Posted by 
When it comes to AI, I’d consider myself a casual user and a curious one. It’s been creeping into my daily life for a couple of years, and at the very least, AI chatbots can be good at making drudgery slightly less drudgerous.
But whenever I start to feel convinced that tools like ChatGPT and Claude can actually make my life better, I seem to hit a paywall, because the most advanced and arguably most useful tools require a subscription. Then came DeepSeek.
The Chinese startup DeepSeek sunk the stock prices of several major tech companies on Monday after it released a new open-source model that can reason on the cheap: DeepSeek-R1. The company says R1’s performance matches OpenAI’s initial “reasoning” model, o1, and it does so using a fraction of the resources. It also cost a lot less to use. That adds up to an advanced AI model that’s free to the public and a bargain to developers who want to build apps on top of it.
While OpenAI, Anthropic, Google, Meta, and Microsoft have collectively spent billions of dollars training their models, DeepSeek claims it spent less than $6 million on using the equipment to train R1’s predecessor, DeepSeek-V3. (Disclosure: Vox Media is one of several publishers that has signed partnership agreements with OpenAI. Our reporting remains editorially independent.)
To get unlimited access to OpenAI’s o1, you’ll need a pro account, which costs $200 a month. DeepSeek does charge companies for access to its application programming interface (API), which allows apps to talk to each other and helps developers bake AI models into their apps. But what DeepSeek charges for API access is a tiny fraction of the cost that OpenAI charges for access to o1. So it might not come as a surprise that, as of Wednesday morning, DeepSeek wasn’t just the most popular AI app in the Apple and Google app stores. It was the most popular app, period.
“The main reason people are very excited about DeepSeek is not because it’s way better than any of the other models,” said Leandro von Werra, head of research at the AI platform Hugging Face. “It’s more that it’s an open model, and coming from a place where people didn’t expect it to come from.”
So as Silicon Valley and Washington pondered the geopolitical implications of what’s been called a “Sputnik moment” for AI, I’ve been fixated on the promise that AI tools can be both powerful and cheap. And on top of that, I imagined how a future powered by artificially intelligent software could be built on the same open-source principles that brought us things like Linux and the World Web Web.
This could be wishful thinking and a little bit naive. After all, OpenAI was originally founded as a nonprofit company with the mission to create AI that would serve the entire world, regardless of financial return. That’s no longer the case.
But this is why DeepSeek’s explosive entrance into the global AI arena could make my wishful thinking a bit more realistic. While my own experiments with the R1 model showed a chatbot that basically acts like other chatbots — while walking you through its reasoning, which is interesting — the real value is that it points toward a future of AI that is, at least partially, open source. It indicates that even the most advanced AI capabilities don’t need to cost billions of dollars to build — or be built by trillion-dollar Silicon Valley companies. That means more companies could be competing to build more interesting applications for AI.
And while American tech companies have spent billions trying to get ahead in the AI arms race, DeepSeek’s sudden popularity also shows that while it is heating up, the digital cold war between the US and China doesn’t have to be a zero-sum game.
DeepSeek’s unconventional, almost-open-source approach
While you may not have heard of DeepSeek until this week, the company’s work caught the attention of the AI research world a few years ago. The company actually grew out of High-Flyer, a China-based hedge fund founded in 2016 by engineer Liang Wenfeng. High-Flyer found great success using AI to anticipate movement in the stock market. That, however, prompted a crackdown on what Beijing deemed to be speculative trading, so in 2023, Liang spun off his company’s research division into DeepSeek, a company focused on advanced AI research.
From the outset, DeepSeek set itself apart by building powerful open-source models cheaply and offering developers access for cheap. In the software world, open source means that the code can be used, modified, and distributed by anyone. In the context of AI, that applies to the entire system, including its training data, licenses, and other components. Thanks to DeepSeek’s open-source approach, anyone can download its models, tweak them, and even run them on local servers.
The major US players in the AI race — OpenAI, Google, Anthropic, Microsoft — have closed models built on proprietary data and guarded as trade secrets. Meta has set itself apart by releasing open models. Conventional wisdom suggested that open models lagged behind closed models by a year or so. DeepSeek apparently just shattered that notion.
DeepSeek’s models are not, however, truly open source. They’re what’s known as open-weight AI models. That means the data that allows the model to generate content, also known as the model’s weights, is public, but the company hasn’t released its training data or code. Von Werra, of Hugging Face, is working on a project to fully reproduce DeepSeek-R1, including its data and training pipelines. One of the goals is to figure out how exactly DeepSeek managed to pull off such advanced reasoning with far fewer resources than competitors, like OpenAI, and then release those findings to the public to give open-source AI development another leg up.
“If more people have access to open models, more people will build on top of it,” von Werra said.
Still, we already know a lot more about how DeepSeek’s model works than we do about OpenAI’s. DeepSeek published a detailed technical report on R1 under an MIT License, which gives permission to reuse, modify, or distribute the software. A similar technical report on the V3 model released in December says that it was trained on 2,000 NVIDIA H800 chips versus the 16,000 or so integrated circuits competing models needed for training. Training took 55 days and cost $5.6 million, according to DeepSeek, while the cost of training Meta’s latest open-source model, Llama 3.1, is estimated to be anywhere from about $100 million to $640 million. But because Meta does not share all components of its models, including training data, some do not consider Llama to be truly open source.
