DeepSeek's Liang Wenfeng Tells Investors Restraint Is the Business Model

Liang Wenfeng has never seemed comfortable in the role of AI billionaire, and on a closed-door call with the investors now backing DeepSeek, he said as much. "We are a group of very ordinary people," the founder told them, according to a transcript of the meeting obtained by The Once Times. "If you like a narrative, it should be ordinary people doing something extraordinary, not geniuses doing something extraordinary."

The video call, which ran more than four hours, came as the Hangzhou-based lab was closing its first-ever round of outside funding. DeepSeek went on to raise $7.4 billion at a valuation above $50 billion, the largest private AI financing in Chinese history, with Tencent, battery maker CATL and a state-backed fund among the participants. The company is now plotting an initial public offering for 2027.

Yet Liang spent almost none of the call on the money. His subject was restraint, a word he returned to again and again.

Ten months, not ten times

DeepSeek prices its application programming interface on what Liang called a "reasonable profit": servers should earn back their cost in about ten months, which works out to roughly six times profit. That sits far below what the market would bear. Demand at current prices is inelastic, he said, so doubling the price would nearly double revenue. He refuses to do it.

He told a story to explain why. When one model launched, the team priced it high out of caution. Liang later cut the price to a quarter of the original level, and the company's internal chat group erupted in cheers. "For our competitors, a price cut is bad news. For us, people were celebrating," he said. "Whoever wants to take more will be beaten by whoever is willing to take less."

The same logic governs open source. DeepSeek will keep releasing its strongest models openly, Liang said, because at six times profit there is no conflict: third parties cannot match its deployment costs, so giving the weights away costs the company nothing. "I don't see what ByteDance gains from keeping its models closed," he said. The lab even helps rivals, including Alibaba, Zhipu AI and Moonshot AI, reproduce its results.

The road to AGI runs through continual learning

Liang laid out a stepped view of AI progress. Last year's step was chain-of-thought reasoning; this year's is agents. The next bottleneck, he said, is continual learning, the ability for a model to keep absorbing the world the way a new employee learns a job over two months. No lab anywhere has solved it.

After that comes a self-improvement loop, where a model can develop its own successor, though Liang stressed this "singularity" would be gradual rather than sudden. Embodied intelligence comes last. The next generation of DeepSeek models, he said, will only deserve the name if they can keep learning.

Until then, the plan is steady iteration. DeepSeek released its V4 family in April, a 1.6-trillion-parameter mixture-of-experts system and a 284-billion-parameter Flash variant, both open source and tuned to run on Huawei and Cambricon silicon as well as Nvidia's. Liang said a comfortable release rhythm is one version every two to three months, and a larger model with around 150 billion active parameters could begin training late this year or early next.

A hardware bet on Huawei

Liang was unusually specific about chips. DeepSeek now fields about 20,000 H-series-equivalent GPUs, most of them delivered in recent months, and more are on the way. "Within a reasonable price, we buy as many as we can," he said. If procurement manages to spend 20 billion yuan this year, he joked, that would count as a superb performance.

The gap with the United States, in his telling, is resources, not talent. Training a model with 800 billion active parameters, the class frontier US labs now work at, would take about 50,000 of Nvidia's GB300 chips or 200,000 Huawei Ascend 950s. Neither figure is within reach. So DeepSeek trains at tens of billions of active parameters and closes the gap with efficiency.

On Huawei, Liang disclosed that DeepSeek has been allocated about 16,000 Ascend 950 cards, an order of magnitude less than what internet giants receive, and equivalent to roughly 4,000 of Nvidia's B-series. Four Huawei cards match one Nvidia card, he said, and the silicon trails by about two years. But he argued Nvidia's real moat, the CUDA software ecosystem, is eroding fast, helped by TileLang, an open-source high-level language DeepSeek uses to rewrite GPU kernels. V3 was trained on Nvidia hardware without Nvidia's software stack, he said, and porting the same approach to Huawei chips is under way. "On this point, Nvidia is digging its own grave," Liang said.

The one thing he will not compromise

Asked what DeepSeek's core interest is, Liang gave a single answer: team stability. Everything else, money, users, market share, is negotiable. The funding round, he acknowledged, relieved the biggest risk because employees now hold meaningful options. Turnover at the lab has long run below industry peers.

The commercial fallback sounds almost casual. Business-to-business revenue could reach several hundred million dollars this year if demand holds, Liang said, and $1 billion in annual AI revenue would make the company cash-flow positive. "In the worst case, selling APIs could support a listed company," he said.

On the global race, he was sanguine. Anthropic's current edge in coding agents is a first-mover advantage that will fade, he said, and he expects OpenAI and Google to trade the lead back and forth. Roughly half of DeepSeek's own staff, he noted, still rate OpenAI's models as the best.

What he would not promise is a timeline. Continual learning is unsolved everywhere, and DeepSeek's approach is to keep many researchers thinking about it rather than throw compute at it. "It's like a lottery ticket," Liang said. "Anyone can buy one. Whether you win, I don't know if that's talent or luck."