Behind The Numbers: DeepSeek: What It Is and How It’s Shaking Up the AI Universe | Feb 03, 2025
25:17 Share

Hey gang, it's Monday, February 3rd. Grace Garjo and listeners, welcome to Behind the Numbers, an eMarketer video podcast made possible by Zeta Global. I'm Marcus, and today we're discussing DeepSeek, what it is and how it has completely shaken up the entire AI landscape. Today I'm joined by two folks we have with us, our technology and AI analyst who lives out in California. It's Grace Harmon.

Hi, nice to be here. Hello there. And we have with us as well, a senior analyst based on the other coast in New York coming to us from our studio. It's Gagio Sevella. Grace Marcus, nice to be here.

We start with the facts of the day. Which land animal has the longest migration? As Kayla Zhu of Visual Capitalist notes, land animals migrate to move between feeding grounds, safer habitats with fewer predators and their home ranges. According to a 2019 study published in the science journal Nature, caribous, as they're called in North America, or simply reindeer, as they're known in other parts of the world, travel 840 miles round trip.

So people are like, what the hell does that mean? So for folks on the East Coast, so for Gajo and Victoria, who edits the show, who's coming to us from the studio as well. And Stuart, who runs the team. Everyone's there. It's like walking from New York to Cleveland or Toronto and then back again. People on the West Coast. So Grace, that's like walking from L.A. to San Francisco and back. And for my European audience, it's like walking from London to Switzerland and then back.

That's too far. What's that? Good for the caribou. So basically you'd have to walk, it's like half a marathon's distance each day with an hour for lunch and it would take you about two months to do that. So by the time you get there, you'd immediately have to turn around. That's so far. I want to say waste of time because I'm sure it's necessary, but it seems like it. Anyway, today's real topic. What is DeepSeek and how has it slash will it change the AI world?

All right, so let's tuck into it, folks. Lots to talk about. ChatGPT maker OpenAI, tech giant Oracle, Japan SoftBank, and MGX, a tech investment arm of the UAE government, just announced that they're planning on building $500 billion of AI infrastructure in the U.S.,

BBC article noted that the new company called the Stargate project was announced at the White House by President Donald Trump, who billed it as the largest AI infrastructure project in history and said it would help keep the future of tech in the US. Less than a week later, Chinese startup DeepSeek had other plans. They released a free AI powered chatbot called R1 that is reportedly at least as powerful as OpenAI's latest R1.

01 model at maths, coding and natural language reasoning, but developed for a fraction of the price, allegedly, of its US rival. Its sudden debut wiped $1 trillion off the value of US tech stocks. AI chipmaker NVIDIA took about half of that hit, falling 17%, the single biggest single day stock decline in history, due to concerns that future AI models could be developed more efficiently, possibly reducing the demand for NVIDIA's powerful GPUs. As Kylie Robertson

Elizabeth Lepato of The Verge put it, it took about a month for the finance world to start freaking out about DeepSeek. But when it did, it took more than half a trillion dollars or one entire Stargate project off NVIDIA's market cap.

So where has DeepSeek come from? Real quickly, a short history. A company was founded in 2023, 2023, how most normal people say it. The southeastern Chinese city of Hangzhou is where it was founded by Liang Wenfeng. He's 40 years old, information and electronic engineer graduate, founded one of China's largest quantitative hedge funds that backs DeepSeek. This past Christmas, DeepSeek released a reasoning model called V3. Its second model, R1, was released last week,

called one of the most amazing and impressive breakthroughs he's ever seen, according to Mark Adresin, a VC and advisor to the president. And it quickly became the most downloaded free app in the US just a week after it was launched. Okay, Gajo, what's your initial reaction to the launch? So two things came to mind. First, it's free. It's a free tool. Anyone can use it. And the second is it's built on open source AI frameworks.

meaning a lot of the work that's gone into it is available for other companies and developers to use, alter, and share if they want to. So I think those two things make it a compelling release, especially on a global scale, given that they've shown that it's been trained on a lot less

a lot less energy, using maybe older GPUs. So it's sort of leveled the playing field, especially for startups or companies that are looking at adapting Gen AI solutions that are pretty much cutting edge at this point. Yeah. Yeah. That's a really interesting part of this because it's

It's open source, people can pick it up, adapt it for what they want to adapt it for, and it does open the door for these cheaper AI alternatives. The French government immediately said, "This message is that we can compete now." That's what the message is. We don't have to rely on raw computing power to determine who wins the AI race. However, that also means that smaller startups, similar to the French startup Mistral AI,

now have a new competitor in DeepSeek, but also probably a bunch of other competitors who are going to look at this and say, oh, we can now get into the market. So it's kind of like almost like with Spotify, for example, before Spotify or YouTube, it was hard to get recognized. But when you created Spotify or YouTube, it meant everyone could get there, which meant it was easy to get recognized because of the power of the platform. But it also meant that everyone was there. And so that

in turn made it harder. So I'm interested to see whether this makes it easier or more difficult for some of the smaller startups to compete. Yeah, I think you can contrast that with Project Stargate. Because what is Project Stargate? They're building massive data centers. These are power-hungry, water-hungry data centers. They're building that for one company. That's for OpenAI.

