Title: Why China's AI Startup 'DeepSeek' is a Sputnik Moment

[music]

Brian Lehrer: It's the Brian Lehrer Show on WNYC. Good morning again, everyone. Have you heard about DeepSeek yet? DeepSeek, China's answer to ChatGPT. This month, a Chinese tech startup launched DeepSeek-R1, a free open-source AI model that holds its own and in some cases surpasses, according to the reviews, the functionality of OpenAI's ChatGPT. Its release spurred panic in the US stock market. You've probably heard about that if you haven't been affected by it, erasing more than a trillion dollars in market capitalization as spooked investors rapidly sold their shares of many leading tech companies, but to understand this moment, here's a historical reference.

October 4th, 1957. Do you know that date? That was when the Soviet Union launched Sputnik, Earth's first artificial satellite, into space. From that point on, Cold War tensions between the United States and the Soviet Union escalated into space as both nations saw technological developments as indicators of their global influence, and the so-called Space Race was on. Why discuss Sputnik in this context?

Well, much of the news surrounding DeepSeek has referred to this development as China's Sputnik moment. Have you heard that? China was seen to be years behind the United States like the Soviet Union was back then. When it comes to tech, particularly AI, and just last week President Trump had signed a $500 billion deal with OpenAI, Oracle, and SoftBank to create something called Stargate, a company focused on building AI infrastructure in the United States.

The investment is seen as necessary according to a policy statement put out by OpenAI because, "If the US doesn't attract those funds, they will flow to China-backed projects, strengthening the Chinese Communist Party's global influence." Yet according to DeepSeek, their model, and this is such a key to this story, their model was built with just $5.6 million of investment, though that is to be debated. Joining us now to explain this moment is Reed Albergotti, technology editor for Semafor. Reed, thanks for joining us. Welcome to WNYC.

Reed Albergotti: Hi, great to be here.

Brian Lehrer: Let me get right to how dubious this claim of how cheap it was to develop DeepSeek really is, because that's why the stock market is freaking out. If it's going to cost them so much less money to make their stuff than it costs us, then all these competitors in the United States would be at such a disadvantage, but is it true? I'm going to play a clip of Scale AI CEO Alexandr Wang in conversation with CNBC's Andrew Ross Sorkin on CNBC's Squawk Box last week. I think we have that clip.

Alexandr Wang: You know, the Chinese labs, they have more H100s than people think. You know, the--

Andrew Ross Sorkin: These are the highest-powered NVIDIA chips that they were not supposed to have?

Alexandr Wang: Yes. My understanding is that DeepSeek has about 50,000 H100s, which they can't talk about obviously because it is against the export controls that the United States has put in place. I think it is true that-- I think they have more chips than other people expect, but also on a going-forward basis, they are going to be limited by the chip controls and the export controls that we have in place.

Brian Lehrer: Alexandr Wang, who is the CEO of Scale AI with Andrew Ross Sorkin on CNBC's Squawk Box last week. Reed, can you put this into plain English for our listeners? Why are DeepSeek's claims of a much cheaper AI product dubious?

Reed Albergotti: Well, I mean, I think what you heard there is a pretty good argument. I mean, we don't really know, and they have every reason to sort of hide the fact that they have more of these massive super powerful NVIDIA chips than they actually say. I would actually argue though that, I mean even if it's true that they found a way to train these models in a cheaper way and run them also, what they call inference, in a cheaper way, it still doesn't actually really change the market dynamics very much.

I think the the big drop in this in the stock market was really kind of unwarranted, and as Alex said in that interview, we still have a long way to go here before we get to, A, the capabilities of these models, where they need to be, and B, the price point. I mean these models actually do need to come down in price, and it's sort of a given that you have these many breakthroughs like we saw with R1 over the years. Otherwise, the industry doesn't really work that well.

Brian Lehrer: This happens with every technology. Right? Like how much did televisions cost in 1948 relative to the cost of living then compared to today?

Reed Albergotti: Yes, that's a great comparison. I also think, you mentioned the Sputnik moment, and people in the AI industry were calling DeepSeek kind of a Sputnik Moment over a month ago, before R1 came out, but they weren't advocating any change in the dynamics of how we should build out this technology. In fact, Sputnik didn't mean the end of spending on the US Space Race. It meant an increase in spending. I think this will be the same.

Brian Lehrer: Listeners, if you have any questions or relevant stories or maybe you work for one of these US tech companies that's freaking out over DeepSeek, give us a call and help us report this story or ask our guest a question. 212-433-WNYC, 212-433-9692. One of Trump's first moves as president was to team up with some of the most prominent tech CEOs to create this company Stargate. What's the purpose of Stargate in the United States' race to develop artificial intelligence, and how does it relate to the DeepSeek story?