When it comes to performance, there’s little doubt that DeepSeek-R1 delivers impressive results that rival its most expensive competitors. A comparison of models from Artificial Analysis shows that R1 is second only to OpenAI’s o1 in reasoning and artificial analysis. It actually slightly outperforms o1 in terms of quantitative reasoning and coding. The big tradeoff appears to be speed. DeepSeek is kind of slow, and you’ll notice it if you use R1 in the app or on the web. It does show you what it’s thinking as it’s thinking, though, which is kind of neat.
Now, the number of chips used or dollars spent on computing power are super important metrics in the AI industry, but they don’t mean much to the average user. The most basic versions of ChatGPT, the model that put OpenAI on the map, and Claude, Anthropic’s chatbot, are powerful enough for a lot of people, and they’re free. They can summarize stuff, help you plan a vacation, and help you search the web with varying results. But chatbots are far from the coolest thing AI can do.
The challenge to America’s global AI supremacy
What’s most exciting about DeepSeek and its more open approach is how it will make it cheaper and easier to build AI into stuff. This is a huge deal for developers trying to create killer apps as well as scientists trying to make breakthrough discoveries. It’s also a huge challenge to the Silicon Valley establishment, which has poured billions of dollars into companies like OpenAI with the understanding that the massive capital expenditures would be necessary to lead the burgeoning global AI industry.
It’s not an understatement to say that DeepSeek is shaking the AI industry to its very core. The stock market’s reaction to the arrival of DeepSeek-R1’s arrival wiped out nearly $1 trillion in value from tech stocks and reversed two years of seemingly neverending gains for companies propping up the AI industry, including most prominently NVIDIA, whose chips were used to train DeepSeek’s models.
It also indicated that the Biden administration’s moves to curb chip exports in an effort to slow China’s progress in AI innovation may not have had the desired effect. Joe Biden started blocking exports of advanced AI chips to China in 2022 and expanded those efforts just before Trump took office. However, China’s AI industry has continued to advance apace its US rivals. DeepSeek is joined by Chinese tech giants like Alibaba, Baidu, ByteDance, and Tencent, who have also continued to roll out powerful AI tools, despite the embargo.
What this means for the future of America’s quest for AI dominance is up for debate. President Donald Trump praised DeepSeek’s ability to come up “with a faster method of AI and much less expensive method.” He added, “The release of DeepSeek, AI from a Chinese company should be a wakeup call for our industries that we need to be laser-focused on competing to win.”
But we’re far too early in this race to have any idea who will ultimately take home the gold. “This is like being in the late 1990s or even right around the year 2000 and trying to predict who would be the leading tech companies, or the leading internet companies in 20 years,” said Jennifer Huddleston, a senior fellow at the Cato Institute.
What is clear is that the competitors are aiming for the same finish line. Liang said in a July 2024 interview with Chinese tech outlet 36kr that, like OpenAI, his company wants to achieve general artificial intelligence and would keep its models open going forward. He added, “OpenAI is not a god.” Liang’s goals line up with those of Sam Altman and OpenAI, which has cast doubt on DeepSeek’s recent success. Microsoft and OpenAI are reportedly investigating whether DeepSeek used ChatGPT output to train its models, an allegation that David Sacks, the newly appointed White House AI and crypto czar, repeated this week.
There is, of course, the chance that this all goes the way of TikTok, another Chinese company that challenged US tech supremacy. It was originally Trump who cited national security concerns as a reason to ban the app, which is owned by ByteDance. Congress and the Biden administration took up the mantle, and now TikTok is banned, pending the app’s sale to an American company.
DeepSeek uses ByteDance as a cloud provider and hosts American user data on Chinese servers, which is what got TikTok in trouble years ago. The concern here is that the Chinese government could access that data and threaten US national security. DeepSeek also says in its privacy policy that it can use this data to “review, improve, and develop the service,” which is not an unusual thing to find in any privacy policy.
Unsurprisingly, DeepSeek does abide by China’s censorship laws, which means its chatbot will not give you any information about the Tiananmen Square massacre, among other censored subjects. But it’s not yet clear that Beijing is using the popular new tool to ramp up surveillance on Americans. At least, it’s not doing so any more than companies like Google and Apple already do, according to Sean O’Brien, founder of the Yale Privacy Lab, who recently did some network analysis of DeepSeek’s app.
“From a privacy standpoint, people need to understand that most mainstream apps are spying on them, and this is no different,” O’Brien told me. “It’s just a question of who’s doing the spying.”
Which brings us back to that paywall question. There’s an old adage that if something online is free on the internet, you’re the product. So while it’s exciting and even admirable that DeepSeek is building powerful AI models and offering them up to the public for free, it makes you wonder what the company has planned for the future.
In the meantime, you can expect more surprises on the AI front. You might even be able to tinker with these surprises, too. OpenAI recently rolled out its Operator agent, which can effectively use a computer on your behalf — if you pay $200 for the pro subscription. This week, people started sharing code that can do the same thing with DeepSeek for free.
