at Air Street Capital, an AI-focused VC in London.
“The only real losers are European companies building general-purpose LLMs. These companies have already been struggling to differentiate themselves on price or performance, so R1 presents a perfect storm for them.”
Prof Alex Ilic, executive director and co-founder of the ETH AI Center in Zurich, agrees the application layer is “heavily underserved”.
‘This is resetting the game and presents a major challenge.’
“The infrastructure layer is more commoditised. There is no real moat and DeepSeek is proof of this,” he says. “There’s a ton of opportunities [in the applications space] where you can really build up a moat.”
Michael Tefula, principal and head of product at Ada Ventures, agrees DeepSeek has left European startups with a “mixed picture”.
“This open-source model presents a huge opportunity for startups at the application layer, so they can now experiment with the power of LLMs much more cheaply,” he says.
“But for startups that have raised large amounts of capital to train foundational models from scratch, this is resetting the game and presents a major challenge. It’s early days and nothing is yet definitive, but foundational model companies will be most unsettled about the DeepSeek release.”
Downturn
Adam Shuaib, a partner at London-based early-stage investor Episode 1, tells Sifted the arrival of DeepSeek has demonstrated how quickly an AI company with limited resources can ship a world-leading product. He says: “DeepSeek is the first domino in a series that will ultimately show LLM costs falling to zero in the short-to-medium term.”
While leading LLM-makers like Anthropic and OpenAI will struggle to keep charging more for incremental improvements, Shuaib says smaller providers will struggle to survive at all. “Many VCs with foundation model bets will have to cope with enormous write-offs.”
DeepSeek caught investors off-guard when it released R1 last week. The model was developed using a stockpile of old Nvidia chips, primarily because US sanctions have limited Chinese companies’ ability to access newer technologies.
‘The only real losers are European companies building general-purpose LLMs.’
The cut-price LLM was made public just days after US president Donald Trump announced “Project Stargate”, a $500bn initiative that counts OpenAI and Oracle among its stakeholders.
R1’s viral success has left many questioning why American and European companies needed so much money to build models with seemingly similar capabilities. Early last year, OpenAI CEO Sam Altman was reportedly canvassing investors for more than $5tn of capital.
“As an investor, I’m taking into consideration the timing of the announcement, less than a week after Stargate and the $500bn figure. This speaks to the role the geopolitical environment will have on the latest AI advancements,” says Johnathan Matlock, partner at Bristol-based VC Empirical Ventures, an early-stage deeptech investor.
“Startups and investors may start prioritising ‘efficient AI’ over the ‘scale-at-all-costs’ paradigm,” Matlock says.
Read the orginal article: https://sifted.eu/articles/deepseek-markets-ai-tech-investments/
