Weekly: Do export controls work?; DeepSeek and the future of AI; Is DeepSeek overhyped?
7 min read.
Highlights
More debate on the chip ban. John Villasenor of Brookings argues that the export controls have forced Chinese chip companies to be more innovative. Anthropic founder and CEO Dario Amodei argues that the export controls are a vital tool for the US to leverage all the advantages it has. We’ll never know the counterfactual.
Though, I don’t think anyone is arguing for the lifting of the bans. Usually the conclusion (on both sides) is that the only way to win is to just make the better product. And as I’ve argued before, it seems to me that to make the better product, you can just do both - invest in domestic innovation AND be selective about who has the best and the most AI chips.
On Jevon’s Paradox and what DeepSeek means for the future of AI industry. Two indispensable semiconductor newsletters, SemiAnalysis and Fabricated Knowledge, cover this in more depth. I think the common take is that Jevon’s is “overhyped,” though cheaper AI will probably drive greater consumption, with the caveat that the timing of abundant supply and subsequent demand has to be just right for it to work in the way that, say, oil did.
Lastly, on whether DeepSeek really is that big of a deal. As time passes and more details and more analysis comes out, the answer seems to be, Not as much as we initially thought. Probably. It’s still too early to tell how this all shakes out, but already, we know that the $6M tag on DeepSeek is a misleading underestimate. DeepSeek also lags 8-10 months behind other reasoning models, which is a long time in light-speed AI development. Yes, Nvidia dropped 17% in one day, but NASDAQ dropped 3%, and since then, the stocks have been slowly recovering.
There are other interesting things like the fact that it’s open-source and that they did it under export controls and that it seems attuned to China’s political sensitivities, but we’ll see how important they turn out to be.
What DeepSeek is, then, is an important moment to step back and reconsider the assumptions we’ve been pricing into the AI hype. We’ll see where it goes from there.
Thanks for reading.
Table of Contents
Dario Amodei, “On DeepSeek and Export Controls,” Dario Amodei Blog, January 2025.
John Villasenor, “DeepSeek shows the limits of US export controls on AI chips,” Brookings, 01/29/2025.
Dylan Patel, AJ Kourabi, Doug O'Laughlin and Reyk Knuhtsen, “DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts,” SemiAnalysis, 01/31/2025.
Doug O’Laughlin, “DeepSeek: Is this Jevon's Cope?,” Fabricated Knowledge, 01/30/2025.
Robert Armstrong, “What DeepSeek’s AI really means for the market,” FT, 01/28/2025.
FT Editorial Board, “DeepSeek defies America’s AI supremacy,” FT, 01/28/2025.
H. Andrew Schwartz and Gregory C. Allen, “DeepSeek Deep Dive Podcast,” CSIS, 01/29/2025.
1.
Dario Amodei, “On DeepSeek and Export Controls,” Dario Amodei Blog, January 2025.
A few weeks ago I made the case for stronger US export controls on chips to China. Since then DeepSeek, a Chinese AI company, has managed to — at least in some respects — come close to the performance of US frontier AI models at lower cost.
I'll focus on whether DeepSeek's releases undermine the case for those export control policies on chips. I don't think they do. In fact, I think they make export control policies even more existentially important than they were a week ago.
Export controls serve a vital purpose: keeping democratic nations at the forefront of AI development. To be clear, they’re not a way to duck the competition between the US and China. In the end, AI companies in the US and other democracies must have better models than those in China if we want to prevail. But we shouldn't hand the Chinese Communist Party technological advantages when we don't have to.
2.
John Villasenor, “DeepSeek shows the limits of US export controls on AI chips,” Brookings, 01/29/2025.
U.S. AI export control rules are designed to impede China’s AI progress, but they may actually be accelerating it.
AI engineers in China are innovating on ways to use limited computing resources more efficiently.
Overly broad export controls on advanced computing chips can harm U.S. companies by reducing global sales opportunities while doing little to enhance U.S. AI leadership.
All of this illustrates that the best way for the U.S. to maintain AI leadership is to outrun the competition through the combination of domestic investment and an innovation-friendly AI regulatory climate. Trying to stay ahead by tripping up rivals can have the opposite of its intended effect.
3.
Dylan Patel, AJ Kourabi, Doug O'Laughlin and Reyk Knuhtsen, “DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts,” SemiAnalysis, 01/31/2025.
The company is not new, but the obsessive hype is. SemiAnalysis has long maintained that DeepSeek is extremely talented and the broader public in the United States has not cared. When the world finally paid attention, it did so in an obsessive hype that doesn’t reflect reality.
