Liang Wenfeng's Three Big Bets: Why Did DeepSeek Open-Source Its 82% Gross Margin?

Title

Liang Wenfeng's Three Big Bets: Why Did DeepSeek Open-Source Its 82% Gross Margin?

Meta Description

Why did DeepSeek^Ad open-source DeepInfer, its inference framework with 82% gross margin, and then launch a hundred-billion-level funding round? An analysis of Liang Wenfeng's three big bets and the business and idealistic logic behind the open-source strategy, from technical roadmap and domestic chips to organizational purity.

Article Content

When a company open-sources its core inference framework with a gross margin as high as 82% to the whole world, and then immediately announces the launch of a hundred-billion-level funding round, the first reaction from outsiders is often "self-contradictory." But if it's Liang Wenfeng, all of this has explanatory clues. Based on an in-depth review by a Bilibili creator and combined with various public information, this article attempts to reconstruct the strategic logic behind this series of DeepSeek actions.

1. Open-Sourcing the "Money Printer": What Exactly Did DeepInfer Deliver?

According to analysis by a Bilibili creator, DeepSeek's open-source DeepInfer inference framework is essentially a scheduling system that makes AI run faster and cheaper. Previously, open-source models were like handing over a recipe—telling you what ingredients to use and what dishes you can cook; whereas what DeepInfer hands over is the cost and efficiency management system for running a restaurant: how to control costs, how to serve food quickly, how to manage the heat, and how to debug the equipment.

The economic value of this framework is extremely astonishing. Calculated at the lowest price, each graphics card can generate $38 per hour in revenue, while the cost of renting a card is about $7 per hour, yielding a gross margin of 82%. It is said that this is the first time "how profitable AI really is" has been brought from a black box onto the table. For any company, such a profit model would be an absolute commercial secret, but Liang Wenfeng chose to make it public.

Immediately afterward, news of a new funding round arrived—approaching 100 billion in scale, with institutions such as Tencent continuing to increase their investment. Liang Wenfeng once stated clearly at an investor exchange meeting: "We will definitely open-source, even our strongest model we will open-source." This sentence is almost impossible to explain in business textbooks: Would Google open-source its search technology before financing? Would SpaceX disclose its rocket landing control algorithms before going public?

Inference framework is the software layer between AI models and underlying hardware, responsible for efficiently scheduling trained models onto GPUs for execution. It determines how a request is split, when to do batching, how to allocate VRAM, how to fuse operators, and directly affects how many tokens can be processed per second and the compute cost per token. Common inference frameworks in the industry include vLLM, TensorRT-LLM, SGLang, etc. If DeepInfer can truly achieve the described 82% gross margin, it means it has significant advantages in scheduling efficiency, and this kind of efficiency improvement is precisely the core bottleneck in the current commercialization of AI.

2. Ban and Countermeasure: Open Source as a Competitive Weapon

The timing of the open-sourcing is intriguing. According to the material, two days before the open-sourcing, the U.S. Department of Homeland Security ordered all government agencies to stop using DeepSeek, requiring deletion of installed versions, and even third-party products developed based on its open-source code had to be investigated. The reason is still "possible risks," but after searching through the documents, no specific case can be found.

On the other side, the U.S. AI community, while loudly proclaiming AGI and giving AI a soul, uses math problems to issue challenges, blocks Chinese users, and accuses distillation. However, DeepSeek's newly released version once again pressed the "kill line" of cost and performance onto the backs of top U.S. AI. Before the suspicion of distillation could be substantiated, the ban arrived.

Facing this situation, DeepSeek senior researcher Chen Deli posted a response: "Open source and closed source, light and darkness, two futures, a decisive battle between two destinies of humanity—however long they want to fight, we will fight for that long." On the same day, DeepInfer was open-sourced.

According to cross-analysis by the Bilibili creator and YouTube tech channels, the core logic of this strategy is: you can block a product, but you cannot block a method. When countless AI companies around the world adopt this inference framework with 82% gross margin, how interfaces are written and how data is stored, DeepSeek's standards may become the industry's default common language. This is similar to placing freezers with your own logo in all competitors' stores for free—the harsher the blockade, the more aggressive the open-sourcing.