Reed Albergotti: Stargate, just to be totally clear, Stargate actually existed before Trump became president. It was written about months before, but what Trump announced was really an increase in funding, really a commitment to fund Stargate. It went from something like $100 billion to $500 billion. Basically, it's data centers. We call them data centers, but really they're not like the data centers you think of, in the past, that serve you things like Netflix movies. These are data centers that are filled with super powerful GPUs, graphics processor units that crunch the numbers for these AI models both in training and inference, which is the term for actually running the models like when you prompt ChatGPT.

We just need a lot more of these things. Right now, there is so much demand for this AI that the companies really can't meet it. I mean, if you sign up for a super expensive subscription to a lot of these plans with Anthropic or OpenAI, if you use it enough, you start to hit these limits, they call token limits. The tokens are sort of the unit of how you measure the amount of data that's going back and forth when you interact with these models. There is just a bottleneck right now in compute power.

Brian Lehrer: Here's a clip of President Trump addressing, or commenting on the release of DeepSeek on Monday as the stock market was freaking out. That really does make it sound like a Sputnik kind of moment. He was speaking to House Republicans in Miami on Monday. Listen.

President Trump: I've been reading about China and some of the companies in China, one in particular, coming up with a faster method of AI and much less expensive method, and that's good because you don't have to spend as much money. I view that as a positive, as an asset. I really think-- If it's fact and if it's true, and nobody really knows if it is, but I view that as a positive because you'll be doing that too, so you won't be spending as much and you'll get the same result, hopefully. The release of DeepSeek, AI from a Chinese company, should be a wake-up call for our industries that we need to be laser-focused on competing to win.

Brian Lehrer: Wake-up call for our industries that we need to be laser-focused on competing to win. What do you make of Trump's statement on DeepSeek?

Reed Albergotti: Well, I think he's being advised by people who really understand the industry. I think he's right that it is probably a good thing as these costs come down, but it does also show that China is not very far behind, and that's not a surprise for people who are really close to this industry. China has a lot of great AI researchers. In fact, there's a lot of back and forth between academia in the US and China, and a lot of this research has been open for a long time. It's only really since ChatGPT came out that the tech companies have started to kind of close up their research and keep it secret.

Before, it was sort of like an academic. Even the people who worked at Google were almost like academics in publishing their research. I think what you're probably going to see is the China hawks say, "Well, if DeepSeek was able to do this with so few resources, and maybe even if their claims are true that they use less-powerful processors, maybe we actually need to expand these export controls.

Brian Lehrer: Let's take a call, who I think wonders, as you say, Trump is being advised by people who understand AI and DeepSeek, who wonders if Wall Street is. Lisa in Manhattan. You're on WNYC. Hello.

Lisa: Hi. Yes, exactly that transition, how well do the investors and analysts and whatnot understand DeepSeek and how much of the market reaction, particularly in those isolated in those chip stocks, was just a piling on of either trading that was automated, trading that had to do with buckets of stock, and when you're on calls, when you're listening to these investor calls and the analysts asking questions, how much do they truly understand these issues here?

Brian Lehrer: Thank you, Lisa.

Reed Albergotti: I think that's really a great question, and something, I think one of the biggest takeaways from all of this is that the market still doesn't really understand this technology. It's still very new. We're all still kind of learning about it. It was very surprising to me to see the market reaction. It seemed to be almost a reaction to social media memes around DeepSeek rather than the fundamentals or a true understanding of the market.

Now, that's not to say there's not a bear case for AI. I could certainly steel-man an argument that all these investments are way too high and the valuations are too high, et cetera, et cetera, but it's just that DeepSeek's R1 model is not part of that argument. It's a side-- It's kind of a sideshow. I think we expect these things to happen. In fact, it has been happening. Companies like Meta or Mistral in France, they've been using almost the same techniques to "distill" these big models down into smaller, more efficient ones, so it's just not a-- I think there was a lot of timing around this, right after Trump's announcement of the Stargate thing.

Then it was right at Davos where you had a lot of people gathered together who were really hyperfocused on AI rather than some truly fundamental shift in the way we view this technology.

Brian Lehrer: Here's a text from a listener who says, "I know you're not the first to make it, but I don't think Sputnik is a perfect analogy. DeepSeek is built off Meta's Llama model. The story here, as I see it, is a triumph of open source over closed source." Then the listener points out, "OpenAI, contrary to its name, does not open source its current models." Would you take a step back for the lay listener and explain what open source is and why it's relevant to the DeepSeek story?

Reed Albergotti: Yes. Open source is, we call it open source. A lot of people call it open weights models because they don't actually give you all the information. They don't tell you exactly how these models are trained, but they tell you essentially how they work, and they allow anyone to run these models anywhere. I can go right now-- In fact, I have-- You can go on these websites, like Hugging Face, and you can download this model, and you can run it on your home computer. Hopefully, you have a very fast one, although there are smaller models that can run even on less-powerful computers.

