Why everyone is talking about China's DeepSeek

Summary

POLITICO Tech: In-Depth Summary of "Why Everyone is Talking About China's DeepSeek"

Podcast Information:

• Title: POLITICO Tech
• Host: Stephen Overle
• Episode: Why Everyone is Talking About China's DeepSeek
• Release Date: January 28, 2025

---

1. Introduction to DeepSeek's Breakthrough

In the January 28, 2025 episode of POLITICO Tech, host Stephen Overle delves into the recent surge of attention surrounding China's DeepSeek and its latest artificial intelligence (AI) model. Over the weekend, Silicon Valley experienced significant unease as DeepSeek unveiled an AI model comparable to top U.S. rivals but at a substantially lower cost. This development has been likened to China's interpretation of a "ChatGPT moment," signaling a robust assertion of its AI capabilities to the global community, particularly the United States.

Key Points:

• DeepSeek's new AI model matches U.S. counterparts in performance.
• Achieves this parity at a fraction of the cost.
• Sparks concern over U.S. technological dominance and market stability.

---

2. Market Impact and Stock Reactions

The introduction of DeepSeek's AI model had immediate repercussions in the financial markets. Notably, tech stocks experienced a significant downturn, with chipmaker Nvidia enduring the largest single-day loss in market history. This volatility underscores the perceived threat DeepSeek poses to established U.S. tech giants and the broader implications for the technology sector.

Notable Quote:

“DeepSeek's breakout arrival scrambles all sorts of preconceptions about US China tech competition.”
— Stephen Overle (03:01)

---

3. Analyzing China's Position in the AI Race with Bill Drexel

Stephen Overle engages Bill Drexel, a tech and national security fellow at the Center for a New American Security, to unpack the ramifications of DeepSeek's advancements.

a. Is China Beating the U.S. in AI?

Bill Drexel assesses whether China's latest AI model signifies a broader victory over the U.S. in AI development.

Key Insights:

• Overall Superiority: Drexel believes China has not surpassed the U.S. in AI overall. "Are they beating us in AI in general? I think the answer on that is no.” (02:19)
• Frontier Models: China hasn’t overtaken the U.S. in frontier AI models but is significantly closing the efficiency gap.
• Efficiency Gains: DeepSeek's model highlights China's ability to optimize microchip usage, reducing training costs and time.

Notable Quote:

“They are definitely, if not beating us, giving us a serious run for our money on efficiency in frontier models.”
— Bill Drexel (02:19)

b. China's Business Model and AI Ecosystem

Drexel draws parallels between DeepSeek's approach and China's broader business strategies, emphasizing state support and cost-competitive practices.

Key Insights:

• State Support Uncertainty: The extent of China's government backing for DeepSeek remains unclear, though political motivations are evident given the model's release timing.
• Innovative Efficiency: DeepSeek isn't merely replicating existing models cheaply; they have pioneered new methods to enhance efficiency.
• Local Talent: The company's workforce is predominantly native Chinese, indicating a robust and self-sufficient AI talent pool without reliance on overseas experts.

Notable Quote:

“They have pioneered some new methods and some new kind of ways to reach this new level of efficiency, and that's different.”
— Bill Drexel (05:35)

c. Effectiveness of U.S. Export Controls

The conversation shifts to the impact of U.S. export restrictions on China's AI progress.

Key Insights:

• Delayed Impact: Current export controls haven’t yet significantly hindered China's AI advancements.
• Future Challenges: The next generation of microchips, which China is increasingly restricted from accessing, will likely exacerbate the technological gap.
• Enforcement Concerns: There are uncertainties about the effectiveness and enforcement of these controls on companies like DeepSeek.

Notable Quote:

“Those controls, they basically have a pretty long delay in terms of when they take effect.”
— Bill Drexel (07:23)

---

4. Authoritarianism vs. Democracy in AI Development

A provocative discussion arises around whether authoritarian governance systems inherently offer advantages in AI development over democratic ones.

