SVTR Editor's Note: An unverified closed-door recording is being circulated in the Chinese venture capital community as a strategic manual for DeepSeek, with the most attention focused on the most unreliable and least appropriate numbers such as card quantities, procurement, and payback periods. SVTR does not circulate such numbers, nor does it treat a voice transcription as an official statement. Instead, we do one thing: break down the same set of decision-making logic that Liang Wenfeng repeatedly uses, for entrepreneurs and investors who want to understand this business. We examine what it has proposed on four issues: domestic computing power, open source, the computing power gap, and the technological endgame, and which of these are solid and which are still just aspirations.
On May 20, 2026, during the first external funding round of DeepSeek, the founder Liang Wenfeng had a four-hour closed-door meeting with investors. According to the SVTR AI venture database, this round was finalized on June 17, with an estimated valuation of about $56.3 billion. The meeting was originally not open to the public, and the transcription of the recording spread in the Chinese venture capital community in July.
The transcription was generated by automatic speech recognition, without speaker differentiation, and has not been confirmed by DeepSeek. This article does not treat it as an official statement, nor does it restate the most sensitive numbers such as card quantities, procurement, and financial forecasts. Instead, we extract four more worthy of long-term tracking judgments and mark how far each has been validated and how it can be refuted. Although the four questions seem scattered, they all point to the same issue: when opportunities arise simultaneously, what does DeepSeek rely on to decide what not to do.
I. What Needs to Be Disassembled Is Not the Chip, but the CUDA Dependency
The most widely shared sentence in the entire conversation was: NVIDIA is digging its own grave.
This sentence is easily misread as domestic chips about to fully catch up with NVIDIA, but Liang Wenfeng's original intention was not that: the most solid software barriers of NVIDIA are being gradually weakened by AI and new programming tools.
In the past, developing high-performance AI operators required engineers to deeply understand CUDA, hardware architecture, and low-level optimization. Migrating to another chip often meant rebuilding an entire set of software support. Now, AI can assist in writing code, and higher-level programming languages are also lowering the barrier for operator development.
TileLang is a representative of this path. It is a domain-specific language for high-performance GPU and accelerator kernels, aiming to allow developers to write and migrate matrix calculations, Attention, and other core operators with less code. DeepSeek has open-sourced the TileKernels operator library based on TileLang, and some of the operators are reportedly used in internal training and inference.
But the boundary must be clear: TileLang currently cannot replace the complete CUDA system, and TileKernels still depends on NVIDIA GPUs and the CUDA Toolkit (currently requires CUDA 13.1 or higher, compatible with Hopper and Blackwell). It can weaken the dependence of high-performance kernel development on hand-written CUDA, but not replace the entire software stack composed of drivers, compilers, communication libraries, and toolchains in one night.
So this sentence means: as AI programming, compilers, and specialized chips develop together, the migration costs accumulated by CUDA may decrease. Whether domestic chips can truly seize the opportunity ultimately depends on capacity, training, and migration costs: whether they can stably produce enough chips, whether they can prove the reliability of clusters in real large model training, and whether the migration costs are acceptable.
SVTR's judgment: the valuation anchor for domestic computing power should shift from "whether it can be compatible" to "whether it can train." The real revaluation point is only one: the appearance of a mainstream flagship model that is completely trained on domestic chips, with publicly available results and reproducible engineering paths. No matter how many compatibility tools are released, they cannot reach this level of proof. Improvements in the software layer are necessary conditions, but not the final proof.
II. The Confidence in Open Source Comes from Cost Advantage, Not Idealism
The first section discusses how barriers are loosening, and this section discusses how DeepSeek itself uses cost as a weapon. Liang Wenfeng groups open source, low prices, and business models into the same logic: when the operating cost of official services is low enough, opening up model weights will not automatically destroy the API business.
After users get the open-source weights, they still need to purchase hardware, build clusters, optimize operators, maintain services, and handle peaks. Most enterprises really compare only one thing: whether the total cost of self-deployment can be lower than the API price of the model company. Whether the weights are free is secondary.
This is also why DeepSeek emphasizes computational efficiency. For resource-rich large companies, some problems can be reliably solved with additional GPUs; for teams constrained by computing power, improving efficiency not only optimizes profit margins but also directly determines how large a model they can train and how many experiments they can conduct. Low cost therefore serves multiple purposes: expanding usage scale, compressing the economic attractiveness of third-party independent deployment, and allowing DeepSeek to use limited computing power to advance larger-scale research.
