Who is the good guy now in AI?

China Translated - Briefing #72

[This article is adapted by my own AI agents from what I personally shared in my recent Baiguan Radio episode with Olivia Plotnick, with my final proofreading and edit.

I enjoy writing, not least because writing is a form of thinking exercise for me, and I try to write everything myself. However, writing is only my part-time job, alongside my full-time CEO job. To strike the right balance between writing on my own as much as I can and publishing in time anything at all, some articles in this newsletter are written by AI, but always with 1) my own core thesis and skeleton, 2) internalization of my style, and 3) my final proofreading of each word before hitting Publish]

---

DeepSeek’s Liang Wenfeng is probably the least publicly visible founder of any major AI company today. Sam Altman and Dario Amodei appear regularly on podcasts, at conferences, and before policymakers. We know their arguments, their mannerisms, and sometimes far more about their internal conflicts than we would like.

In contrast, Liang has given only a handful of interviews, mostly in writing. We don’t even know what Liang’s voice sounds like. Yet, DeepSeek has somehow become one of the most consequential AI companies in the world while its founder remains almost entirely silent.

That is why an alleged transcript of a nearly four-hour meeting between Liang and DeepSeek’s investors caused such a sensation when it circulated online late last month. The transcript contains answers on everything from artificial general intelligence to Huawei, Nvidia, open-weight models, pricing, corporate governance, and the future structure of the AI industry.

DeepSeek has not confirmed the transcript, so every quotation attributed to Liang should still be read with that qualification. I cannot prove that it is real either.

My instinct, however, as soon as I got my hands on it, is that it probably is.

The simplest reason is that it sounds remarkably consistent with everything DeepSeek has done, or with whatever little clues Liang has previously let out into the public domain. In a rare earlier interview, Liang described open source as a cultural behavior rather than a commercial calculation. The company has repeatedly released model weights, papers, and production-tested code. Its official model repositories allow commercial use of many releases. It has kept prices extraordinarily low even when demand would probably have supported much higher ones. Its own website still describes the company as being “dedicated to exploring the essence of AGI”.

Then, the plot thickens. A few days after the transcript appeared, Bloomberg reported that DeepSeek had paused its second fundraising round, driven in part by Liang’s frustration over the leak. A new alleged story piles on another.

Who is the “good guy” in the AI race?

The American debate about China and AI usually begins with a familiar premise: the United States must win because it cannot allow the Chinese government, or people working under the Chinese system, to control such a powerful technology.

Yet, if we compare Liang with some of the leading American AI founders purely in terms of disposition, restraint, and the moral values we would want behind the wheel of AGI, it becomes difficult to say who the “good guy” is.

Again, we know very little about Liang. Secrecy naturally creates suspicion, especially in the current geopolitical environment. Yet the limited evidence we do have presents a founder who already had enough money, preferred difficult scientific problems to a celebrity life, and built a company around a narrow mission. In the leaked transcript, he repeatedly says DeepSeek will avoid attractive businesses that do not advance the path toward AGI. The company does not use conventional KPIs for researchers. It wants employees to retain time for unscheduled exploration. Liang also famously describes restraint itself as a strategy.

This kind of person is a minority in China. But it’s a growing minority.

Most Chinese entrepreneurs, like most entrepreneurs everywhere, care a great deal about profits. But I increasingly see a newer generation of founders, who were born into relative comfort and for whom money becomes the means rather than the end. (POP Mart’s Wang Ning is another case in point.)

Liang’s self-made wealth as a quant boss gave him the freedom to start DeepSeek without asking investors for permission. More importantly, it gave him the freedom to define success in his own way.

That freedom shaped the company before outside capital arrived.

“Open weights” as the objective

Many technology companies use open weights models to recruit developers, win attention, or pull users toward a paid product. For them, openness is one stage in a commercial funnel.

DeepSeek, however, seems to judge success partly by how widely its models can be used, including by people who will never become DeepSeek customers.

The alleged transcript makes that philosophy unusually explicit. Liang says DeepSeek intends to release even its most advanced models, wants other companies to deploy them successfully, and does not keep a superior private version for itself.

What’s remarkable in the transcript is that he laid out a clear maximum profit goal (10-month payback and 6-fold return on investment), so the open-weights stance does not seem like an empty promise but an intentional, calculated bet. In his own telling, the more restrained company has a better chance of reaching AGI because it avoids wasting energy fighting for every adjacent source of revenue.

In other words, open weights are part of the mission itself.

This is close to what the name OpenAI originally meant. In its 2015 founding announcement, OpenAI described itself as a nonprofit intended to build value for everyone rather than shareholders. Its researchers would publish papers, code, and patents, and collaborate freely with other institutions. Its charter still says its mission is to ensure that AGI benefits all of humanity.

OpenAI has faced pressures and responsibilities that DeepSeek has not yet encountered, and DeepSeek’s own ideals may eventually collide with commercial reality. Still, judging from their present behavior, the Chinese company looks surprisingly close to the original OpenAI proposition.

The Chinese government has also begun to recognize the geopolitical value of this approach. At the 2026 World AI Conference in Shanghai, Xi Jinping emphasized openness, inclusive development, and support for AI capacity in the Global South. China’s Foreign Ministry subsequently described the country’s open-source AI ecosystem and AI-related international public goods as tools for narrowing the global AI divide.

That language is self-interested, of course. All soft power is. But cheap, downloadable Chinese models offer something tangible. A developer in Indonesia, Brazil, or Nigeria can run the weights locally, modify them, and avoid permanent dependence on a handful of American platforms. This is a more useful form of soft power than another nice speech about “win-win cooperation.”