Open source has really flourished in the AI industry, has for a long time. Meta's strategy here, they make the Llama model that DeepSeek is based off of, is that they're going to create this technology, spend a lot of money on it, and open source it, make it free for anybody to use for the most part, and they will benefit by by sort of building this ecosystem around their technology. That's sort of the classic open-source strategy. OpenAI, on the other hand, does not tell people how their models work. They don't allow people, anybody, to just download it and use it on their home computers. They make you pay a subscription or pay for an API key to use their models.

That's how they want to keep their-- they keep their research secret. There are two different philosophies and strategies here. I think it is-- I mean, you definitely do see-- I mean, this is how open source is supposed to work. You have one company come up with a breakthrough, everybody else learns from that. There's sort of competitive spirit there, and they improve on it, and that's how it works. I think both of these things, though, will exist. They both have a place in the market. Again, we've sort of known this, I think it's just going to continue along those lines.

Brian Lehrer: Yes, but I guess when we're talking about open source, and the listener says, "It's an imperfect analogy, Sputnik," I guess that means because NASA wasn't sharing its technology with the Soviets, and here, in effect, we are. Right?

Reed Albergotti: Right. [crosstalk]

Brian Lehrer: Go ahead.

Reed Albergotti: Yes, I think you have to make sort of a leap to call it a Sputnik moment. You have to say, sort of stepping back in the abstract, it just shows that China has a lot of brain power, and if they have the chips which we're currently trying to prevent them from having, they might leap ahead of the US, and that's really kind of the-- that holds true, I think, even if the details aren't correct. I mean, there's a lot of things that probably are imperfect about any historical analogy, but for one, Sputnik, the US was probably farther behind. I mean, in this case, it's China sort of almost maybe catching up rather than surprising everybody with the first satellite around space.

Brian Lehrer: Now, there's also an environmental and potentially climate implication to this Chinese AI product DeepSeek, and Ann in Rivington wants to ask about that. Hi, Ann.

Ann: Oh, hi. Can you hear me?

Brian Lehrer: I can hear you.

Ann: Oh, good. Yes, I'm very interested in the aspect of the Chinese technology that is so much more energy-efficient. Donald Trump, one of the first things he did in office, of course, was open up fossil fuel production and drilling so that the big tech companies could have cheap fossil fuel, but it will be interesting to see, I wonder what impact the energy efficiency aspect of this will have on the development of Stargate and AI in the US.

Brian Lehrer: Reed?

Reed Albergotti: Well, I'm happy to address that. Unfortunately, there's something called the Jevons paradox, which I think has really held true, which is that as this technology becomes cheaper and more accessible, the demand for it just increases. We've seen the cost of these models come down by well over 100% since ChatGPT came out, and that has only increased demand, and so what you see is, yes, the AI companies and the chip companies like NVIDIA and Amazon and Google, they're all trying, they're racing, really. Actually, the race, and we can get into this in more detail.

The race is really to make this stuff more efficient, but as they do that, as they make it cheaper, then people just buy more compute. I just see no end in sight of the desire for more and more of these tokens, as we call them. Unfortunately, we are going to, in order to satisfy the demand, we are going to have to add something like, some estimates are like the country of France to the US energy grid, so hopefully, we can find a way to do that in an environmentally friendly way and not have to burn coal or gas or other fossil fuels. Hopefully, there's breakthroughs in battery technology and maybe nuclear as well, because we are about to see a massive jump in energy production.

Brian Lehrer: On DeepSeek sharing its code, being open source, listener writes, "DeepSeek is open source because it's almost certainly stolen source." You addressed this before, but what could the upshot of this be? Are some of the US tech companies, maybe OpenAI, which makes ChatGPT, maybe even others, going to sue them? Is there that kind of recourse, shut them down because they have stolen their code?

Reed Albergotti: Yes. I mean, I think if they sued them, it would be sort of an interesting copyright question. Obviously, OpenAI and other AI companies are facing their own copyright lawsuits, and there are really interesting copyright questions in this industry. What we're talking about here is this process of distillation. This is something that researchers have been doing now for years.

Microsoft Research has a project called Textbooks is All You Need. What they do is they take OpenAI models and they instruct them to essentially create data that you can use for training. They call them Textbooks, but it's really just synthetic data. Then they use other models to kind of check that data and really hone it and distill it down into really all you need, so these models become much, much more efficient. That's allegedly what DeepSeek did. They allegedly used OpenAI, ChatGPT, to create this synthetic data in this process called distillation.

We don't actually know that for sure, but if they did, it would violate OpenAI's terms of service, and so we might see lawsuits, but I think more likely what we're going to see is attempts by OpenAI and the closed-source foundation model companies like Anthropic and others to try to prevent this practice. They might try to algorithmically figure out, "Okay, what are the type of prompts that might be associated with this distillation product and stop it from happening."

