What China Does Not Want the U.S. to Know

In a leaked call transcript, we learn what DeepSeek's CEO thinks are China's weaknesses and strengths in the AI race with the United States | Edition #309

Guan Yu, circa 1700 (painting, modified)

Today, a call between DeepSeek's founder and CEO, Liang Wenfeng, and a group of investors was leaked.

A few hours after the leak, the WeChat links referring to this call had been removed from the internet.

It is still unclear whether the Chinese government ordered it due to political, economic, or regulatory risk, or whether it was a private removal request.

My guess is that it was a private request, but the Chinese government is glad it was removed, as it contained a Chinese AI CEO's thoughts on China's strategic positioning in AI.

What is DeepSeek and why do its CEO’s opinions matter?

DeepSeek is a leading Chinese AI company that caused a massive shock to the U.S. stock market last year when its R1 model closely matched the capabilities of OpenAI’s latest o1 model at the time.

Last year’s event led to the public acknowledgment, for the first time, that contrary to what many thought, Chinese capabilities in AI are approaching those of the U.S.

This has been further noted in global AI reports, including the 2026 Stanford HAI Index Report.

Why does this call matter to AI policy professionals?

The call leaked today took place on May 20 and lasted 3 hours and 44 minutes.

The information shared in this call reflects, among other topics, the views of the CEO of a leading Chinese AI company on China's main strengths and weaknesses in the AI race with the U.S., as well as its global strategy.

These assessments are strategic for shaping American (and global) AI policy efforts, such as export controls and bans on Chinese AI systems.

Remarkably, what DeepSeek’s CEO says is similar to what Anthropic’s Dario Amodei said in 2025 and reiterated this year: compute is the most important bottleneck in the AI race between China and the U.S.

(According to this view, export controls are probably an effective way to curb Chinese competition in AI).

The most interesting excerpts:

Below, I selected 11 excerpts that reflect DeepSeek’s CEO's thoughts on China's weaknesses, strengths, and its competition with the U.S.

If you work in AI policy, or are interested in the political, economic, and technological AI race between the U.S. and China, these excerpts will help you understand the Chinese point of view.

You can find the full transcript here.

1. DeepSeek lacks resources

“What I just said is that our company’s most important issue is personnel stability. From another dimension, what are we lacking? What’s the gap with the U.S.? Only one thing: resources.”

-

2. The talent gap is due to the compute gap

“Our gap with the U.S. is mainly in resources; in people, the gap isn’t large—almost none, because it’s the same pool of Chinese talent. Some stay in China, some go abroad; it’s not that the smarter ones leave.

Talent isn’t the bottleneck—resources are. Resources affect talent development because less compute means fewer experiments, so our talent lags behind the U.S. The talent gap is essentially due to the compute gap.”

-

3. “We’re still at tens of billions—an order of magnitude difference”

“For the largest models now, we can’t afford to train them. Even spending all 50 billion, we can’t train them; even if we could stack, we can’t afford to use them. The largest model now has about 800B activated parameters; domestically, we’re still at tens of billions—an order of magnitude difference.

To train a model as large as the top U.S. ones, we’d need 50,000 GB300 or 200,000 Huawei 950 cards—just for training, excluding research. So our biggest gap with the U.S. is resources.

Our current resources—within this year and the coming months—only allow experiments in the tens-of-billions-activated scale. We still have many experiments to do and many things to figure out at that scale. We’re far from being able to train an 800B model—there’s plenty of time, and not enough cards.

So I think our difference with the U.S. is purely resources. All observed differences—talent, model capability, applications—can be attributed to compute resource differences.”

-

4. “Huawei’s production is limited”

“Compute resource issues are twofold: domestic cards are hard to buy, and our capital investment is lower than the U.S. We invest much less; talent salaries are a small fraction—even if someone is paid $100 million, it’s still a tiny share compared to compute.

This issue is largely unsolvable right now because Huawei’s production is limited. To train an 800B model, we’d need 200,000 of Huawei’s latest cards—just for training, not research.

So we won’t even consider competing with the U.S. at that scale now. We focus on the scale we can afford—tens of billions activated—and do it well. With more resources later, we’ll move to 150B, 156B, or 250B activated.

So we have a gap with the U.S. that currently seems hard to bridge. You could force-train a large model, but you can’t do sufficient research before training.”

-

5. “About half of our people think OpenAI is better”

“Is Anthropic’s current lead over OpenAI long-term? I don’t think so—it’s definitely episodic. OpenAI and Google will likely trade leads in the future. Actually, Anthropic’s advantage in Code Agent isn’t that big—it’s not crushing OpenAI.

About half of our people think OpenAI is better. Anthropic has first-mover advantage, but that should fade quickly—not a durable edge. These three are all strong; among them, ours is the most efficient, with the lowest costs and burn rate.

In the global AI division of labor, Chinese companies are likely to play the role of largest producer. Typically, our production capacity is largest—including chips—and we have the most electricity. So China’s AI is likely to be one of the three major players.