It's not for all the AI companies in the US or even in North America. And that tells you, you know, it's a closed system. You know, they're looking at really heavily resourced, heavily funded investments to sort of get to the next level, which is AGI. And on the other hand, you have, you know, DeepSeek, who's taking a completely different approach here.

you don't hear about their plans to get investors, to get massive data centers. And they're really just focusing on the building of the AI through their models, through refining what's already existing. Whether or not they've taken bits and pieces from companies like OpenAI or other open source AI companies, the point is,

It looks like DeepSeq has become what OpenAI was when it first started, meaning, you know, it's a project. It's about building the best AI possible with what you have. And so that really shows you the dichotomy between those two events that have happened in a relatively short time. It does seem like...

data centers of the investment project du jour uh there were some numbers from mckinsey saying global demand for data center capacity would more than triple by 2030 growing between 19 27 annually so um lots of appetite for for that um i'm also wondering uh

Two things I guess, Grace. One, is the Stargate project dead on arrival? Because Ben Berkowitz of the Axios was saying that if China can do AI better and faster at one one thousandth of the cost, kind of cast a shadow on the rationale for the Stargate project investing $500 billion in AI infrastructure. And then kind of alongside that, how are shareholders of US tech giants kind of looking at this? Because they're looking and saying, hey, why are you spending billions and billions of dollars

Is that an efficient use of money when there's other companies done it for a lot cheaper? Meta's planning to spend over $60 billion in capital investment. That's 50% more than it did last year. Microsoft's planning to earmark, or has earmarked 80 billion on AI data centers in its current fiscal year. So how do you think this announcement has disrupted those larger, more established, at least in the West, AI players?

Sure. Well, I think that one thing that DeepSeq is doing that could impact the project is it's kind of upending the idea of scaling law, which is the more data and the more computing power that you put into an AI model, the smarter it gets. And like Gajo said, DeepSeq is operating with older hardware and slightly weaker GPUs and being able to produce these really high quality models.

With investors, there's been concerns for a while, but whether there's an investment bubble growing because of the sheer amount of money going in and investors are becoming a little less patient. You have CEOs of these big tech companies saying that the risk of under-investing is greater than the risk of over-investing, but that's not necessarily a sentiment that's shared by investors.

I think that one thing that DeepSeq could have done that might shoot itself in the foot a bit is that OpenAI and Microsoft are already looking into whether DeepSeq used OpenAI

Open AIs APIs to train its models. So if that's something that's true, I think that kind of undercuts the idea that they were able to do so much with so little. If that's what's true, I'm not saying it necessarily is. I mean, Microsoft's jumping in on that investigation because they are the exclusive licensor of Open AIs APIs. So I think that if...

If that's accurate, that's something that could be a really big issue. I think that what DeepSeek is doing also just alters the standards for AI development. And there's pros and cons for whether that'll let more people get into the market, like you said, like Mistral. But for investors, I think that that is why this had such a big impact on the stock market is that

investors don't want to be paying as much as they are paying for AI considering how long the timelines are for AGI, for AI to become so good that you can really progress it and really ask consumers to pay a lot for it. So I think that's why there was...

the scaling back on the stocks. The US tech stocks, I mean, we talked about Nvidia dropped 17% in a single day, wiped half a trillion dollars out of it. But others also took a significant hit. Oracle fell on 14%, Supermicrocomputer, which makes servers used for Gen AI, fell 13%, Chipmaker Broadcom fell 17%. So a lot of folks were affected by this. What

One of my questions, Gaju, is what does this do to AI spending? Does it supercharge it? Does it make people kind of reconsider all of this money that is being pumped into it? Because Ion Stoica, co-founder and executive chair of AI software company Databricks, said the lower cost of DeepSeq could spur more companies to adopt AI into their business,

As opposed to people saying spend less because it shouldn't cost as much, people might spend more because look at what we can do with the advent of this R1 model. It could also help with the kind of huge energy hurdle facing AI. However, because it uses less energy, it could mean more investment, which means that inevitably it will use more energy. The energy problem basically is the 2024 Energy Department reporting that AI will account for 7% to 12% of US electricity by 2028.