Brian Lehrer: Andrew in Astoria, you're on WNYC with Reed Albergotti from Semafor covering the release of this Chinese AI product called DeepSeek, which is freaking out the US stock market. Hi, Andrew.

Andrew: Hey, Brian and guest. Good morning. I'm calling just to give a little more context, one, that DeepSeek is one of the hottest downloaded apps on the App Store, Apple's App Store, and it's, like your guest had spoke about, developed by a Chinese company. That means a lot of topics such as Tiananmen Square protests, Taiwan's political status are completely off limits.

You get like a vague response or a redirection when you try to ask it about it. Being an app from China, according to their privacy policy, the company does collect various types of user data like text, audio inputs, prompts, chat history, and that's stored on secure servers in China, but maybe your guest can speak more on the ability, because it's an open-source model, to download it locally and sever its ability to send data back to China.

Brian Lehrer: [crosstalk] It's exactly why TikTok might wind up being banned in the-- is under the law. That's on a 75-day reprieve. This is exactly why TikTok might be banned in the United States ultimately, right?

Reed Albergotti: Yes. I think, that app, if I had to guess, is probably not long for this world. If you thought TikTok's data gathering was bad, the type of data you get from these chatbots is really problematic. I mean, people actually fall in love with these things and date them, as you might have read, some funny stories in big publications lately. I think that the app will for sure-- I mean, it is, it is drawing alarm from lawmakers already.

Then, as your caller, your listener, mentioned, if you download this app and run it on your local computer, there's no data that's going back to Chinese servers. I do wonder, though, and this is speculation, whether lawmakers will also take issue even with the open-source model because they might fear that there could be some sort of backdoor in the model that might allow attacks.

There's something called the prompt engineering attack where you can sort of trick these models into giving out information that they're not supposed to give out. This is all very frontier. We don't really know how these security vulnerabilities are going to materialize in these models, so I suspect there is going to be a lot of discussion and debate on this in the coming weeks and months.

Brian Lehrer: The caller also mentioned censorship by the Chinese government, and I want to play a clip of Victor Gao, who used to work under the former Chinese leader Deng Xiaoping, discussing censorship on DeepSeek, but as you will hear, he also has something to say about China's role generally in AI development.

Victor Gao: Well, first of all, I think the breakthrough by DeepSeek means AI is not American intelligence. It means AI should be artificial intelligence for mankind. It also means that there is not one single pathway in the development of AI and no one has a monopoly of AI, and countries like China and many other countries all have a right to participate in AI and make their contribution to the development of AI.

Now, DeepSeek is also very intelligent in itself, and it is really thinking deeply and seeking deeply, and it also knows the sensitivities. Now, we need to acknowledge that for every culture, every civilization anywhere in the world, there are sensitivities. From DeepSeek's perspective, they recognize the sensitivities and they want to further explore all the other possibilities, and that means DeepSeek is very sophisticated rather than just a blundering machine.

Brian Lehrer: That DeepSeek recognizes the sensitivities of various countries. That could either be reassuring or it could be taken as very threatening. Let's end kind of the way we began, because in my intro to this segment, I quoted, "OpenAI's policy memo on US investment into AI development." Here's that quote again. "If the US doesn't attract those funds, they will flow to China-backed projects, strengthening the Chinese Communist Party's global influence." Hearing this quote and that clip, what would it mean, from your perspective as a reporter, for the world if China were to sort of win the AI race?

Reed Albergotti: Well, I think whoever wins the AI race, first of all, they're not going to win it by very much. This is a very close race, and I think we can also argue about how to define what is winning the AI race. I don't think we really quite even know yet how good these models will ultimately get and what they will mean, but let's say you get to something like, people call it artificial general intelligence or superintelligence. That is definitely a very powerful military tool, and we know that the Chinese government, it has expressed desire to invade Taiwan.

If that happens, we could see a very bloody war between two superpowers. I don't think anybody in the US wants to see that. I think that's really what pushes these China hawks to want to "win" that race. The recording also talked about AI is a technology for humanity. I think that's simultaneously true. I mean, all the technology developed during the Cold War, post Sputnik, ultimately became consumer technology that really is now spread around the world and really has changed humanity, so I think both of those things will be true. I think also, the last point, is we have a long way to go. I don't believe that next year we're going to-- that the race is going to be won or lost. I think it's a multi-year thing.

Brian Lehrer: Reed Albergotti, technology editor at Semafor. Thanks a lot, Reed.

Reed Albergotti: Great to be here. Thank you.

Copyright © 2025 New York Public Radio. All rights reserved. Visit our website terms of use at www.wnyc.org for further information.

New York Public Radio transcripts are created on a rush deadline, often by contractors. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of New York Public Radio’s programming is the audio record.