Years of feverish hype around artificial intelligence technology have convinced many that it’s Silicon Valley‘s next speculative bubble — and prompted questions of how long giants like OpenAI can keep burning through billions of dollars in their quest for a true breakthrough AI. Now, a Chinese company has unveiled a cutting-edge AI model that it says it developed in under two months, with end-stage training costs of less than $6 million, figures that significantly undercut the levels of investment from U.S. firms in bots such as ChatGPT and the infrastructure to run them. And the market may never be the same.
DeepSeek — the name of both the lab and its model — emerged as a side project of Liang Wenfeng, co-founder of the hedge fund High-Flyer, who began importing processing chips from Nvidia in 2021 for the project. (The California-based manufacturer is a major supplier of the high-performance graphic processing units, or GPUs, hardware that provides the tremendous computing power necessary to train and run AI.) In 2023, High-Flyer launched a version of DeepSeek as an internal tool to help identify and predict market trends to improve its trading decisions.
But last week, the company released an “AI assistant” bot, DeepSeek-V3, a large language model that has since become the most-downloaded free app on Apple devices (ahead of OpenAI’s ChatGPT), and a reasoning model, DeepSeek-R1, that it claims hits the same benchmarks as OpenAI’s comparable model. The lab has also debuted a multimodal model capable of generating images that appears to outperform competitors DALL-E 3 and Stable Diffusion.
These various upstarts alone might have sent ripples through venture capital firms and major tech players that have bet billions on AI, including Microsoft, Meta, Google parent Alphabet, Amazon, and Nvidia. What really shook these investors on Monday, however, was the efficiency touted by DeepSeek: it reportedly uses a limited number of reduced-capacity chips from Nvidia, in turn substantially lowering operating costs and the price of premium models for consumers. By contrast, OpenAI CEO Sam Altman acknowledged just weeks ago that the company loses money even on pro subscriptions that cost $200 a month, thanks to the astronomical cost of the processing power their software requires. With an alleged price tag of around $5.5 million for its final phase of development, DeepSeek-V3 also represents a relatively cheap alternative to models that have cost tens of millions to engineer. On top of all that, DeepSeek’s codes are actually open-source, freely available for users to distribute and modify, or run on a private device without giving away personal data. U.S. firms, meanwhile, tend to keep the inner workings of their AIs cloaked in as much secrecy as possible.
The upshot of all this was a sudden loss of faith in industry leaders, including several who are collaborating on a $500 billion project to expand AI infrastructure under President Trump, known as the Stargate Initiative. The chipmaker Nvidia was hardest hit, losing $600 billion in market capitalization as its share price plummeted 17 percent — the largest single-day drop for a U.S. corporation in history. This sell-off indicated a sense that the next wave of AI models may not require the tens of thousands of top-end GPUs that Silicon Valley behemoths have amassed into computing superclusters for the purposes of accelerating their AI innovation. Sharply reduced demand for chips and massive data centers like those Trump has proposed under Stargate (in an announcement that propelled AI stocks higher just days ago) could entirely reshape this sector of the economy. By Monday’s market close, the tech-heavy Nasdaq 100 stock market index had shed an incredible $1 trillion from its value on Friday.
Of course, DeepSeek’s big splash also made it a target, and the company limited registration on Monday during what it called “large-scale malicious attacks” on its services (though without limiting access to existing users). At the same time, certain analysts argued that the market panic about the long-term competitiveness of the U.S. in the AI race could be an overreaction. After all, DeepSeek may point the way for increased efficiency in American-made models, some investors will buy in during this dip, and, as a Chinese company, DeepSeek faces some of the same national security concerns that have bedeviled ByteDance, the Chinese owner of TikTok.
On the other hand, it’s hard to ignore the questions that DeepSeek raises about the staggering sums of capital that U.S. tech giants and Silicon Valley VCs have poured into AI with the expectation of revolutionary (and profitable) results. To some observers, it will start to appear that such spending is not only unsustainable but ultimately wasteful, considering how much a foreign startup has accomplished with far less. We haven’t seen the bubble burst just yet, but with this many investors rushing to unload assets that suddenly seem a lot riskier, you can practically hear it deflating.
From Rolling Stone US