The price can be seen as an example of this approach. The transcription was made during the V3.2 period, and V4 has been recently released: according to the official 2026 July pricing, the cost per million output tokens for V4 Flash and V4 Pro is $0.28 and $0.87, respectively, less than one-tenth of the same-tier models from OpenAI, and at the lowest level of global mainstream models.
However, the specific multiples of "ten months to break even," "six times profit," and third-party deployment costs in the transcription lack cost structure and calculation cycles, and are the口径 of the parties involved, and cannot be directly understood as standard financial indicators. A reliable conclusion is: DeepSeek is trying to set a price that can cover R&D and infrastructure investment while also compressing the profit space of intermediate deployers.
Following this logic, an industry fact is taking shape: the model API layer is actively giving up on high profits. This raises more stringent questions for AI application investments: when model capabilities continue to improve and prices continue to fall, what assets do middle-layer companies have that will not be swallowed by the base model? Simply reselling tokens and packaging workflows is difficult to form a long-term barrier. Valuable application companies must have independent data loops, business execution capabilities, or result delivery capabilities.
III. The U.S.-China Gap Is Not Just Computing Power, But Computing Power Determines the Right to Experiment
Liang Wenfeng attributes most of the gap between U.S. and Chinese cutting-edge models to computing power. His self-narrative is: behind the U.S. by about two years, and achieving the same results with only one twentieth of the U.S. computing power.
This is not without reason. Computing power determines not only the scale of training a large model once, but also how many experiments a research team can complete before formal training and how many wrong paths they can eliminate. Ten times the computing power is not just making the model ten times larger, but also trying more architectures and data solutions, and quickly restarting after failure. Computing power essentially buys the right to experiment and time advantage.
DeepSeek's past narrative was to narrow the model capability gap with far less resource investment than U.S. top laboratories. The next challenge is to further shorten the time lag in capability while maintaining efficiency advantages.
However, attributing all gaps to computing power still narrows the problem. In the same conversation, Liang Wenfeng also admitted that high-quality data and post-training are another key constraint, and mentioned that a considerable portion of the company's core researchers' efforts are invested in data annotation. Computing power can be rapidly increased through procurement, but high-quality data production, researcher collaboration, and training method accumulation take time and cannot be immediately supplemented by capital expenditures alone.
SVTR pays more attention to the shift in competition standards: the capital competition of Chinese model companies is shifting from "who raises more" to "who can turn computing power into model capabilities." When evaluating a basic model company, it should not only look at the number of GPUs, but also at how many effective experiments each unit of computing power generates, whether new computing power is truly converted into capabilities, and whether data and post-training teams are expanding in tandem. Buying cards as assets only holds when the cards can run stably, be efficiently utilized, and continuously produce model progress. Otherwise, expensive clusters can also be a rapidly depreciating burden.
IV. After the Agent, the Real Step Is Continuous Learning
The first three sections discuss the current offensive and defensive strategies, and this section is Liang Wenfeng's judgment on the endgame, which is the most worthy of long-term tracking.
He understands the development of large models as a series of continuous steps: language models solve knowledge expression, chain-of-thought improves the reasoning limit, agents expand the range of tasks, and the next step is continuous learning.
Today's agents can call tools, break down tasks, and execute complex processes, but they still highly depend on humans providing context. After two months of work, a new employee gradually understands the relationships between people, internal rules, and implicit experiences, but most AI systems do not convert these experiences into long-term capabilities after the task ends. Continuous learning aims to solve: how can AI absorb new experiences without complete retraining, retain long-term memory, and transfer to the next task, while not destroying existing capabilities.
Liang Wenfeng also admits that this is still a global problem that has not been solved. DeepSeek internally refers to this kind of research as "hunting for prizes," meaning that it does not necessarily require the largest scale of training resources immediately, but requires a large number of researchers to continuously explore possible breakthrough paths.
This also explains why DeepSeek places "the model first serving its own R&D" in an important position. If AI can participate in data processing, write training code, analyze experiments, and help design the next generation of models, AI R&D will no longer be a linear process: each generation of models not only provides product capabilities, but also improves the R&D efficiency of the next generation. This is closer to the real goal that frontier laboratories are competing for, which is who can establish a positive feedback loop of AI accelerating AI research first.
It is also necessary to remain restrained here. Continuous learning does not equal increasing memory databases or expanding parameter scale; a model can remember conversations, but it does not mean it has formed transferable new capabilities. To determine whether it is fulfilled, it cannot only look at whether the product has a "long-term memory" button, but also whether the model can continuously absorb feedback in long-term tasks and stably improve behavior without large-scale retraining.
