（预告）Xi Jinping 与中国科技公司、芯片禁令的长期影响，以及对台湾的悲观展望

Speaking of DeepSeek, one of my thoughts—and I mentioned it in passing when writing about it—is that there’s going to be some impact that’s hard to know now, but probably significant for China, almost more from a psyche-and-belief perspective. You go back to the idea that these are very hard problems that need to be solved. Money isn’t enough. Incentive is also not enough. You sort of need the belief that we can do this.

Is that a good read? Is there a bit where DeepSeek—which is a very good model; it’s not the leading model, but it’s in the class of the leading models, both V3 and R1, and beyond—has had, or will have, this positive impact? Do you anticipate this sense that, look, even the stuff the West is supposed to be best at, we’re just as good?

1. DeepSeek Builds Chinese Confidence

DeepSeek really created that view. There are other models: Alibaba has a model, Apple apparently is going to use it for Apple Intelligence in China, and Baidu has a model. But DeepSeek kind of came out of nowhere, and they open-sourced it.

Of course, after the DeepSeek story percolated for a few days, it crashed the US stock markets, crashed NVIDIA, and caused a real melt-up in some of the AI- and tech-related stocks that trade in Hong Kong. It was very much a turning point from a psychological perspective, in the sense that, yeah, we can do this. Even though we’re struggling under this chip blockade, DeepSeek showed that they could find very creative ways to maximize the hardware they had and build an internationally competitive model. Then they made it open source, so now everyone is using it. Baidu has integrated it, and I think Tencent has integrated it.

Are you running it yet on your local machine?

Oh, yeah. I downloaded it, or a smaller version. I don’t have beefy enough hardware to run the full model.

Right.

But, yeah, why is it open source?

It's a bait.

People overestimate the extent to which the central government knows or cares about this. I think DeepSeek may have made their own decisions all along. Is there a sense that, oh, this is actually really valuable—should we be open-sourcing it?

To what you said, I think they made their own decisions. They were originally a hedge fund. They actually got in a little bit of trouble around a crackdown on quant trading. They were a quant fund, and they had bought all this hardware—all these NVIDIA chips.

And Xi Jinping is now saying, “This is my quant.”

Right. It’s amazing how quickly he’s risen. Li Wenfeng, the CEO, was at this meeting with Xi on Monday. He met with the premier a week, 2 weeks, or 3 weeks ago.

But no, I think they just did it. They open-sourced it. Now, though, I think there’s a realization that this is an incredibly powerful thing for China because it’s a very good model and it’s open source. Anyone, any country, anywhere around the world can download it and have this Chinese model running instead of having to pay for Anthropic, for Claude, or for OpenAI. It’s a really fascinating way for a Chinese AI model—at least one Chinese AI model—to go global very quickly.

2. OpenAI Faces a Brand Threat

The reaction to it has been really interesting, because most people’s encounter with it is not downloading it to their local machine and running it. It’s using the DeepSeek app. But it speaks, just from a business perspective: I think that OpenAI—number 1, I said from the very beginning that ChatGPT was just an accident in many respects—had achieved the most valuable and difficult thing in tech, which is a consumer brand with meaningful market share.

Part of that is that your inevitable end state is advertising, and they need to get there fast so that they can give free users the best possible models. People got DeepSeek and thought, “Wow, this is amazing. It’s so much better.” Well, yeah, because they weren’t paying for the better OpenAI models. It wasn’t the best, but to a lot of people it felt like it was.

I don’t know: Was the propaganda effect of DeepSeek greater in China, or on people in the US and the West?

That’s a great question. I think what’s interesting in China is how quickly it changed the market, because now all these other companies that were trying to charge for their models have to go free, too. It’s not at all clear what the business model is around these models in China now.

That’s a question in the US, too. Don’t worry.

At least in the US, OpenAI has revenue. Anthropic has revenue from subscriptions—not enough to pay for it, not enough to cover costs, but still.

3. Why DeepSeek Went Viral

I’m a fairly skeptical person. I’m curious about the sudden surge of DeepSeek on social media, like on X and in the App Store. I do wonder how much of that was totally authentic and how much of it was inorganic.

Yeah, I know you mentioned that. I feel like it was pretty authentic. I think the meta bit is that V3 came out over Christmas. They had documented a lot, and they’ve been publishing papers and models for several years, so this wasn’t out of nowhere by any means.

I think V3 had some of those cost estimates, which were totally twisted and warped by everyone driving their own agenda. They were very clear in the paper that the cost they published was for the specific training run. It wasn’t for all the experimentation, all the R&D, and all those sorts of things. They never said otherwise. People are trying to paint it as if they were trying to trick people, but the paper is very clear. It lists all the things that the cost did not include.

So V3 comes out. It was the one that actually had that dollar figure, $6 million or whatever it was. It was a very, very good model that was very, very cheap. Then R1 comes out, and I think it was a combination of the fact that people hadn’t used reasoning models yet because they were paywalled.

Number 1, it was people’s first access to a reasoning model. Number 2, the UI—or the UX, I should say—for DeepSeek was better because it actually laid out its thinking. If that was the first time you used a reasoning model and you saw the model talking to itself and trying to figure out the answer, it was kind of charming. It was like, “Oh, look at my little AI friend trying to help me out and figure this out.”

OpenAI did not expose that for competitive reasons. They were saying, “We’re not going to list what we’re doing.” So you had a double whammy: it was behind a paywall, and it was behind a competition wall, or whatever you might want to call it.

Then you layer on the general angst about China—the idea that at least we have AI, and that’s our great hope. And then there’s the bit where we’re spending billions and billions of dollars. The stock market is resting on these investments of billions and billions of dollars. Is this all kaput?

