The DeepSeek Investor Q&A with Liang Wenfeng: Decoding Restrained Pricing and the AGI Roadmap

A company that doesn't do overtime, has no KPIs, lets employees spend half their time on whatever research they want, prices its products to earn only a reasonable profit while actively pushing prices down, open-sources all its models—and plans to open-source even its strongest one. By any normal business logic, such a company should be struggling to survive. But this company is called DeepSeek.

Recently, an audio recording from a Liang Wenfeng investor Q&A—nearly four hours long and transcribed into 35,000 characters—was fully analyzed on Bilibili. This article is based on that transcript (figures follow Liang Wenfeng's own on-site statements, with the possibility of minor transcription errors) and, from an engineer's perspective, distills the parts of this material most likely to change how you think. This isn't a pile of industry platitudes—it's a set of logic that interlocks tightly.

The Counterintuitive Pricing: 6x Profit and Still Too Much

The Q&A was an internal meeting aimed at investors. Liang Wenfeng first spoke for two and a half hours, then opened the floor for questions—topics ranging from vision, pricing, and open source all the way to the AGI roadmap, compute, and domestic chips. The most counterintuitive part was—money.

Liang Wenfeng set a pricing rule for DeepSeek's API: once a batch of hardware is purchased, recover the cost within 10 months. Financially, hardware is typically depreciated over 3 or 5 years, but he believes that from a business standpoint, recouping costs in 10 months is enough. By his own estimate, this corresponds to roughly 6x profit.

Many people would question this: you call 6x profit restrained? The key lies in his next judgment—in this price range, demand is inelastic; doubling the price again would make little difference to token consumption. In other words, he could have nearly doubled his revenue, but chose not to. His explanation is blunt: lowering the price further wouldn't increase demand, because at this price point everyone can already afford it. Cutting prices further brings the company no more revenue, and society no more value.

There's a telling detail in the transcript: for one particular model, they initially worried demand would be too high and set the price on the higher side, which made the team unhappy. Later, Liang Wenfeng cut the price to a quarter of the original, and the company chat erupted in cheers. A company cutting prices, and employees cheering—that's not common in engineering circles.

Why Open Source Doesn't Hurt Revenue

Many investors asked: does open source hurt revenue? Liang Wenfeng's answer was crisp: it doesn't—provided you only earn 6x profit.

This logic is actually quite airtight. When you price for a 10-month payback, a third party wanting to deploy the model themselves to steal your business can't match that price or that cost—the barriers to deployment are high and costs can't be driven down. So even after you open-source, others can't take your business. But if you wanted to earn 100x profit, open source would genuinely hurt you, because even a third party operating at 20x your cost would still be cheaper than you.

Every link in this restraint corresponds to a calculation: open source doesn't hurt revenue because pricing is kept low; pricing is kept low because a 10-month payback is enough; a 10-month payback is enough because the AI market is simply enormous—his exact words were, "even a tiny slice of the pie is already huge." That's why he dares to say even the strongest model can be open-sourced: even if you hand over the entire model, the barriers to using it well remain high, and matching the same low cost is even harder.

Following this logic, many things make sense. Last Chinese New Year, the consumer side suddenly exploded, with users "impossible to chase away"—yet they didn't seize the moment to monetize, didn't grab users, didn't build the next super app. As Liang Wenfeng put it: everything ahead is just sesame seeds; the bigger watermelon lies behind. Last year's consumer side, this year's enterprise side—both were picked up along the way to AGI. They pick them up, but they don't stop.

In this entire playbook, only one thing cannot be compromised—team stability. He says this is the sole core interest: as long as no one leaves, this will eventually get done. Money and resources aren't the problem; everything else can be conceded.

The AGI Roadmap: The Path to Ladder-Step Development

This is the most valuable part of the entire transcript. Liang Wenfeng views AI development as a ladder, where each step stands on the one below it.

Language models sit at the very bottom. Last year came CoT (chain of thought), letting the model think for itself and raising the ceiling of intelligence. This year came Agents, capable of handling multi-step tasks and expanding the scope of capability one more notch. Agents rely on CoT, and CoT relies on language models—no step is wasted.

On Chain-of-Thought (CoT): Chain-of-Thought is a prompt engineering technique systematized by Google Research in 2022. The core idea is to have a language model explicitly output intermediate reasoning steps before giving its final answer, mimicking the human process of solving problems step by step. Traditional models map directly from input to output, whereas CoT introduces the "let's think step by step" paradigm, significantly improving model performance on complex tasks like mathematical reasoning and logical inference. Models like DeepSeek-R1 deeply integrate CoT into the reinforcement learning training pipeline, enabling the model to autonomously generate reasoning chains rather than relying on manually designed ones—this is the key mechanism behind its outperformance of comparably sized models on reasoning benchmarks.