What Liang Wenfeng Truly Delivered Is a Set of Trade-off Principles
The answers to the four questions are different, but the underlying logic is the same. The most worth recording in this closed-door meeting is not a set of unverified card quantities or revenue numbers, but how DeepSeek understands its own boundaries.
It wants to keep basic models, computing efficiency, and AGI research in-house, and to leave more application opportunities for external partners; it lowers industry resistance through open source, expands usage through low prices, compensates for computing power gaps through efficiency, and keeps commercialization within a range that does not change the research priority. The four judgments are actually answers to the same trade-off question on four battlefields.
This logic is not inherently correct. Domestic computing power may take a long time to verify through flagship models, continuous learning may have no breakthroughs for years, and low-price APIs may not be sufficient to cover rising R&D costs. As the team, capital, and customer scale grow, DeepSeek will also face the same issues that all tech companies cannot avoid: how to balance research goals and commercial responsibilities.
But the rarity of Liang Wenfeng lies in his ability to explain all choices with the same goal, rather than packaging open source, low prices, domestic substitution, and AGI into independent slogans. This rarity has already spread into a phenomenon: in the past two months, SVTR has encountered more Silicon Valley funds asking about DeepSeek's organizational model than about its model capabilities, and this will only increase after the release of this transcription.
This set of trade-offs is documented in the SVTR AI venture database. On the capital side, the A-round investors of DeepSeek are not led by financial VCs, but by the National Artificial Intelligence Industry Investment Fund, with strategic capital such as Tencent, JD.com, and NetEase, and Ningde Technology, which aligns with the statement in the transcription that financing should choose "the most consistent interests and least hostile." The investment of the power battery giant Ningde Technology coincides with Liang's analogy of BYD batteries and the calculation of computing power as electricity.
On the talent side, Liang Wenfeng's relationship map shows that the core team of DeepSeek is almost entirely from his quantitative fund Huafang and the same source as the Zhejiang University School of Information and Electronic Engineering, with very few external hires, which is the structural foundation of his statement that "team stability is the only core interest."
For entrepreneurs and investors, the key is to learn this decision-making approach, as the answers themselves cannot be replicated: when opportunities arise simultaneously, which things can increase the probability of reaching the endgame, and which are just early realization of unestablished advantages?
Signal Tracking
Can domestic chips complete the training of flagship models (Verification window: before mid-2027). Observe whether a mainstream flagship model that is primarily trained using domestic chips appears, with publicly available model results and technical explanations. Only when the training results appear will the claim that "domestic software stack is not available" be truly refuted.
Can TileLang cross the single-point operator optimization (Verification window: 12 to 18 months ahead). Observe whether TileLang and similar tools can cover the entire cross-platform compilation, communication, and deployment process from kernel development. Whether it can form large-scale production cases on non-NVIDIA hardware is more important than the project's popularity.
Can continuous learning go beyond "long-term memory" (Verification window: by the end of 2026 to 2027). Observe whether cutting-edge models can absorb new experiences in long-term tasks, stably update behavior, and transfer learning results to subsequent tasks. If the changes are still limited to context expansion, memory retrieval, and parameter scale increase, it means the next step has not been crossed yet.
Can low-price APIs support long-term R&D (Verification window: after the next round of financing or the release of public financial information). Observe whether the company's revenue, infrastructure expenses, and R&D investment can form a sustainable cycle. If the API and enterprise business can cover the main operating costs, the model of "commercialization as a byproduct of research" will be initially validated; if not, the capital market will require it to redefine its commercial priorities.
SVTR Comment: The value of this transcription is not in the numbers, but in the method. All four judgments may be overturned by subsequent facts, but the fact that they are all explained by the same set of trade-off logic is something most hot-topic companies cannot provide. To see how this logic is implemented on the capital and talent sides, you can check the DeepSeek Capital Map and Liang Wenfeng's Relationship Map in the venture database: the former is the shareholder structure and valuation trajectory, and the latter is the team continuity from Huafang to DeepSeek.
Material Notes: This article is based on the transcription of a closed-door investor meeting between Liang Wenfeng and investors on May 20, 2026. The transcription was generated by automatic speech recognition, without speaker differentiation, and some proper nouns and numbers may be incorrect, and has not been confirmed by DeepSeek. The references in the article are for research and commentary purposes only and do not represent the official position of DeepSeek. The API pricing is taken from the official July 2026 version of DeepSeek, and the TileKernels information is taken from its public code repository. The financing, valuation, and shareholder data are taken from the SVTR AI venture database (A-round, as of July 2026); the database also records an unconfirmed $7.1 billion valuation round in July, which is not cited as a fact in this article.