Liang Wenfeng (DeepSeek) Meets Investors: Restraint as a Strategy

The transcript is based on an audio recording of about 3 hours 44 minutes that was transcribed automatically and edited with AI, so some names and figures may have been misrecognized. Below is the position of Liang Wenfeng, founder and CEO of DeepSeek, following his own line of reasoning.

Vision Instead of KPIs

According to Liang, nobody thought about making money, going public or an IPO when the company was founded: the team took on the work “with enormous goodwill toward the world.” Following Jack Welch, he believes the most important thing for a company is its vision, not a slogan on the wall but how the company actually operates. DeepSeek, he says, has no formal organization and no KPI system, and the vision is not even written down: it lives in the way people work.

Liang calls himself and his colleagues “a group of ordinary people who did something extraordinary.” Two years ago they had no big money, no GPUs and no reputation, and he describes himself as a graduate of a university that was “not even a top one.”

Restraint as a Strategy

The key word of the meeting is restraint. Liang estimates that AI may eventually account for around 10% of global GDP, and such a market cannot be monopolized. Whoever wants to take 5% will lose to someone content with 1%, who in turn will lose to someone willing to take 0.1%. So the company aims for a “reasonable profit”:

• API prices are set so that purchased hardware pays for itself in about ten months, which corresponds to roughly a sixfold profit; for accounting purposes servers are depreciated over three to five years;
• according to Liang, demand in this price range is inelastic: doubling the price would nearly double revenue, while further cuts would not bring more users;
• once the company cut the price of one of its models to a quarter, and employees cheered in the company-wide chat;
• he believes DeepSeek’s optimization is what makes this cost level possible; by his estimate, large competitors’ costs are several times higher.

The company took the same approach to its sudden popularity during last year’s Spring Festival: it served the users but did not monetize them or fight ByteDance and Tencent for the market. Liang cites the Chinese saying about the sesame seed and the watermelon: today’s market is sesame seeds, and the watermelon of AGI lies ahead.

Open Models and the Business

Liang describes open source as part of the vision and a sound business decision at the same time. He says the company’s strongest model will probably be open too, and the open version is identical to the one DeepSeek deploys itself. At a sixfold profit, openness does not hurt the business: it makes no sense for a third party to deploy the model on its own, because matching DeepSeek’s cost is very hard. Openness would only get in the way of someone who wants a hundredfold profit.

According to Liang, DeepSeek helps everyone deploy its models, including competitors Alibaba, Zhipu and Moonshot AI, and only worries that their quality or cost might suffer. He calls the consumer and enterprise segments by-products on the road to AGI: a few people are enough to support the API, and there is no customer service or sales team. Still, he estimates API revenue this year could reach several hundred million dollars, and at a billion dollars the company’s cash flow would turn positive.

The Roadmap to AGI

Liang describes AI progress as a staircase: the language model, then chain of thought (CoT), and this year agents. Each step builds on the previous one. Today’s AI already outperforms humans when given full context and instructions, but it cannot replace an employee who needs two months to settle into a company. One thing is missing: continual learning. From there, he hypothesizes, the order is:

• continual learning;
• a “singularity,” meaning a model that can develop its own next versions, and a gradual process rather than a single point;
• embodied intelligence that enters the real world: housework and elder care.

The company does not consider video generation or world models part of the main line of intelligence, although it acknowledges they are good business. Multimodality is a component, like search: V4 and later versions will support it natively. Liang does not yet see a ceiling for scaling language models. The first goal of every new model is to help DeepSeek’s own development, because that is the fastest path to AGI.

Chips, Huawei and the Gap with the US

Liang names one key interest for the company, essentially the only one: team stability. Money and resources, he says, are not a problem, and DeepSeek’s staff turnover is lower than its peers’. He sees the gap with the US primarily in compute:

• DeepSeek has about 20,000 cards in H-equivalent terms, most of them received in the last month or two; the company buys as many cards at a reasonable price as it can, since NVIDIA cards pay for themselves in ten months while money in the bank earns a couple of percent a year;
• the largest US models have about 800B active parameters, Chinese ones a few dozen B; training such a model would take 50,000 GB300s or 200,000 Huawei 950 cards;
• Huawei allocates the company about 16,000 of its 950 cards, equivalent to roughly 4,000 B-series cards; Liang believes the Huawei 950 supernode can replace the GB200 and GB300 at the cost of “four cards for one plus a two-year lag”;
• the in-house high-level language TileLang removes dependence on the CUDA ecosystem, and Liang considers a 1-2% efficiency loss acceptable.

In Liang’s view, the CUDA “moat” is eroding fast: AI can write code and helps recreate the ecosystem, languages such as TileLang have appeared, and specialized chips are no longer tied to the legacy of gaming GPUs. He expects practice to show within a year that the Chinese chip ecosystem works fine, leaving only a shortage of manufacturing capacity.

He estimates that Chinese models lag the US by 6-18 months, or “a year or two with one-twentieth of US compute,” and aims to narrow the gap to 3-6 months. In the optimistic case, training of the next model with 150B active parameters could start at the end of the year.

The Market and How the Company Works

Liang predicts that competition among large models will ultimately come down to cost, time to market and user experience. Too many companies in China are building foundation models, and the market will consolidate to three or four players. In his view, China will become the largest producer of AI by volume and make it cheaper, and Anthropic’s current lead over OpenAI is temporary.

Among practical priorities, Liang singles out the coding agent: in China, he believes, it makes the most sense to focus on a general-purpose agent, while financial and medical agents can wait. Half of the company’s key researchers are now busy with data labeling, because solving AI problems at this stage depends on data. He sees hallucinations as a problem that better post-training can solve but classifies it as a product issue. The shortage of AI specialists, he says, is temporary and has already eased noticeably.

DeepSeek is managed along two lines: top-down for company-wide projects such as the V4 release, and bottom-up for free research. Formal work takes no more than half of an employee’s time, overtime is rare, and releases come out roughly every two to three months. As headcount grows, some units will get a clearer structure. Even if technology stalls, Liang says, API sales alone could sustain a public company: “We want a bigger dream, but we also have a guaranteed minimum.”