On Agents: An AI Agent refers to an AI system capable of perceiving its environment, autonomously planning, and executing multi-step tasks—distinct from a language model that merely does single-turn Q&A. An Agent's core architecture typically comprises four modules: a perception layer (receiving input), a memory layer (storing context and history), a planning layer (task decomposition based on CoT), and an action layer (invoking tools, APIs, or code executors), forming a closed "think-act-observe" loop. Coding Agents are currently the most mature Agent direction, because code execution results can be precisely verified (whether it compiles, whether tests pass), providing a natural reward signal for reinforcement learning—this is the underlying technical logic behind Liang Wenfeng's judgment that "coding agents will mature first."

Beyond Agents: The Critical Leap of Continual Learning

What comes after Agents? He says they'll solve every problem that can be solved, reaching the ceiling of their own capabilities—but they still can't replace employees. The missing piece is called continual learning.

Continual learning, also known as lifelong learning, refers to an AI system's ability to keep learning from new experiences without forgetting what it already knows. The fundamental limitation of current large language models lies in their "static weights"—once training is complete, the parameters are frozen and cannot automatically update knowledge from post-deployment interactions. They can only rely on a limited context window to temporarily remember the current conversation. In academia, this problem is known as "catastrophic forgetting": when a model updates its parameters on a new task, it often overwrites old knowledge.

He gave an example worth everyone's careful attention: you hire a new employee, and they spend two months getting to know the company and the work. Afterward, you say "go get Xiao Wang," and they know who Xiao Wang is. But if you tell the AI "go get Xiao Wang," you first have to tell it who Xiao Wang is, what their role is, where they are, how to find them, and what to watch out for—you have to feed it all the context. This is unrealistic. So until continual learning is solved, AI still can't replace those jobs that "live on context."

Further along comes self-iteration: once a model can continually learn, it can develop its own next version. Liang Wenfeng believes there's no critical threshold here—it's a gradual but nonlinear process, because AI can accelerate AI research. Only after self-iteration does it become the turn of embodied intelligence—AI stepping into the physical world to do housework and elderly care. Embodied intelligence requires AI to have not only powerful reasoning and planning abilities, but also to build a "world model" of the physical world—predicting the consequences of actions and forming a perception-action loop with the real environment. This is precisely why Liang Wenfeng places it at the very end of the roadmap: it presupposes the full maturity of all capabilities in the digital world.

The Practical Significance of the Roadmap

This roadmap has a very practical benefit: each step adds very little new material, which is why he says this path is the most relaxed—no overtime needed. Looking back at "no overtime, no KPIs," the root cause lies in the choice of roadmap. He also calmly judges: video/world-model generation is not on the main line of intelligence. When Sora came out, companies large and small all rushed to do it, but the small ones later abandoned it—it has nothing to do with the ceiling of intelligence. Commercially it's a good business, but it's not their goal.

Real-World Constraints: Compute, Financing, and the State of Chips

No matter how beautiful the roadmap, reality is a bottleneck—chips. Liang Wenfeng laid out the family assets quite concretely: right now they have roughly 20,000 H-equivalent units of compute, most of which arrived just one or two months ago. How is the funding spent? "At a reasonable price, we buy as many chips as we can; spending the money within six months would be the ideal scenario." But in practice it's not achievable—chips are too hard to buy.

The Gap with the U.S.: Scale and Resources

On the gap with the U.S., he's blunt: it's a resource gap; on talent there's almost no gap—they were the same batch of people to begin with, some stayed in China, some went abroad, and the distribution was basically random. In terms of scale, foreign players are training models with 800B active parameters, while domestic players are still at the tens-of-billions active scale—an order of magnitude behind.

Here it helps to understand the background of Scaling Law and the MoE architecture. The Scaling Law was systematically articulated by OpenAI in 2020, with the core conclusion that model performance follows a power-law relationship with parameter count, data volume, and compute, and no clear ceiling has yet been observed within the known range. "800B active" refers to the number of parameters actually activated per inference under the Mixture of Experts (MoE) architecture—for example, an MoE model with 2 trillion total parameters activates only 800B of them per token, making ultra-large-scale training feasible while keeping inference costs manageable. Training an 800B-active model requires 50,000 GB300 units, or 200,000 units of Huawei's latest chips—and that's just training. He says even spending all of the 50 billion yuan wouldn't be enough to train it. So the strategy is very pragmatic: run experiments thoroughly at the tens-of-billions active scale, and only push toward 150B and 250B once resources allow.

As for whether scaling has run its course, his answer is crisp: he believes in scaling—the larger the scale, the better the results, and he hasn't seen a ceiling yet. The only thing holding them back is compute. "Silicon Valley says scaling is over—that's Silicon Valley being over. China is still far from that wall."