Is “intelligence” a commercial product, or a public good?

It’s tempting to compare China’s AI strategy with its manufacturing playbook: enter at a low price, improve quality quickly, and eventually offer a product good enough that the price gap becomes decisive.

That’s the more friendly way to put it. The unfriendly way will call this strategy “dumping”.

The analogy is useful, but AI may go further. A vacuum cleaner or an electric vehicle is just a consumer product. But “intelligence” could become a basic input used by every person and every business, closer to what the internet, electricity, or water are today. If that happens, the social value of making intelligence extremely cheap may outweigh the value captured by any one model company.

On this topic, Liang sounded almost Hegelian. In the leaked transcript, he said:

This differs from open-sourcing a piece of software in the past, because that software’s market wasn’t so large, while AI is simply too big. If we tried to monopolize the benefit, we would certainly be discarded by history. Most fundamentally, I think this is an objective law — a view of history.

这跟之前开源一个软件可能是不一样的，因为那个软件的市场没那么大，而 AI 这件事实在太大了。如果我们想独占这个利益，那一定是要被历史抛弃的。我觉得最主要的是，这是一个客观规律，这是一种历史观。

DeepSeek’s pricing suggests that it is already operating with this possibility in mind. According to the leaked transcript, its API prices are designed to recover equipment investment in roughly ten months and earn what Liang considers a “reasonable return.” He says the company could charge substantially more without losing much usage, but chooses not to, and when DeepSeek once cut a model’s price to one quarter of its original level, the team “celebrated.”

The Nvidia triangle

The leaked transcript becomes more politically awkward when Liang turns to semiconductors. He reportedly says Nvidia is “digging its own grave” by allowing export controls to accelerate Chinese substitution. He also claims that DeepSeek used Nvidia hardware to train V3 while reducing its dependence on Nvidia’s software stack, and that the company is working closely with Huawei’s emerging ecosystem.

Some of the specific technical claims in the transcript, especially its comparisons between Huawei and Nvidia hardware, remain unverified. Yet, the direction is much easier to see. DeepSeek is helping Chinese models run efficiently on a technology stack that depends less on CUDA and advanced American chips.

I recently described this as the “galapagosization” of Chinese technology. Export controls create short-term pain, then give Chinese companies a powerful reason to develop alternative hardware, software, interconnects, and standards. DeepSeek sits on the software side of that emerging ecosystem.

The interests of the three main actors now point in different directions. China wants distance from Nvidia. The US government also wants Nvidia to keep its distance from China, but for a different reason. Nvidia itself wants to remain at the center of both ecosystems.

This is why Jensen Huang sounds so frustrated when American policymakers tell him the China market will be lost anyway. Jensen understands that technology competition is perpetual. If Chinese developers keep building on Nvidia, the American ecosystem retains scale, influence, and visibility. Forcing them away does not stop development. It helps create a second ecosystem that the United States can neither control nor easily observe.

Liang’s reported language is more combative than Jensen would like, but the two men appear to agree on the mechanism.

Capital comes after the mission

DeepSeek now faces a contradiction of its own. I previously argued that its success was possible precisely because the Chinese government did not create it and venture capital did not control it. I went further in Baiguan: had DeepSeek taken conventional VC funding at the beginning, investors might have pushed it toward user growth, monetization, and a faster exit before its research culture had time to mature.

DeepSeek has now taken outside money. According to The Information, as relayed by Reuters, its first round raised more than RMB 50 billion, or about $7.4 billion, at a valuation above $50 billion.

I do not think this invalidates my earlier argument. Sequence matters. DeepSeek accepted investors after its mission, culture, and technical credibility were established. The reported deal placed most investors into a limited partnership controlled by Liang, imposed a five-year lock-up, and gave them no voting rights. (China’s national AI investment fund was the notable exception.) These are unusual terms for an unusual company.

The money also serves a clear purpose. DeepSeek needs compute, and it needs to keep a small group of researchers who could command much higher compensation elsewhere. In an earlier account of the fundraising meeting, Liang’s most important demand to investors and large technology companies was simple: do not poach his people. The leaked transcript says team stability is the company’s main non-negotiable interest, and that financing has reduced the risk by giving employees meaningful equity.

Capital has become a means of protecting the mission. Whether it stays that way is now the central test.

Bloomberg reported that the paused second round could resume later and that DeepSeek had been seeking a valuation around RMB 500 billion. Reuters has also reported early deliberations about a possible listing on Shanghai’s STAR Market. Neither step is a confirmed company plan. Yet a domestic listing would make sense if DeepSeek needs liquid equity to retain talent and finance compute without surrendering strategic control.

The company does not need to remain profitless to remain mission-driven. Liang’s stated goal is a reasonable profit and a durable commercial foundation. The real question is what the company chooses to optimize when revenue, market share, research openness, and its AGI mission eventually pull in different directions.

That test has broken many idealistic organizations before.

When OpenAI was founded in 2015, it said a leading AI institution should prioritize a good outcome for everyone over its own self-interest. Nearly eleven years later, one of the companies making the most serious attempt to operationalize that idea is a secretive Chinese laboratory created by a quant billionaire and now financed partly by Chinese state capital.

DeepSeek may still disappoint me. Its culture could change as it grows. Political pressure could narrow its openness. The alleged transcript could contain errors or inventions. But the vision inside it is coherent with the company’s behavior: make the models open enough to spread, cheap enough to become infrastructure, and profitable enough to keep the research going.

I hope DeepSeek ends up becoming the real OpenAI. That is what I am rooting for.