Background
DeepSeek was launched as a research lab in December 2023 by AI enthusiast Liang Wenfeng, who had been using AI trading algorithms in his High-Flyer hedge fund. Like Meta’s Llama models, DeepSeek models are open source, meaning anyone can download, modify, and use them commercially.
Although DeepSeek previously claimed to have no detailed plans for commercialization, the company is now selling API access to a hosted version of its latest model—and at prices far below those of other AI companies. Its low prices forced its Chinese competitors to drastically lower their API pricing. Whether US companies will have to follow suit is unclear.
DeepSeek’s latest releases include a conventional LLM used in its chatbot, plus a “reasoning model” called R1 that is similar to OpenAI’s advanced o1 model. Reasoning models spend more time thinking through problems before answering, resulting in higher accuracy and problem-solving abilities. Reasoning models aren’t good at language-oriented tasks or creating chatbots. But they can perform better for coding, math, science, and other advanced problem-solving tasks.
The sudden rise of DeepSeek as a major AI player caught the industry off-guard. DeepSeek’s iPhone chat app rocketed to the top spot for AI app downloads within a few days of its release (although the company claims that shortly thereafter it was hit by a cyber-attack that forced it to limit new signups). However, it’s not just the quality of the model that took the industry by surprise. It’s DeepSeek’s claim that it trained its new models at a fraction of the time and cost of US model providers and without state-of-the-art GPUs.
Implications
In the days after the announcement of the latest DeepSeek models, the US stock market lost over a trillion dollars in value. Most of the losses came from NVIDIA and other AI-related companies. The market has rebounded slightly, but DeepSeek has cast a pall over tech companies and has raised questions of national security. Here are the main issues.
Changing economics of AI
It’s not possible to verify DeepSeek’s claims that it created its model for approximately one-tenth of the cost of US model producers. But if true, it represents a fundamental shift in the economics of the AI industry. It’s good news for the environment, as it suggests that fewer data centers might be needed. However, it raises questions about whether the billions of dollars that tech companies are spending on AI infrastructure will be a profitable investment, creating significant economic risk.
Data privacy
While DeepSeek’s stated privacy policies are not dissimilar to those of US companies like OpenAI, its servers are located in China, which means that the Chinese government almost certainly has access to any data collected. Since AI models are used by research organizations and major corporations, the privacy risks associated with an AI model make those related to a social media app such as TikTok seem like small potatoes. There are also concerns about the attention that companies located in China pay to protecting data. It took security research firm Wiz Research only minutes to discover a publicly accessible database that leaked highly sensitive information from DeepSeek users.
Censorship
DeepSeek’s models follow Chinese government guidelines on political issues. Queries about Taiwan, Tiananmen Square, and other sensitive topics are blocked or answered with propaganda. This raises the specter of DeepSeek becoming a vector for China-based influence campaigns based on disinformation.
National security
US leaders consider the AI race an issue of national security. US chip restrictions were supposed to keep China from catching up to the US in AI. The fact that a Chinese company created a leading model without state-of-the-art GPUs suggests that chip restrictions aren’t the handicap the US believed them to be.
The release of the latest DeepSeek models is being widely referred to as a “Sputnik moment” for the US, referring to the shock in the US when the Sputnik launch showed that Russia was actually ahead of the US in the space race. This is not necessarily a bad thing, as it will focus national attention from both the US government and AI companies on efforts to catch up. The increased attention could massively spur innovation, as it did with space technology. But given AI’s future impact on global economics and security, it’s hard to overstate what’s at stake.
Now what?
By all appearances, DeepSeek is a good actor. As a research lab, it has made genuine innovations. And it has been far more open about its techniques than most US AI companies. DeepSeek has pushed the entire AI industry forward and perhaps even contributed to lessening the AI data center energy crisis. It just happens to be based in a country where there is a blurry line between the government and private businesses. And that government happens to have a complicated, but mostly adversarial, relationship with the US.
It’s too early to tell whether DeepSeek’s models are as good as the early reports. Benchmarks don’t always tell an accurate story about model performance. Even if they are, it seems unlikely that DeepSeek models will get widespread use by US businesses, at least at the enterprise level. The cost is tempting. But the risks are high. And there’s even a chance that the US government will step in to prevent it, as it did with TikTok.
That means the stock market reaction is likely overblown. However, the ability of a scrappy foreign startup to compete with top-tier AI vendors without access to the massive resources mainstream AI companies possess is a preview of AI’s likely future. And it is an uncertain one. It will force US companies to scramble to find ways to keep their technology lead—if they can.