It's up from 4% in 2023. So where does AI spending go from here? I think there's going to be a lot of, there's going to be a reassessment along AI spending. You know, investors are probably not going to want to consolidate into, you know, big tech players that much. As DeepSeek has shown, you know, it's worth looking at startups and,

It's worth looking at that spirit of innovation over kind of a spirit of profit. And I think we're going to see that going a lot deeper. So if there have been stalled investment cycles in startups, I think that might spin up again. At the same time, I mean, companies like NVIDIA might see less investment because...

We've now seen a world where maybe you don't need the greatest, biggest, most expensive, hardest to get hardware that's decided on and supplied by one company. And so that changes the game a little bit. And if you had money to invest, you'd probably want to spread it out and bet on...

maybe longer term, but sure bets, right? Yeah, it's a chance to hedge your bets. Yeah, absolutely. So let's talk a bit about this model. How much more efficient is R1? Well, DeepSeq claims its model can be trained on 2,000 specialized chips versus an estimated 16,000 for leading models. And it also says that it costs $6 million to train.

a fraction of the over 100 million alluded to by OpenAI boss Sam Altman when discussing GPT-4. There was, I saw one ranking, I don't know if you say IMSYS or IMSYS, but it's a crowdsourced ranking of chatbots. They put R1 seventh, higher than any other open source model,

and the highest produced by any company other than Google or OpenAI. But I mean, how much of an impact do you think this should be making to the market given, yes, how relatively cheaply it was developed and with less processing power? But what's your assessment of how good this model is as it stacks up against some of the ones out there in the market? Okay, I think I'll start with this. A lot of the original chatbots

are one solution for a number of things. With DeepSeek, it came out of a quant lab. So basically, it was tuned for very specific types of logic, right? Which means it doesn't need a whole amount of data. It knows to make the most with what it has. And that speeds up

the processes and the training and it also gives you I think more accurate solutions. The one thing that stood out to me was in terms of coding, they compared it with a number of tools and Deepsea consistently provided code that was good to run without any debugging needed. Whereas other solutions like the latest GPT still had a bit of a clunkiness to the code they produced.

DeepSea kind of leapfrogged them in that respect. And code is actually a good test to see that it's good in math, it's good with that sort of logic. So it's not going to be a one-size-fits-all, but for the things that matter to businesses, I think it's spot on for now. Yeah. Yeah.

Yeah, there's, I mean, there's, this has come out, a lot's happened in a very short space of time. So we had, we had 01, which came out at the end of last year, that was kind of the first reasoning model, quote unquote, that hit the market. And that basically just means something that has to stop and think a bit more about, you know,

what the answer could be. One of the examples I was reading was if someone said to a human, what's the capital of France? Capital city, you would say Paris pretty quickly. But if someone said, you know, what is the

second biggest city in France, you might be like, oh, you have to stop and think. You might be like, Lyon might be up there, Nice, Marseille, and you'd have to kind of think through what the answer might be. And then you eventually be like, through kind of logical reasoning, you might figure out what the answer is. Google came out with a reasoning model called Gemini Flash Thinking in December. OpenAI then came out with

uh oh three a few days later um china's version of amazon alibaba they released a new version of quen chatbot qwq with the same reasoning capabilities as well so i mean this this r1 projects hit the market but there are a lot of others out there grace i wonder what you think of this because um

I thought when this came out, I was like, okay, this is probably going to be a rallying cry for the president. He's already called it a wake-up call and big tech using it to kind of make folks a bit more nervous about what China's doing and letting them charge ahead with AI investment, kind of shooting also shooting down any legislation that might get in the way. DeepSeek, their privacy policy makes sense.

It makes it pretty clear that the company stores the information that they collect in secure servers located in the People's Republic of China. That was the fear over what TikTok was doing, collecting a ton of Americans' data and with the Chinese government, collecting a ton of Americans' data, which the Chinese government could potentially allegedly have access to, and is the reason that their future in the US is in limbo. What do you make of how...

successful R1 might be allowed to be in this country given America's reaction to TikTok. I think we're seeing really mirrored issues between TikTok and DeepSeek. DeepSeek is storing all user data on servers in China, which doesn't necessarily mean that the Chinese government, it doesn't mean that the Chinese government is getting instant access to everything that's going through, but it does mean that the Chinese government would have a much easier time

requesting and accessing that user data from a company. We're getting some issues reported from users with censorship in terms of being able to get answers to questions about territorial disputes with the South China Sea, questions about the Chinese government.

