With a monthly income of US$27,000 per card, Liang Wenfeng pierced the window paper of the AI industry.
In addition to new models, DeepSeek has launched self-developed search, reasoning frameworks and encoding and decoding components within a week. Although the search service will start charging on November 1, the rest are free and open source.
These three products can actually be strung together in a line, that is, no matter what computing platform, harness, or model you use, as long as you use these three, you can use the same DeepSeek product and call the same DeepSeek interface.
Liang Wenfeng’s view on open source and inclusiveness is exactly the same. He once said in a closed-door meeting: “We will not say that open source is bad, and then use a better one when we deploy it ourselves.”
At the same time, DeepSeek also disclosed the specific hourly income of DeepSeek’s single card in the article.
And in the latest article on the DeepSeek official account, there is this sentence: “Do what is good and do it for the public, do it for the sake of skill and do it for the public.”
This sentence comes from “Yin Wenzi·On the Great Road” by Yin Wen, a famous writer in the Warring States Period. Translated into modern language, it means to do good deeds and bring everyone together; to learn skills and skills, everyone must be able to master them.
This is also the most fundamental motivation for Liang Wenfeng to release these three products. Users can use them for free, and DeepSeek can also obtain data streams and entrances, ultimately achieving a win-win situation.
Liang Wenfeng’s approach seems “maverick”. After all, OpenAI, Anthropic, and even all AI companies in the world have never dared to be so candid about open source.
But if you zoom in a little higher, you’ll find that Liang Wenfeng’s style of play is surprisingly consistent with Coca-Cola’s.
01
What are these three products used for?
Let’s briefly talk about these three products first. In one sentence, Search and Fetch are the “purchasers” of the restaurant, responsible for buying ingredients from the outside market; DeepInfer is the “back kitchen”, where the model works; deepseek-recipe is the “multilingual recipe”, so that people from different countries can also recognize what they are eating.
Let’s start with Search and Fetch. This is a search for AI.
The search engines we usually use are for people to see. There are ads, recommendations, and a bunch of links on one page of results that people need to judge by themselves. When you click on them, there are navigation bars, pop-up windows, and sidebars. People can ignore these, but the model cannot. Every extra word will occupy its context. These are tokens and cost money.
DeepSeek’s Search and Fetch are two original JSON interfaces. Their usage is the same as calling the DeepSeek API, and they use the same Bearer key for authentication.
Search is responsible for “searching”. In addition to the title and URL, each result also has a short summary, a long summary, optional web page text and release time. There are also two scores specifically for model reference: rerank_score (relevance re-ranking score) and authority (source authority).
Fetch is responsible for “grabbing”: give it a URL, and it returns the cleaned Markdown text.
In other words, it is not a traditional “search box”, but “material that can be directly inserted into the model context.”
The pricing of Search and Fetch is also designed according to the usage of Agent. Each 1,000 successful requests costs 5 yuan, which translates into only 5 cents per search.
Official documents show that among the three tools of DeepSeek Harness, Claude Code, and Codex, Search and Fetch used through server-side tool calls will not be billed before 00:00 on November 1, 2026, Beijing time, and will be charged on a per-use basis.
In addition, Search and Fetch are not bound to the DeepSeek model. You only need a DeepSeek API to call Search and Fetch on any model.
The second one is DeepInfer.
DeepInfer is the inference framework currently used by DeepSeek’s large-scale online services. What is open sourced this time is its offline inference version and core capabilities. The protocol is Apache 2.0 and can run on NVIDIA and domestic computing platforms.
Many people think that once a model is released and the weight file is downloaded to their computer, the matter is over.
Not really. Weights are just a bunch of numbers, and there is a whole set of projects in between to make them run efficiently. This set of projects is called the reasoning framework.
A model is like a book full of words. Even though the book has its own beauty and golden house, it cannot speak by itself, let alone answer your questions.
DeepInfer is the program that makes this book open. You send it a sentence, and it feeds the sentence to the model file. The numbers in the model file are mobilized, the answer is calculated, and then returned to you.
DeepSeek officials stated that the capability boundaries of the model are delineated by the model itself, and how to enable these capabilities to serve a large number of users at a reasonable cost is the core problem to be solved by the inference framework.
The third one is deepseek-recipe.
Just like greeting in different countries, China says “Hello” and the United States says “Hello”. The speech formats of different models are also different. OpenAI’s format is called Chat Completions and Responses, and Anthropic’s format is called Messages.
Although these are all “asking questions to the model and getting answers from the model” in the end, the formats are different, and even the same dialogue cannot be used between different models.
