Resource Is the Only Gap: Compute, Cost, and Endgame Competition

If I could spend all this money within half a year, that would be ideal. Because turning money into NVIDIA cards is definitely better than leaving it in the bank. In the bank, I'd get maybe two percent interest, that's it. But buying NVIDIA cards gives me a societal cost over ten months.

So you definitely buy as many as you can. Once I have the cards, I have a lot of room to generate cash flow by providing services or something. With cash flow, I can survive, and I don't need to keep a lot of money on the books.

So our only worry is not being able to buy enough cards. If we could turn all the money into cards, we would do it without hesitation, and we'd even be willing to pay a premium. It's just too cost-effective. Even after paying a premium, we might still not achieve that goal. Objectively, if I can spend 20 billion this year, that means our procurement department has done an awesome job.

The gap between us and the US is mainly in resources, not really in people. There's almost no gap in people because they're the same people, Chinese. Some Chinese go abroad, some stay domestic, some go abroad-it's not that the smart ones all go abroad, that's not the case.

It's actually pretty random. The smartest ones aren't necessarily more than half going abroad; maybe slightly less than half stay in China. We don't lack talent domestically, and we have a large base with so many new graduates each year.

Talent is not the bottleneck; resources are the biggest bottleneck. Resources first affect talent development because with less compute, we have fewer opportunities to run experiments, so our talent overall lags behind the US. The talent gap is essentially due to the compute gap.

We can't really train the largest models now. Even if we spent all 50 billion, we still couldn't train them. Even if we could stack up the compute, we couldn't afford to run them. The largest models today have around 800 billion activated parameters; domestically, we're still at the tens of billions scale-the largest domestic model might have tens of billions activated, an order of magnitude difference.

To train a model as big as AI's, I'd need about 50,000 GB300s or 200,000 Huawei 950s. That's just for training, not even research. So the biggest gap between us and the US is in resources.

Our current resources, including what we'll have in the next few months, even with the big influx soon, are only enough to run more experiments at the ten-billion activated parameter scale. We still have many experiments to do and things to figure out at that scale. We're still far from being able to train an 800-billion-parameter model; we need more time and cards.

So I think the difference between us and the US is purely a resource difference. We can argue that all the differences we see-in talent, model capability, applications-are essentially due to the compute resource gap.

On the compute side, part of the issue is that we simply can't buy enough cards domestically, and on the other hand, our capital investment is much lower than that of the US. In terms of capital investment, we're far behind. Talent salaries account for a very small share-look, they pay salaries of a hundred million dollars, but even then, talent costs are a fraction; the bulk is compute.

This problem is basically unsolvable for now because Huawei's production is limited. To train an 800-billion-parameter model, I'd need 200,000 of Huawei's latest cards, and that's just training, not research. So we don't even think about competing with the US at that scale. We should first do well at the scale we can afford-tens of billions of activated parameters. Then, when we have more resources, we can move to 150 billion, 156 billion, or 250 billion activated parameters.

So there's a gap between us and the US that currently seems hard to close. You could force-train a large model, but you wouldn't be able to do sufficient research beforehand.

Another question people care about is where the endgame differences in the large model competition will manifest. When it comes down to it, what will distinguish the winners? I think in the end, there may not be huge differences. The final gaps will be in three areas: cost, time, and user experience. Beyond that, probably nothing else.

Cost is straightforward: providing the same service with the same quality at what cost. Take BYD's batteries for example-can other companies offer the same technology at the same price? That's tough. It's not easy; there's certainly a moat. So cost is definitely a differentiator, and I think it's the primary one. Second is time: when you can achieve it. Being a few months earlier or later makes a difference.

Third is user experience-there are some differences in user experience. There will be some user stickiness and barriers, but that might not be fundamental. The fundamentals are cost first, then time. Whether you make it first or later, and how fast you deliver the same thing.

As the long-term commercialization path and product lines get richer, how should we price? We think the most worthwhile thing to focus on right now is still AGI. We need to push AGI forward, to raise the lower bound of intelligence. At this stage, that is more worthwhile than building more product lines or considering more commercialization paths-it's a higher-return path. I think this has been true in the past and will be true for the foreseeable future. That is, if we spend a lot of time thinking about how to enrich our products, what use is that? What was the commercialization plan we discussed six months ago? It must have been about running ads, then e-commerce, embedding e-commerce into products, and integrating deeply with local services. That's definitely useless because things change too fast. If you are ahead of the curve at this stage and spend a lot of time on commercialization considerations and paths, your product line will have a very short life cycle.