I think that one thing that could be a block is being able to access the same level of success that OpenAI has. And some AI companies have had with large enterprise clients in the U.S. and with the U.S. government. OpenAI has contracts with the Department of Defense. The U.S. Navy has already banned use of DeepSeek.

It's already been pulled from app stores in Italy, App Store and Google Play Store. They're already launching a probe into it with concerns over data privacy. So I think that that is going to be very easily a big block. I think in terms of seeing a ban

With the new administration, I think there would be a newer motivation. With the Biden administration, I think that a clampdown would come more from the angle of wanting guardrails to protect user privacy. With the Trump administration, I think it would come more from the angle of wanting to protect U.S. innovation. I think it's likely that we will see some motions to control or maybe ban deep-seek. I don't think they will move that fast. We've seen how long the process has been

with TikTok. It's been a long, long time coming. It might be a lot longer. But it's already causing issues and it's been mere days. I was reading that Dr. Richard Whittle from the University of Salford in England was saying that he had various concerns about data and privacy with the app, with DeepSeek, but said that there were plenty of concerns with the models used in the US as well, because they are both

hoovering up an incredible amount of data about folks. Gajo, let's close out with this. How does this change things moving forward? Given everything we've talked about, everything that's happened, what are you looking for? What are you paying most attention to in terms of how you expect the landscape to shift as a result of this R1 model? Okay, so we know that NVIDIA took a hit, huge hit.

But I think in the long term, it's going to be open AI that will have to be on the defensive. They are the leading AI company. And now they have a rival that's more affordable, open source, easier to adopt and use. And so they've just started to charge, I think, $200 for their pro plan. And performance-wise, people might just say,

Do I really have to pay that much? So they've spent the past year convincing us that you have to pay for access to AI. And they're losing money. And they're losing money. And now we have a situation where we might just have a new model to play with every month, seeing how level the playing field is. So it's going to be difficult for them to, even if they're, say, the next GPT,

is really good, do we really need really good or just good enough, right? Yeah. And DeepSeek has shown us that it's really good. So that bar has kind of been changed. And I think other AI companies will have to adjust as well, whether to lower the cost or...

or cost of admission to their products and services. Since a lot of them are shifting towards AI agents, I think that area is safe, especially since these are focused on enterprise and security and privacy is addressed because all that information is kept in the US or within those company servers. But for general use, I could see DeepSeek becoming

Basically what OpenAI was two years ago. Something everybody wants a piece of. Yeah. Grace, how about for you? How does R1 change things moving forwards?

Well, we've already gotten a response from OpenAI. Sam Altman said on X, I think it was yesterday, that they will deliver much better models, that it's legitimately invigorating to have a new competitor, whether or not that's a genuine statement. Yeah, I think it's a really big pressure to step things up. Like Adjo said, I mean, they're losing money on their pro subscriptions. They're $200 a month.

They, you know, OpenAI has its own issues with whether or not, you know, it's using users data for model training and whether or not users can consent or not consent to that. But yeah, I think it really shakes up the idea that if you want a better model, you have to pay more money for it and we have to pay more money for it. I thought it was quite ironic that so the question being, did the restriction of AI chip exports to China backfire?

because Mark Cieslak of the BBC was pointing out that all of this has been achieved using lower end technology, at least reportedly from them, from Deepsea, because of chip export restrictions. Those restrictions may have initially helped to cause this AI bubble, but they're also what led to the structural integrity of the AI bubble in the US being called into question. So it's quite

And China has its own successful companies, whether they have big funding or little funding. I mean, you know, they have pretty powerful AI companies with backing from Alibaba or Baidu. You know, their AI chatbot has the biggest market share for search engines bigger than Bing. There's a lot of pure play startups there. So even if you're just looking within China, whether or not they're relying on U.S. chips, there's a really good market there already for U.S. or for, excuse me, for AI developers. Yeah.

The Economist writing the competition nipping at American AI heels may yet spur it to greater things. We shall see. That's all we have time for this episode. Thank you so much to my guests for hanging out with me today and explaining what the hell is going on with R1. Thank you, First of Garjo.

Thanks. This was fun. Yes, sir. Thank you to Grace. Thanks, guys. Nice to talk with you. Yes, indeed. And thank you to the whole editing crew, Victoria, John Lance and Danny, Stuart who runs the team and Sophie does our social media. Thanks to everyone for listening in to an eMarketer video podcast made possible by Zeta Global. Tune in Wednesday for the Reimagining Retail Show where the gang will be discussing TikTok shop.