Key Insights:

• Surprised Reaction: Drexel expresses surprise at a comment suggesting communism outpaces capitalism in AI creation, highlighting the complexity of attributing AI success to political systems alone.
• Multi-Dimensional Factors: AI competitiveness depends on several dimensions, including compute power, algorithms, institutional efficacy, and energy infrastructure.
• Institutional Advantages: The U.S. excels in rapidly commercializing innovations, while China's strengths lie in scaling technologies efficiently.

Notable Quote:

“I think a useful paradigm is that it's up to four dimensions.”
— Bill Drexel (09:17)

---

5. Policy Responses and Strategic Recommendations

In light of DeepSeek's advancements, Drexel offers strategic advice to policymakers on bolstering the U.S. position in AI.

Key Recommendations:

• Enforce Export Controls: Strengthen the implementation of existing restrictions to impede China's access to advanced microchips.
• Diversify AI Investments: Balance investments in large-scale AI training runs with enhancements in algorithmic efficiency.
• Expand AI Scope Beyond Frontier Models: Explore and invest in non-frontier AI applications that could offer economic and strategic advantages.
• Promote Institutional Efficiency: Enhance the U.S. ability to translate AI innovations into practical applications swiftly.

Notable Quote:

“We need to do both. We need to compete in all areas.”
— Bill Drexel (13:53)

---

6. Consumer Behavior and National Security Implications

The episode examines how consumer preferences intersect with national security concerns, particularly with DeepSeek's app topping the Apple App Store.

Key Insights:

• Consumer Preference for Quality: Regardless of origin, consumers prioritize the best and most cost-effective technologies.
• Strategic Challenge: The U.S. faces a dilemma in promoting its technology while countering competitive pressures from China.
• Call to Action: To maintain technological leadership, the U.S. must prioritize creating superior products that naturally attract consumer adoption.

Notable Quote:

“We just need to be better.”
— Bill Drexel (16:41)

---

7. Conclusion: Building Competitive AI Ecosystems

Stephen Overle concludes the episode by reinforcing the importance of strategic competition and the need for the U.S. to enhance its AI development framework. The discussion with Bill Drexel emphasizes that while China's DeepSeek presents a formidable challenge, it also serves as a catalyst for the U.S. to refine its AI strategies and policies to ensure continued leadership in the technological arena.

Final Takeaway:

• Focus on Excellence: Success hinges on the ability to produce superior, cost-effective AI technologies that meet consumer needs.
• Comprehensive Strategy: Addressing AI competition requires a multifaceted approach encompassing policy enforcement, investment diversification, and institutional efficiency.

---

This episode of POLITICO Tech provides a comprehensive analysis of China's DeepSeek AI model, its implications for U.S. technological and economic landscapes, and strategic pathways for maintaining competitiveness in the evolving AI race.

Transcript

[American Express]

Flex your business with an American Express Business Gold Card. You'll earn four times Membership Rewards points on your top two eligible spending categories like transit and electronics each month on up to 150k in combined purchases per year. Plus, you can now earn 3 times Membership Rewards points on flights and prepaid hotels booked on amextravel.com terms and points cap. Apply learn more at americanexpress.com business gold amex business gold Built for business by.

[Stephen Overle]

American Express hey, welcome to Politico tech. Today's Tuesday, January 28th. I'm Stephen Overle. Over the weekend, Silicon Valley had a collective freakout. A Chinese company called Deepseek released a new artificial intelligence model on par with top US Rivals, and it appeared to do it at a fraction of the cost. For China, this is a chatgpt moment, a clear flex to the rest of the world and the US in particular, that its AI shouldn't be underestimated. And for the US It's a major wake up call that trade restrictions have not hampered China's AI ambitions and that high flying American companies are not invincible. In fact, tech stocks took a big hit on Monday, most notably chipmaker Nvidia, which lost more value in a single day than any stock in the history of the market. DeepSeek's breakout arrival scrambles all sorts of preconceptions about US China tech competition. So I called up Bill Drexel, a tech and national security fellow at the center for a New American Security. On the show today, Bill breaks down the significance of Deepseek, starting with whether it means China is winning the AI race. Here's our conversation. Bill, welcome back to Politico Tech.