The Reality and Prospects of Domestic Chips

This is the section he discussed in the most detail. Huawei gave them 16,000 units of the 950, and he bluntly says this only amounts to 4,000 units of the B series—4 Huawei 950s equal 1 GB300, they're two years behind in timing, and Huawei is also 50% to double the price. But he says it doesn't matter, because "you can't buy NVIDIA's chips anyway." Buying Huawei is, to a large extent, helping it build its ecosystem.

From an engineer's perspective, "4 to match 1, two years behind" has another side: alternatives that were unimaginable two years ago now exist—the cost being buying more chips and waiting longer. The chip problem is essentially a money problem, and money is something that can be figured out in China.

More critical is the ecosystem. In the past, the fatal weakness of domestic chips was the CUDA moat—CUDA (Compute Unified Device Architecture) is a parallel computing platform NVIDIA launched in 2006, which over nearly two decades has accumulated thousands of optimized libraries like cuDNN, cuBLAS, and NCCL, with the underlying operators of nearly all mainstream deep learning frameworks deeply dependent on it. The reason you buy a domestic chip and can't put it to use isn't hardware performance—it's this ecosystem lock-in.

Liang Wenfeng believes this moat is crumbling, and the key weapon is TileLang—a high-level operator description language whose design goal is to decouple operator logic from specific hardware backends: developers describe computation logic in TileLang, and the compiler is responsible for translating it into optimized instructions for different hardware (NVIDIA GPUs, Huawei Ascend, AMD GPUs, etc.). AI can write code, which makes building an ecosystem easier than before; TileLang then provides a technical path to systematically rewrite NVIDIA's entire operator ecosystem.

DeepSeek's approach on NVIDIA chips is: V3 uses NVIDIA's chips, but not the official software libraries—the operator layer is written by themselves using TileLang, with an efficiency loss of about 1% to 2%. As for how efficiency holds up when ported to Huawei chips, there's no public measured data yet. But by Liang Wenfeng's account, doing the same thing over again on Huawei chips would, within a year, overturn the perception that domestic chips have an "ecosystem problem." He even judges that, given the reality of not being able to buy NVIDIA chips, "NVIDIA is digging its own grave."

Data, the Endgame, and Your Moat

There's another figure that makes you sit up straight: half of DeepSeek's core researchers are labeling data—half of the most important people are doing annotation. The reason is that high-quality data annotation is too expensive, and China's capital structure can't support the way the U.S. does labeling, so they walk on two legs, first labeling what's cheap. He emphasizes that at this stage, AI companies rely on data.

The Final Form of Large-Model Competition

On the endgame, his judgment follows naturally: after large-model competition converges, only three things separate players—cost, time, and user experience. Cost comes first, time second (a few months earlier or later makes a difference), and experience has stickiness but isn't fundamental.

Then comes that widely circulated judgment: suppose AI eventually accounts for 20% of global GDP, and you say "I want to take 5%"—in theory that math works out. OpenAI is running the same calculation. But he'll be beaten by someone who says "1% is enough for me," and then beaten again by someone who says "0.1% is enough."

"If your vision is to take a lot, you've already lost." The earlier restraint, the price cuts, the open source—here it all connects. The same logic, two sides of the same coin.

The Competitive Advantage of Chinese Models

He's also clear about his own position: one to two years behind the U.S., getting things done with 1% of the compute, with the next step being to shorten the time gap to 6 months, then 3 months. He summarizes the chip gap as "add two years, 4x," and says the ecosystem gap will eventually disappear. Chinese models can, this very year, become a substitute for the existing foreign paradigm—note the wording is "substitute for the current paradigm," not AGI. AGI requires at least continual learning.

He also poured two buckets of cold water: there are too many companies doing foundation models domestically, spread too thin on resources, and it will inevitably converge to three or four; the shortage of AI talent is temporary, and in two or three years the market can cultivate enough.

Your Moat and the Future of Your Career

The single sentence in the entire transcript most relevant to your survival is: until continual learning is solved, AI still can't replace you. The reason is very concrete—it doesn't have your two months of context. Your work contains a large amount of default knowledge: who's responsible for what, how the processes flow, how that pit was filled last time—none of this is in any document, and AI can't access it. This is your most tangible moat right now, but it has an expiration date.

He also hints: at this stage, the most important Agent is the coding agent—do programming first, and push other vertical directions to the back. The field of writing code will be conquered first—because code correctness has an objective standard, the reinforcement learning reward signal is clearest, and model iteration is fastest. So the pragmatic advice is: take stock of your work—which parts are held up by "two months of context," and which parts AI could do once it has the context. The former is your position; the latter is what you should learn to let AI do.

Conclusion

At the Q&A, an investor commented that "10-month payback—the profit is too high," and Liang Wenfeng replied, "There is indeed still room to cut prices." This sentence is basically a microcosm of the company: earning 6x profit and still calling it too much, two years behind and using only one-twentieth of the compute, and still figuring out how to shorten the time to three months.