Latest / China Economy with Fexingo: Beijing's Policies, Trade, and Asia-Pacific Markets / How Chinas AI DeepSeek Is Upending the Model Cost Curve
Transcript
- Lucas: There's a number that's been stuck in my head this week, and it's five point six million dollars. Luna: That's the cost to train a large language model, I'm guessing? Lucas: Exactly. Specifically, that's what DeepSeek — a Chinese AI lab you may not have heard of — says it spent training their V3 model. To put that in perspective, a frontier model from OpenAI or Google can run a hundred million dollars or more. So DeepSeek is claiming they did it for roughly five percent of the cost. Luna: That's a massive gap. Is that real, or is it creative accounting? Lucas: That's the right question. DeepSeek published a technical paper with enough detail that independent researchers have done reality checks. The consensus so far is that the number is credible, and the trick is architecture. They're using a 'mixture of experts' approach — MoE — where only part of the model activates for any given input. That slashes the compute required. Luna: Right, so instead of firing the entire brain every time, you only use the relevant neurons. Lucas: Exactly. It's an idea that's been around in academic papers for years, but DeepSeek implemented it in a particularly efficient way. They also optimized their memory usage during training — things like using low-precision arithmetic more aggressively than most labs do. Luna: So this isn't really a breakthrough in fundamental AI capability. It's more of a cost breakthrough. Lucas: That's a useful distinction. The model's performance is competitive with GPT-4 and Claude on many benchmarks, but it's not necessarily better. The real story is the cost curve. If you can train a model that's good enough for a fraction of the price, it changes the economics of the entire industry. Luna: And it also has implications for the export control debate. The US has restricted high-end chips like NVIDIA's H100 going to China. But if Chinese labs can achieve strong results with less advanced hardware, or with clever software tricks, those controls become less effective. Lucas: Right. DeepSeek used NVIDIA H800 chips, which are a reduced-bandwidth version allowed under export rules. They used two thousand of them for about two months. That's a lot of chips, but it's not the ten thousand plus clusters that US hyperscalers deploy. Luna: So they got around the spirit of the controls, if not the letter. And now they're publishing the recipe for free. Lucas: The model is open-weight, which means anyone can download it and fine-tune it. That's a deliberate strategy. DeepSeek is essentially saying: we don't need to sell a product. We just need to make AI cheap enough that it becomes a commodity, and then the scale of China's manufacturing ecosystem takes over. Luna: That's a very different mindset from OpenAI, which is trying to build a moat around its technology. One is playing the pricing game, the other is playing the performance game. Lucas: And in the long run, the pricing game has a track record of winning. Think about what happened with solar panels. China didn't invent the technology, but they drove the cost down so aggressively that now they dominate global production. Luna: I've been thinking about this same dynamic. And it makes me wonder — if AI compute costs plummet, what happens to the investment thesis for the big cloud providers? Amazon and Microsoft have been spending billions on data centers, betting that demand for AI inference will be enormous and profitable. Lucas: That's the key question. If inference becomes cheap, margins compress. The cloud business might end up looking more like a utility than a high-margin tech service. That's a huge shift for the stock market, where these companies are valued partly on the promise of ai driven growth. Luna: And it also changes the trade story. The US trade deficit in goods has been widening — the latest data shows seventy-seven point six billion in May. If China can export cheap AI capability as a service, that could show up in the services trade balance too, which is traditionally a US surplus. Lucas: The yuan has been weakening slightly, down to 6.78 per dollar. That makes Chinese exports in general more competitive, and digital services are no exception. If a Chinese company can offer AI inference at half the cost of a US provider, the dollar strength actually amplifies that advantage. Luna: So we have a convergence: a US trade deficit that's growing, a dollar that's still relatively strong, and a Chinese AI sector that's learning to do more with less. That's a potent combination. Lucas: And the political response is uncertain. The previous administration focused on hardware export controls. But if the bottleneck shifts from chips to algorithm efficiency, then the leverage point changes. The US might need to think about restricting software or know-how, which is much harder to enforce. Luna: That's a more complicated regulatory challenge. And it's one that cuts against the open-source culture that a lot of US AI researchers value. Lucas: Yeah, it's a real tension. You don't want to stifle innovation at home, but you also don't want to hand your competitors a playbook. DeepSeek's paper is essentially a playbook. Luna: Speaking of the value of open sharing — that's one reason we keep this podcast ad-free. We think having clear, specific conversations about the economy without sponsored interruptions is worth preserving. Lucas: I agree. We deliberately don't run ads. If that kind of independence matters to you, you can support it at buy me a coffee dot com slash fexingo. No pressure, just if you find the show useful. Luna: Exactly. And that support is what keeps us free to follow stories like this one — where the details really matter. Lucas: So back to DeepSeek. One thing I find striking is how little attention this has gotten in mainstream financial media. The stock market reaction has been muted. But if this cost curve is real, it should affect the valuations of companies that are betting big on proprietary AI. Luna: Which companies are most exposed? Obviously NVIDIA, because lower compute demand could mean fewer chip sales. But also the cloud giants we mentioned. Lucas: And also some of the AI startups that have raised huge rounds on the premise that proprietary models are a durable advantage. If open-weight models are nearly as good and cost a tenth to run, that advantage erodes. Luna: But isn't there a counterargument that cheaper AI expands the total market? More businesses can afford to use it, so total chip demand might still grow. Lucas: That's the bull case, and it has merit. Jevons paradox — if something becomes more efficient, usage increases. But the margins on that usage are lower. So the revenue per chip sold may drop. Luna: And for Chinese companies, lower margins might be acceptable because they're playing a volume game, as you said earlier. Lucas: Exactly. That's the pattern we've seen in steel, solar, and now EVs. China exports cheap, others struggle to compete. The question is whether AI services become the next battleground. Luna: I can see a scenario where US companies respond by doubling down on the highest-end models, aiming for capabilities that DeepSeek can't match. Frontier models that are ten times smarter, not just ten percent cheaper. Lucas: That's one plausible path. But it requires continued breakthroughs in model architecture, which aren't guaranteed. And it also requires massive capital expenditure, which is hard to justify if the market is simultaneously being commoditized from below. Luna: So the strategic landscape is shifting. And it's happening quietly, without a big trade war headline. Lucas: Right. The headline this week is about Trump and Canada and wildfire smoke. But the structural story that will matter in five years might be this little-known lab in Hangzhou showing that you can train a world-class AI model for five point six million dollars. Luna: That's the kind of detail that makes you rethink assumptions. And it's exactly why we dig into one specific number per episode. Lucas: Next time, I want to look at what DeepSeek's approach means for the carbon footprint of AI. If training costs drop, does energy consumption drop too, or does it rebound? That's a fascinating follow-up. Luna: I'll be reading the paper in the meantime. For now, thanks for listening.