Chinese companies will make products cheapest, and in terms of effectiveness, many Chinese goods are not much different from U.S. ones. Future AI may be similar, but Chinese AI will likely be cheaper—systemically lower, like other Chinese services.

When I decide what to do, I habitually think: what gives the highest return now? If product development has the highest return, I’ll do that; if achieving AGI first has the highest return, I’ll do AGI. Clearly, I think product isn’t highest-return now.”

-

6. “NVIDIA is digging its own grave”

“By redoing this process on Huawei cards, we complete the transition. This could be a historic mission—completely overturning the perception of poor domestic chip ecosystems. Conditions are mostly right; it just takes time. Within a year, many will see this issue resolved—leaving only production capacity.

I’m optimistic about domestic compute. On this point, NVIDIA is digging its own grave. Huawei’s 950 supernode can fully replace NVIDIA’s GB200/GB300 in performance and price. Price will be higher but limited—50%, 100% even 200% doesn’t matter.

If 100% higher, it’s still a price replacement. Task-wise, it’s also a replacement—everything GB300 can do, Huawei’s supernode can do, with similar latency. The only cost: four Huawei cards equal one NVIDIA card, with a two-year lag.

Four-to-one is understandable. Two years behind means four Huawei 950s equal one GB300. Huawei 950 supernode ships Q3/Q4 this year; NVIDIA GB200 shipped Q3 two years ago—two-year gap.

NVIDIA may have a new generation in Q3 this year. So our chip gap with the U.S.: I think the ecosystem gap will disappear, but chip-wise it’s four times plus two years.”

-

7. “We don’t maximize profit or price for maximum revenue”

“OpenAI initially thought it could monopolize, but faces many challengers—it’s not easy. The U.S. will face challenges, including from China, where people are willing to take less to provide services.

In China, some will take even less. But there’s a balance: too little and the business model fails. Take too little, you can’t survive; take too much, you’re beaten by those taking less.

So for us, we don’t maximize profit or price for maximum revenue—we only take a reasonable return. That’s an explanation.”

-

8. “Our gap with the U.S. might be (...) roughly two years, but using one-twentieth the compute”

“Our gap with the U.S. might be 12 to 18 months, or 6 to 12—roughly two years, but using one-twentieth the compute. The narrative: one to two years behind, one-twentieth the compute. Future: we’ll shorten the gap to six months or three months with fraction-of-the-compute—that’s a goal.

And we might surpass them in some areas, though overall, with an order-of-magnitude compute gap, comprehensive catch-up isn’t realistic. But in focused areas, we can.”

-

9. “I see no upper limit to language model scaling”

“I see no upper limit to language model scaling—neither at our level nor the U.S.’s.

Interesting: the U.S. can train an 800B model but can’t afford to use it widely—so it’s hard to serve.

Humans gained language ability only recently, yet training AI might reverse the order, but eventually physical models or embodied intelligence may be needed.

Yes, embodied intelligence is inevitable. For normal people, needs aren’t computers—they need embodied intelligence to solve physical labor needs. So embodied is unavoidable.”

-

10. “Our capital structure can’t support high-cost labeling like that of the U.S.”

“Another question: Hallucination affects user experience. It’s solvable through better post-training—improveable. But we see it as a product issue—not the top priority.

On data labeling: our capital structure can’t support high-cost labeling like the U.S. Chinese labeling has no cost advantage—especially high-end data—so we can’t invest as heavily.

So we use two approaches: label low-cost data first. About half our core researchers are labeling data—focused on that. Solving AI at this stage relies on labeling data—it’s all about data.”

-

11. “In China, the most reasonable approach seems to be to focus on general Agents, with Coding Agent as the top priority”

“First, you mentioned Chinese models are more efficient than U.S. ones. In what other aspects might we surpass the U.S.?

In many user experience aspects—our product experience is good—we may not be worse. Product capability probably not worse. Costs will be lower—China can be competitive.

Other structural advantages? Maybe none. But cost and product are structural advantages.

Cost advantage: U.S. companies don’t focus on cost—they don’t develop that capability. We prioritize it; they don’t. Product-wise, Chinese companies have strong product capability—so structural advantages exist.

Second, you mentioned post-training is costly—Anthropic and OpenAI spend huge sums. After funding, will we increase post-training investment?

The gap is in high-quality data labeling—mainly in AI research. We will increase investment, but labeling quality data is bottlenecked by time, not capital.

OpenAI, Anthropic started earlier, with more capital and cards. China only started seriously in the last six months—so time is needed. Capital isn’t the constraint—we already expand at max speed.

So within a year, high-quality data should be manageable—expected. It takes time, but not too cold.

Third, Anthropic uses its model to build vertical financial, legal…

…even medical products. Will we consider vertical applications at some stage?

I haven’t thought it through. China’s business models and path may differ from the U.S.—hard to judge now.

In China, the most reasonable approach seems to be focusing on general Agents, with Coding Agent as top priority—other vertical agents like finance or medical come later.”