[Bill Drexel]

Thanks for having me. Great to be here.

[Stephen Overle]

Let's start with the question that everyone is debating right now. Does this new AI model mean that China is beating the US at AI?

[Bill Drexel]

Well, it's a big question. I think that it definitely sends some sort of shot across the bow, but I'm not sure that it's we can split that question in a few things. First, are they beating us in AI in general? I think the answer on that is no. And it's hard to even talk about that because AI is such a kind of varied and diverse field. Are they even beating us in frontier models? I think the answer to that is also no, still no. But we can say that they are definitely, if not beating us, giving us a serious run for our money on efficiency in frontier models. And that's a really important aspect of this overall competition.

[Stephen Overle]

Well, that seems to me from everything I've seen to be kind of the big breakthroughs here are Deep Seq was able to use microchips more efficiently and sort of train AI more quickly. And it was able to do all of that pretty cheaply based on what we know about this model. Are those the big breakthroughs as you see them? And for the non techie crowd, why does that matter so much?

[Bill Drexel]

Yeah, I think, I think you're right on. They really hit the United States frontier AI ecosystem where it hurt, which is especially cost efficiency. So even before this sort of deep sea freakout, it was the case that we were investing incredible amounts of capital into these massive training runs to kind of birth these new AI models that are incredibly resource intensive. And investors were already getting a little bit nervous whether the juice was worth the squeeze on this. So there was this question of if we're funneling tens, hundreds of billions of dollars into these models, are they actually going to contribute that much back? And so that was the kind of existing insecurity. And now Deepseek comes along and just very quickly after our kind of industry leader comes out with a model, comes out with a fairly comparable model at a small fraction of the cost and offers it at a small fraction of the cost. So that really is going to, it hits the business models hard. It hits kind of our expectations of, of how, what our strategy is for this competition pretty hard. So it's very important in that respect.

[Stephen Overle]

In some ways that sounds to me similar to kind of the Chinese business model in other areas as well. Right. China is known for kind of funding the development of technology and, and products in general and then sort of selling them low cost globally.

[Bill Drexel]

Yeah, I think it's, it's definitely, it's in that mold. So you're right that it follows a certain pattern in that regard. I guess two points off the bat from that one is that it's unclear exactly how much or what sort of state support they're getting. And I'm sure that we're not going to really know because obviously China has a really vested interest in this competition. And at least from the timing that they released this model during inauguration week seems to suggest that there's politics playing a role here and there's actually a lot of debate about exactly what resources they are using to, to get these results.

[Stephen Overle]

Right.

[Bill Drexel]

So you're right that it's, it's China doing a similar thing at a much lower price. What is interesting about this and, and should be a little concerning to national security policymakers to, to those in AI labs, etcetera, is that they've really done some, some, from what I can tell, really serious conceptual innovation here. So it's not just that they are following on the coattails of an American leader and just making it way cheaper. It is the case that they have pioneered some new methods and some new kind of ways to reach this new level of efficiency, and that's different. And it's also notable that deepseek is sort of remarkable in how native Chinese it is, how local it is. There aren't a lot of engineers in Deep Seq who return from abroad, which often when we talk about sort of the China's copycat model, it's that you have people coming from abroad who come back and kind of replicate what they got. In this case, at least, the CEO, I believe, reported that in their version 2 model, they didn't have anyone, no one working on it, who returned from China from overseas. So this does suggest that China's AI ecosystem, in terms of algorithm talent, is natively really competitive. So that's something to watch.

[Stephen Overle]

Part of the surprise does stem from the fact that the US has imposed a variety of sanctions on China, most notably restrictions on powerful microchips that are needed to develop AI. This Deep SEQ model suggests maybe that hasn't inhibited China as much as people thought. Does it mean those policies failed?