I think all of those things created a perfect storm. It just became a current thing for a weekend. We’ve seen that happen before. There were so many factors that made sense for this to explode that I’m inclined to give the benefit of the doubt to it being organic, as opposed to inorganic.

Okay. I think it was some mix, but I will defer to you on that. You made a pretty compelling case.

4. DeepSeek Disrupts the AI Market

I will say what’s interesting is that DeepSeek disrupted, obviously, stock prices here and some videos.

And to be clear, everything is almost back up to where it was. It was very much a current thing.

Right. Yeah.

But the disruption was that they also disrupted the Chinese AI market, which is really interesting.

Right.

Right?

Yep.

This is where they went: They disrupted globally, and frankly, I think, good for them.

I think the U.S. AI companies needed to be disrupted. They were really fat and happy.

Oh, yeah. No, I mean, people were comparing OpenAI pricing. That's because their margins were super large. The pricing's already come down. They've already gotten more aggressive, I think, in releasing things. The GPT-4o update over the weekend appears to be significantly reduced in terms of the HR voice, like scolding you for things; it's more open.

And I think we're actually seeing a pretty compelling competitive response. By the way, Google has models out there that are even cheaper and arguably just as good or better. Again, it was just this perfect: everyone's perception got—it was a bubble—

Reordered.

—that got pricked, but if you were paying attention, it wasn't totally shocking.

Now, I hesitate—I almost feel bad saying that—because DeepSeek deserves so much credit, and the engineering they did was amazing. All their work, if you go back 2 years and read their papers—and I haven't read all of them, but I've read 3 or 4 of them—is really good stuff, with some genuine breakthroughs that are going to be, or have been, adopted globally.

But that almost makes the point. The myth of AI has always been a bit different from the reality, but the reality is also fairly spectacular and not fully appreciated either. So there's just this crazy mishmash.

No, it's interesting. Again, I think the U.S. Silicon Valley firms should thank DeepSeek for a lot of what they did, right? Because ultimately, even though OpenAI, Anthropic, and xAI could buy as many NVIDIA chips as NVIDIA can make, won't they be able to make their models run much more efficiently and much better if they learn from DeepSeek?

Well, this is the interesting thing. With Grok, Grok 3 just came out this week. It appears to be the state-of-the-art model. At least, o3 may be better, but o3 is this very distinct sort of thinking model that I don't think anyone is ever going to release directly. It is in Deep Research, which is incredible but has very clear flaws, to be clear.

But this is, at least for someone like me, a very visceral feeling: Yeah, there's a lot of jobs that are really screwed looking forward. I know the people who program have felt this way for a while because AI has made such a difference there. And so it's a state-of-the-art or state-of-the-art-adjacent model. What's incredible about xAI is that it was founded 19 months ago.

Right.

And now they have a state-of-the-art model, and it's almost the inverse. It's the flip side of the DeepSeek story, which is: it's incredible, these optimizations DeepSeek did. They completely rethought how you do a mixture-of-experts architecture, which is definitely better for inference, but it had all this training overhead. They changed how you do the training to be able to scale that much more gracefully because of their bandwidth limitations; they couldn't handle too much overhead.

And I also believe, by the way, they were using H800s. They weren't using H100s because—

Right.

—they did so many things with how they designed the model. That speaks to a company struggling with bandwidth limitations, which were exactly the sort of limitations they shared.

And they said that. I mean, they've said—the CEO has said, other employees have said—their biggest constraint is chips.

Right, which I think totally lines up with the way the model's designed.

Access to chips.

So I actually think DeepSeek—again, with China, with everyone in general, you just should be skeptical—but this is another case where I believe them. Everything around this story lines up with that.

Yeah.

5. XAI Takes the Capital Intensive Route

But xAI comes in, and they deliver this state-of-the-art model in 19 months. A big part of that is they've raised $16 billion or $12 billion, and they bought a whole bunch of NVIDIA chips and wired them all together. And how could they do that? Because they had access to the chips. Also, NVIDIA—one of their big differentiators is all the networking stuff they do, where they make it easy and possible to tie a ton of chips together to get this sort of performance.

And so you can look at American AI companies and say, “Wow, why didn't you do this optimization?” On the other hand, if you look at it from a comparative-advantage perspective, it's like—I always mock big companies trying to copy a startup. A startup invents something, and they're like, “Oh, we can do that, too.” Then you get Facebook releasing the Poke application. It's like, why are you trying to do that? Inventing something's really hard. You're almost capturing lightning in a bottle. When you're small and a startup, you do it because that's the only way to do it. And, by the way, most startups fail.

Right.

If you're a big company, you have large amounts of cash. You can de-risk by just going and buying the startup. Go and buy the people inventing it, bring them in-house. Or, in the case of Facebook, Poke was a response to Snapchat. What they actually did is just, “Okay, we'll just rip off Stories and put it in Instagram,” and basically stopped Snap in its tracks.

And it's not very glamorous, but it's actually recognizing your advantage. I think that's what we saw with xAI. Did they do the grunt work of DeepSeek to heavily optimize around a limited number of chips with low bandwidth? No, they just bought a bunch of chips because they had a bunch of money, but it also got them where they wanted to go.

Right.

Right? And so xAI and DeepSeek have totally different approaches, but both of those approaches are rational given their circumstances, and that in and of itself, I think, is an interesting takeaway.

6. Chip Access Shapes the Race

And then one of the questions, right, is, going forward, you push out 1 year or 2 years. If DeepSeek continues to not have access to the best NVIDIA chips and effectively can only buy Huawei's Ascend chips, whereas xAI or OpenAI can keep buying the better NVIDIA chips, do you start seeing a real separation?

I mean, that is the big question. There are a couple of concerns that I have about this, and I think we've talked a bit about this offline.

Yes.

So let's buckle up and get into it.