The narrative now is that DeepSeek is so efficient that we don’t need more compute, and everything has now massive overcapacity because of the model changes. While Jevons paradox too is overhyped, Jevons is closer to reality, the models have already induced demand with tangible effects to H100 and H200 pricing.
We believe DeepSeek has access to around 10,000 of these H800s and about 10,000 H100s. Furthermore they have orders for many more H20’s, with Nvidia having produced over 1 million of the China specific GPU in the last 9 months.
Our analysis shows that the total server CapEx for DeepSeek is almost $1.3B, with a considerable cost of $715M associated with operating such clusters. Similarly, all AI Labs and Hyperscalers have many more GPUs for various tasks including research and training then they commit to an individual training run due to centralization of resources being a challenge.
DeepSeek’s price and efficiencies caused the frenzy this week, with the main headline being the “$6M” dollar figure training cost of DeepSeek V3. This is wrong… We are confident their hardware spend is well higher than $500M over the company history.
4.
Doug O’Laughlin, “DeepSeek: Is this Jevon's Cope?,” Fabricated Knowledge, 01/30/2025.
As intelligence gets cheaper, we will throw more brute-force intelligence at every one of the world’s key problems. In a world with massively lower costs, instead of becoming more efficient, we often throw more resources at the problem. This is the thrust of the Jevons Paradox as a bull case and what will almost assuredly happen in the long run. It took ~16,000 transistors to get a rocket on the moon. Now, ~16,000 transistors probably could run any modern application.
There is one little problem. Lead and lagging times can create kinks, producing very drastic scenarios. Maybe the Jevon paradox is correct, but the issue is not if we will use more, but a bit about timing. That is an example of a supply-and-demand overshoot that the market is quickly contemplating. The terminal value is acceptable for everyone involved, but the pesky short run escapes us.
In this case, the bear case is that yes, Jevons happens, but supply overwhelms demand so much in the short term that the rapid addition of supply overwhelms the long-term demand.
The reality is likely never as bad as feared and never as good as dreamed.
5.
Robert Armstrong, “What DeepSeek’s AI really means for the market,” FT, 01/28/2025.
What happened in markets yesterday was not an out-of-nowhere surprise, nor a panic, nor a bubble popping. It was the pricing in of slightly higher odds that AI is not a winner-take-all game.
It was always conceivable that the AI revolution would be like the invention of the automobile or the aeroplane. Those revolutions did lead to the creation of huge, durably profitable companies, but also lots of competition, ensuring that much of the value created went to consumers rather than shareholders. Now that looks like a more significant possibility for AI.
Nvidia itself is hardly on its knees. Its proprietary coding language, Cuda, is still the industry standard. And just because the DeepSeek model is more efficient does not mean that leaner models will not benefit from the higher computational power offered by Nvidia’s best chips. While its shares dropped nearly 17 per cent yesterday, that only brings it back to the (very, very high) level of September.
We should not overstate the market’s reaction to R1. The Nasdaq fell 3 per cent. That was a bad day, not a panic. The winner-take-all view of AI is wounded, not dead. The AI bubble, if that’s what it is, may yet pop, but it didn’t pop yesterday. Not even close.
6.
FT Editorial Board, “DeepSeek defies America’s AI supremacy,” FT, 01/28/2025.
America’s superinflated tech stocks had seemed due for a correction for months, but the trigger has come from an unexpected source. The latest large language model from China’s artificial intelligence start-up DeepSeek may not be quite a “Sputnik moment”.
Like Meta of the US but unlike OpenAI or Google’s Gemini, it is open source — ready to share the recipe for its secret sauce rather than keep it locked away in hope of extracting maximum financial gain. That makes it appealing for developers to use and build on.
Chinese AI start-ups have been compelled to find inventive ways of extracting the most juice from the chips they do have. Far from stifling Chinese innovation, Washington may have stimulated it.
The open question now is not necessarily who will develop the best AI models but who can apply them best to real-world tasks. Kai-Fu Lee, a Chinese AI pioneer, has long argued that China excels on the application front even if it may lag behind in infrastructure. That was before the Chinese start-up world was squeezed by the political clampdown on tech entrepreneurs and the surge of investment in US AI start-ups. But after DeepSeek’s achievement, it looks a much more even game.
7.
H. Andrew Schwartz and Gregory C. Allen, “DeepSeek Deep Dive Podcast,” CSIS, 01/29/2025.
In this crossover episode with Truth of the Matter, we discuss the origins of Chinese AI Company DeepSeek (0:55), the release of its DeepSeek R1 model and what it means for the future of U.S.- China AI competition (3:05), why it prompted such a massive reaction by U.S. policymakers and the U.S. stock market (14:04), and the Trump administration's response (24:03)