[Bill Drexel]

Right, so that's, that's kind of one of the big questions circling around here. And from all that I can tell, it's actually not suggesting that those export controls aren't working. It is suggesting they just haven't really bitten yet. So those controls, they basically have a pretty long delay in terms of when they, when they take effect. The chips that were purportedly used to train the model that we're discussing aren't that different from what peer competitors were using. The next generation of chips that China is in theory going to be blocked from accessing, at least within its own borders, that differential will be much wider. And we have indications too. I mean, the CEO of the company has been open about the fact, and there are other indications from Deep Seq as well, that compute is their largest constraining factor going forward. Now, all that to say that's if export controls work. There's also this other issue, which is are they actually going to be enforced in a way that keeps them away from companies like DeepSeq? We don't know that either, but at least in theory, this kind of stage in the competition hasn't really been hit too hard by export controls yet. It's what we'll be seeing in the next couple of years where we'll see how successful or unsuccessful that's been.

[Stephen Overle]

You know, I got a text message from someone in tech about the deep SEQ news who said, this sort of stuck with me who said, at some point we need to admit communism is better than capitalism at building AI. I wonder if the real question there is whether authoritarianism is more efficient than democracy in some ways at facilitating AI. But regardless of how you frame it, is that sort of a hard truth or a hard reality of the competition with China that we should be discussing at this point?

[Bill Drexel]

I wouldn't say so. That quote actually surprises me a little bit. A couple things here. I mean, they've tried to send a signal with this model based on its timing, and it's very similar actually in that sense to, you know, the big kind of freak out. When Huawei launched that phone, when Secretary Raimondo was visiting, and initially everyone was like, oh man, they're gonna surge ahead with their own chips. And then, you know, in the weeks after things cooled off a little bit, their yield wasn't that great. Concerns about production delays, et cetera. It was still an advancement, it was still closing the gap. But I think right now, particularly with the stock market behaving the way it is, we're really at peak freak out. We can say this, they've gone a lot faster than we expected. They've closed more gaps than we anticipated. But it's still going to be difficult for them to compete in the years ahead, particularly with the chip technology differential. That said, what this shows, I think, is that they're hungry. They've been fiercer than we have in pursuing more efficient algorithms while we've been really laser focused on these sort of massive fundraising rounds, kind of hypnotized by this idea that if we just do a huge moonshot training, that that fixes everything. We need to do both. We need to compete in all areas. I think, to the question of which system is actually better for developing frontier AI, you kind of have to look at it in a few more dimensions. I think a useful paradigm is that it's up to four dimensions. So you have your compute. And like we say, we're still waiting for our actions on that to really bite. You've got algorithms. So, yes, this does show that in some ways the Chinese are. They certainly seem to want it more. They're pursuing things faster on that front than we are. Another major thing that matters is institutions. And that's where I think the system of government matters more. So how quickly can you translate innovations to practical applications and in which areas? And that's where, you know, you can make cases in both directions. You can make cases that actually the US ecosystem is really, really effective at commercializing innovations quickly. And there's been some work by Jeffrey Ding and others that suggest that our system has a, has an edge there. There is the counter argument that we've seen time and again that competition in China is so fierce in tech that they've historically been less adept at going from 0 to 1, but scaling 1 to 100, they're really excellent. So I think the jury's still out there. Oh, energy. Energy is the fourth component there and there. It's true, it depends what we do. But there are certain advantages to a more autocratic system where they can just pursue what they want to do without a lot of red tape on massive energy infrastructure.

[Indeed]

This episode is brought to you by Indeed. When your computer breaks, you don't wait for it to magically start working again. You fix the problem. So why wait to hire the people your company desperately needs? Use indeed sponsored jobs to hire top talent fast. And even better, you only pay for results. There's no need to wait. Speed up your hiring with a $75 sponsored job credit@ Indeed.com podcast. Terms and conditions apply.