3. Three Big Bets: Technical Roadmap, Competitive Qualification, and Organizational Purity

Bet One: Open-Source Technical Roadmap

Before entering AI, Liang Wenfeng was already a quantitative tycoon equally famous for controversy and wealth. To train models, he hoarded tens of thousands of graphics cards in one go and spent billions on a supercomputer. In July 2023, DeepSeek was founded, and he turned to large models. In January 2025, R1 was released, with all weights made public—anyone could use, modify, and commercialize it for free. Its performance was comparable to the then most expensive OpenAI O1, and the training cost was only $294,000, less than 1% of GPT-4.

What was open-sourced was not just the weights. The chain of thought—the capability OpenAI locked in a black box and charged by the word—DeepSeek directly opened up the thinking process, and even wrote "welcome to distill your own model with R1" into the license agreement. In terms of caching algorithms, the MLA multi-head latent attention method compresses the KV cache by more than 90%, making long contexts no longer expensive and multiplying inference throughput several times. Microsoft brought it onto Azure, Amazon onto Bedrock, the open-source community reproduced it based on R1, and companies in various countries used it as a base to modify their own models. According to the material, 11 days after R1's release, OpenAI urgently launched O3 mini to respond—the first time a closed-source giant changed its schedule for an open-source model.

MLA (Multi-head Latent Attention) is an attention mechanism improvement proposed by DeepSeek in V2 and continued through R1. Traditional Transformers, when generating long texts, need to cache key-value pairs for each token (i.e., KV Cache). The longer the context, the larger the VRAM usage. MLA greatly reduces the cache size by compressing keys into a low-dimensional latent vector space, with a compression rate reportedly reaching over 90%. This design significantly lowers the VRAM cost of long-context inference, and throughput increases accordingly. It is also one of the key technical pillars that enables DeepSeek to push API prices extremely low.

Bet Two: Unfettered Competitive Qualification

Facing all-around containment, a Chinese tech worker who wants to touch the most cutting-edge AI has only two paths: go to the U.S. to obtain top capital and chips, or stay in China and start from scratch. Liang Wenfeng chose the latter.

After completing a 50 billion funding round earlier this year, DeepSeek announced the construction of a super-large data center in Ulanqab, Inner Mongolia, deploying at least 160,000 Huawei Ascend chips, with orders worth about $2.56 billion. There, electricity costs a little over 0.3 yuan per kWh, and the annual average temperature is 4.3 degrees Celsius—naturally suitable for building computer rooms. This is a bet on China's globally unmatched power infrastructure, and also an ultimate test of domestic AI chips.

Liang Wenfeng has said that Nvidia has no technical secrets; the real moat is CUDA—millions of programmers worldwide have written code on it for 18 years. Chips can be caught up with, processes can be changed, but the software ecosystem cannot be dismantled. On September 30, DeepSeek open-sourced the entire set of tools running on Huawei chips and placed them in the Huawei community, giving every company in the world that wants to switch to Chinese chips a possibility to bypass CUDA. The cost is also obvious: the Ascend ecosystem has just started, operators need to be adapted one by one, frameworks need to be rewritten, and performance of the same model needs to be tuned bit by bit after migration. Training remains a tough bone highly dependent on Nvidia; the 160,000 Ascend chips mainly handle inference tasks, and due to production capacity constraints, delivery may take more than a year to complete. Other companies can directly start running with ready-made CUDA, while DeepSeek has to start by building roads.

CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model launched by Nvidia since 2006. Developers use it to call GPUs for general-purpose computing. After nearly two decades of accumulation, CUDA has formed a huge ecosystem including compilers, math libraries, communication libraries, and debugging tools. The vast majority of AI frameworks worldwide (PyTorch, TensorFlow, etc.) prioritize CUDA adaptation. This means that even if other manufacturers' chips are close to Nvidia in paper compute power, software migration and tuning costs still constitute a huge barrier. For domestic chips to be truly usable, in addition to hardware performance, the more critical thing is to establish a software stack that can support mainstream frameworks—this is exactly the core difficulty referred to in the text as "bypassing CUDA."

Bet Three: Organizational Purity

The first two bets are only means; the truly difficult one is the third bet—internal organizational purity. DeepSeek has more than 100 people, no assessment metrics, no clock-in in the morning, and most employees leave work at six or seven in the evening. In contrast, in Silicon Valley, core researchers at OpenAI and Google work 70 to 80 hours a week, and camping beds are already prepared.

