DeepSeek, a hedge fund-slash-tech company, sent shockwaves through the U.S. tech industry and the financial markets by releasing two new large language models (LLMs) whose performance reportedly matches that of OpenAI’s latest version of ChatGPT — and are open-source and free. Even more remarkably, DeepSeek claims to have developed these models using just $6 million worth of computing power and lower-cost AI chips — a tiny fraction of the resources invested by industry giants like OpenAI and Google’s parent company, Alphabet. DeepSeek continues to send shock waves through Silicon Valley, Wall Street, and Washington D.C.
The title of a recent podcast episode from the New York Times captured the consensus and hyperbolic view of DeepSeek: “How DeepSeek Caught Us All Off Guard.” But the reality is that DeepSeek’s January 2025 release should not have caught anybody off guard. DeepSeek has been transparent about its research and ambitions in AI development, publishing its research in a series of more than 10 papers on arVix and GitHub prior to the January paper.
DeepSeek also has been sharing its research on X and YouTube since October 2023. HuggingFace started listing DeepSeek’s research in December 2023. Scores of posts on Reddit around the same time reveal both an awareness and appreciation of DeepSeek, In June 2024, one poster wrote “DeepSeek is one of the best OSS coding models available — I’ve been using their models pretty much since they dropped and there’s very little they can’t do honestly.”
Even a relatively obscure online blog predicted DeepSeek’s technical capabilities in the middle of last year: “DeepSeek AI is a game-changer in the AI industry, offering advanced reasoning capabilities, cost-effective solutions, and open-source models. …As DeepSeek AI continues to innovate, it will play a crucial role in shaping the future of AI.”
Around the same time, a few mainstream industry publications got wind of DeepSeek’s Gen AI chops. Bloomberg’s Catherine Thorbecke wrote, “Separately, Chinese quant firm High-Flyer Capital Management quietly released an open-source AI model earlier this month dubbed DeepSeek Coder V2 that impressed much of the global tech community with its ability to write code and do math. Its developers also claim it beat competitors at common benchmarks — at a fraction of the cost of other tools developed by U.S. tech giants, and despite the intense restrictions on access to chips from Washington.”
In mid-2024, Andrew Carr, chief scientist at Cartwheel, an AI animation start-up based in the U.S., told the Financial Times that the DeepSeek “model’s architecture is very unique,” allowing it to achieve similar performance as Meta’s latest Llama 3, OpenAI’s GPT-4, and Anthropic’s Claude 3 Haiku — all at a lower cost.
Despite DeepSeek being an open secret to many, the release of its R1 models did blindside U.S. tech Illuminati and market whizzes. And this was the case even though AI companies closely monitor competitors’ published scientific papers, technical blog posts, and open-source releases for insights into their technical approaches, capabilities, and strategic directions. Similarly, equity analysts form their investment recommendations by digging into AI research and technical publications, looking for insight into the drivers of AI companies’ capabilities, competitive position, and future growth.
So why were they caught by surprise?
I offer three possible explanations:
- American technological dominance rested on a dangerous assumption: that the U.S. held an insurmountable lead in frontier AI technologies. Recent developments disprove this myth. As Miles Brundage, former Open AI’s senior advisor for AGI Readiness, succinctly noted, AI capability depends primarily on skill and computational scale -- not secret technical advantages. This reality means no nation can maintain a permanent monopoly on advanced AI development.
- The assumption that capital alone creates a sustainable advantage in AI development has proven recklessly naive. Bloomberg reported that in 2024, the CEO of a high-profile AI company was asked at a private dinner how his company was different from peers that were building “the systems underpinning chatbots like ChatGPT. Did he have a moat? Yes, he answered, according to another CEO who was there. No one else had raised the billions of dollars that he had. That was his moat.” Apparently, this CEO failed to read the leaked internal Google document that proclaimed, “We have no moat and neither does OpenAI.” DeepSeek shows that technical talent, computing infrastructure, and execution speed often matter more than financial resources alone.
- The U.S. government, tech companies, and the financial markets overestimated the deterrent effect of restricting advanced AI chip exports to China. In an op-ed in the Financial Times, Angela Zhang, professor of law at the Gould School of Law, University of Southern California, expands on this argument, “It [DeepSeek’s achievement] ]also lays bare the limits of U.S. export controls designed to slow China’s AI progress. While these measures may deliver short-term disruptions, their impact diminishes over time as China innovates to adapt. The inconvenient truth for US policymakers is that strict export controls have forced Chinese tech companies to become more self-reliant, spurring breakthroughs that might not have occurred otherwise.”
In sum, Silicon Valley and the markets were caught off guard because they assumed the core principle of the American tech industry — that innovation can happen in a garage — was exclusive to the United States. However, as DeepSeek’s success demonstrates, the Chinese have their own garage.
Angelo Calvello, Ph.D., is the founder of C/79 Consulting, LLC.