[Stephen Overle]

I'm curious. You know the freak out about Deepseek and its new model really started in Silicon Valley. We saw that hit Wall street, obviously in terms of stock prices. Inevitably it will hit Washington. How do you predict policymakers will respond and how should they respond?

[Bill Drexel]

I think that this is going to wake up Washington to the fact that the enemy really gets a vote on the AI race. I think it should really mobilize a lot of attention on this issue. I think the, the most strategic, the wisest thing to focus on in the immediate would be to make sure that the enforcement of existing export controls is sufficient to actually have the effect that we're hoping for. That's one area that I imagine could use some investment. I also worry though that we may draw some of the wrong lessons from this.

[Stephen Overle]

Like what?

[Bill Drexel]

So you could say this kind of series of events may suggest that US companies have been over fixated on these really capital intensive moonshot training runs. And as, as we discussed earlier, that could mean, okay, we need to do that in addition to algorithmic enhancement, working on, on greater efficiency with those big capital intensive runs. Another way you could interpret it and, and I'm also very sympathetic to this argument is that we might just be over fixated on these Frontier models as kind of the end all, be all of the AI competition. Right. So if China suddenly looks more competitive in frontier models and we were already over fixated on these big moonshot frontier models, we might become more fixated on them to the exclusion of other forms of AI that might actually prove more decisive on the battlefield economically, et cetera. So to give you some examples, you know, a lot of this frontier fixation is, is in the hopes of, of achieving this sort of artificial general intelligence superintelligence, which would be a multiplier for everything. Right. But it's still, you know, the proof is in the pudding. We don't exactly know how true or in what ways that may or may not be true. And so it may be the case that China is pursuing a more diversified approach to in other forms of AI that could actually prove decisive. And if we get over fixated on one kind of holy grail and just pursue that with everything we've got, and then it ends up not panning out, they might have a more diversified portfolio in the end. So I think that yes, we need to be more serious about algorithmic efficiency while we're also pursuing these moonshot runs, but we also need to continue thinking about what's the non frontier AI that's nonetheless extremely important and where we need to be ahead of China.

[Stephen Overle]

Let me ask you one more question here, Bill. One of the most significant takeaways to me in some ways is that over the weekend the Deep Sea app became the number one app in the Apple App Store. And we've been talking a lot on the podcast recently about the whole TikTok ban. Consumers who were leaving TikTok going to RedNote, another Chinese app. And I sort of take these together and I can't help but think like the average American does not really care whether their technology was developed in China or not. And I don't know how you reconcile that with concerns about national security or US competitiveness, et cetera, except it seems like the only way to win is to just build the best technology so that people use it.

[Bill Drexel]

Yeah, I think that's exactly right. I think that at the end of the day it will be a massive strategic problem for the United States and a massive hit to human flourishing. If China kind of has the strongest AI ecosystem in the years ahead. And we can say that in the abstract, but it's such a bird's eye level view on the day to day. The tool that works best is the one that you're going to want to use, or the tool that's most cost effective. Right.

[Stephen Overle]

Right.

[Bill Drexel]

And so that's the goal. We just need to be better. And the kind of arguments that it's from China, it's hard to get that across and it's easier to get that across in the United States than it is in the rest of the world.

[Stephen Overle]

Right, Right.

[Bill Drexel]

We're really fighting an uphill battle if we think we can convince everyday folks that just because a model or a tool is from China, they shouldn't use it. Maybe with some very particular exceptions, but broadly speaking, they're going to go for what's best and most cost effective. And so we just need to find ways not only to curb their development where we must, but also just boost ours as much as we can.

[Stephen Overle]

Well, Bill, appreciate you being back here on Politico Tech.

[Bill Drexel]

Thanks for having me. And always a pleasure to talk.

[Stephen Overle]

That's all for today's Politico Tech. If you enjoy Politico Tech, be sure to subscribe. And for more tech news, subscribe to our newsletters, Digital Future Daily and Morning Tech. Our managing producer is Annie Reiss. I'm Stephen Overle. See you back here tomorrow.