Liang Wenfeng's explanation is very plain: it is hard for a person to work at high quality for more than 6 to 8 hours a day, and confused judgments from overtime fatigue waste precious compute instead. How to compete with Americans without overtime? His answer is open source itself. Research results are downloaded, modified, and deployed worldwide, and employees can directly see how their work affects the industry—this sense of achievement, more than KPIs, can cultivate a group of pure researchers. Researcher Liu Shengyu once wrote: "The most cutting-edge intelligence should be supplied to everyone in an open and cheap way." In the hearts of these researchers, light and darkness are not literary rhetoric, but a contest over two futures.

KPI (Key Performance Indicator) is a quantitative assessment system commonly used by enterprises, evaluating employee output and allocating incentives by setting measurable goals. In AI research, KPIs easily lead to short-term, publishable, quantifiable results while avoiding high-risk, long-cycle basic exploration. DeepSeek's practice of canceling assessment metrics and not requiring clock-in is essentially using organizational design to protect research freedom, shifting incentives from "completing metrics" to "results being adopted by the industry." This model is not without precedent in small, high-density research teams, but scaling it to over 100 people and continuously producing frontier models is still rare, and its sustainability is also a focus of external attention.

4. Restraint Is Attack: The Business and Ideals Behind Open Source

Compared with OpenAI and Anthropic, DeepSeek's choice appears almost pure: not chasing user volume, not building a super app, and temporarily not touching specialized fields such as finance and healthcare. Liang Wenfeng does not even want excess profits beyond electricity costs. He does not want to become the next traditional big company, and even less wants to lose the original intention of openness like OpenAI and become a capital monster that devours resources.

But open source by no means implies a lack of aggressiveness. If DeepSeek were to build a super app, force enterprise entry points, and charge all companies with closed models, it would immediately become the common enemy of Alibaba, Tencent, ByteDance, Zhipu, and Moonshot. Giving up part of direct control and technological monopoly in exchange for the entire industry viewing DeepSeek as an industry lighthouse^Ad. It does not occupy all users, but lets more and more products be built on its results; it does not control every application, but influences the entire industry's technology costs and research directions.

According to the material, Liang Wenfeng said at a closed-door meeting in 2026: "DeepSeek has only one main line, which is AGI; everything else is strategy. The more restrained you are, the more likely you are to accomplish this." The three big bets are step by step: the first bets on the open-source technical roadmap, the second seeks unfettered competitive qualification, and the third aims at the ultimate goal of AI.

5. Multiple Perspectives and Points to Be Verified

A note: the above narrative mainly comes from the Bilibili creator's review and carries a strong opinionated tone. Regarding specific figures such as "82% gross margin," "100 billion funding," and "160,000 Ascend orders," there is currently a lack of complete confirmation from DeepSeek official sources or authoritative financial media, so readers should remain cautious. At the same time, although details of the U.S. ban and the motives behind OpenAI's emergency schedule adjustment have been mentioned in multiple reports, the rigor of the causal relationship still leaves room for discussion.

The confirmable consensus is: by continuously open-sourcing underlying technologies, DeepSeek has substantively influenced the cost structure and development paradigm of the global AI industry; its binding with the Huawei Ascend ecosystem is an important experiment in China's AI compute autonomy. As for whether this big bet can ultimately reach AGI, the answer is still on the way. Liang Wenfeng's purity and ambition may be exactly as described at the end of the material—what he wants has never been spectatorship, but constantly reaching unknown lands that humanity has never explored.

One-line Summary

With an 82% gross margin inference framework open-sourced as its spear and hundred-billion funding as its shield, DeepSeek's three big bets point directly at AGI.

Paragraph Summary

Based on a Bilibili creator's review, the article reconstructs the strategic logic behind DeepSeek's series of actions: open-sourcing the DeepInfer inference framework with a gross margin as high as 82%, which seems self-contradictory with hundred-billion funding, is actually a competitive weapon using open-source standards to bypass the U.S. ban and fight for industry discourse power. Liang Wenfeng's three big bets point respectively to the open-source technical roadmap (full disclosure of R1 weights and chain of thought), unfettered competitive qualification (160,000 Ascend chips in Ulanqab and a breakthrough in the CUDA ecosystem), and organizational purity (no KPIs, no clock-in). The article argues that restraint is attack: DeepSeek gives up direct monopoly in exchange for the position of industry lighthouse, with AGI always as the main line, while noting that many specific figures still await official confirmation.