For example, if you want to run the DeepSeek model on your own server, and your client software is written in the OpenAI format, if you ask the DeepSeek model about the OpenAI format, it will not understand it. At this time, someone has to translate the words.
What deepseek-recipe does is the middle layer of translation. You send a request in OpenAI format, which is translated into internal words (prompt words) that the DeepSeek model can understand, and then compiled into words (tokens) that the model recognizes. After the model answers, it translates the answer back to OpenAI format and returns it to you.
You don’t need to change the format of your input for two different models to still understand it.
The official technical article mentioned two real problems it solved.
One is anti-injection: special tags are originally used to mark roles and boundaries, but users may also deliberately type the same string in the input to pretend to be the system giving orders. What the recipe does is that only the tokens explicitly inserted when the prompt word is constructed will be mapped to the real special token, and the same text in the user input will be treated as ordinary text encoding.
The other is to speed up long text: ordinary text between special tags can be encoded independently, so it first locates the boundary, then encodes each paragraph in parallel, and finally spells it back in order, reducing the word segmentation delay of long context requests.
In real scenarios, these three products can be connected in series. For example, if a company wants to deploy V4.1 Flash by itself, then it needs to use DeepInfer. If you want to match other models, hang up the recipe to convert the format.
The execution layer then uses this call to request Search and Fetch, inserts the cleaned results back into the conversation, and the model continues to output answers.
02
Why is DeepSeek like Coca-Cola?
Another thing these three things have in common is that each of them leaves a door open so that you can use DeepSeek’s products no matter what the scenario is.
If you have enough computing power and want to deploy the same inference capabilities as official DeepSeek, then use DeepInfer and add the already public model weights.
And whether you have an NVIDIA GPU or a domestic GPU, DeepInfer is compatible with them all.
If you already have your own inference backend and want to run DeepSeek’s model, but don’t want to rewrite protocol conversion and prompt word rendering yourself, then use deepseek-recipe.
If you are using another company’s model or Harness, and the network capability is not enough, then connect Search and Fetch.
In the final analysis, it comes down to two words: “arrangement”. Don’t worry about anything, Liang Wenfeng will make arrangements for you.
In addition to this, there is another interesting thing. Liang Wenfeng showed his trump card.
In the deployment practice of DeepSeek, it not only writes “what should we do and how to deploy”, but also writes the specific costs: the throughput of a single cluster is 155 trillion Tokens per day, and the overall revenue is an average of US$38 per card per hour based on the valley price; Pref ill has 60,000 to 1.5 million tokens per second per card; Decode has 14,000 tokens per second per card. The output speed seen by a single user is about 230 tokens per second, and there are about 300 code scenarios; the average first word delay is 400 milliseconds, and P99 is 1,000 milliseconds.
This is equivalent to telling the industry: how to build hardware, how long to keep the cache, and how much revenue a card can generate in one hour. Multiplied by 38 US dollars by 24 hours and then 30 days, one card earns approximately US$27,000 per month.
And on average, 96% of the 77K tokens input in each request hit the cache. After hitting the cache, the throughput per card increases 25 times.
Liang Wenfeng once said at a closed-door meeting with investors: “Our API pricing is a reasonable profit. It probably means that we buy a batch of equipment from the market and recover the cost in 10 months. I think this is a reasonable profit.”
The money earned from each card is essentially the same as our salary, which is usually kept confidential no matter what unit it is in. OpenAI and Anthropic have never disclosed the unit economics of each of their cards.
Liang Wenfeng pierced this layer of window paper, which is also in line with his own philosophy. At the closed-door meeting, Liang Wenfeng said that DeepSeek only makes reasonable profits, and at the same time, people who get less will defeat those who get more.
Judging from the income and the price of DeepSeek’s own API, DeepSeek obviously belongs to the group that gets less. The disclosure of this set of figures also proves what Liang Wenfeng said: “We only earn a reasonable profit, not maximize profits.”
On the other hand, the reason why Liang Wenfeng dared to disclose this is also to explain: “DeepSeek’s moat is not at all based on cost.”
DeepSeek is the Coca-Cola of the AI era. The formula is fully disclosed, but it is not the formula that makes money at all.
In 2011, the Coca-Cola formula was leaked, but nothing happened.
Coca-Cola stated that it does not admit that this is the current formula in use. It is an early historical manuscript and the current mass-produced version has been iteratively modified.
The “This American Life” podcast team asked the flavorists from the Jones Soda Soda Factory to help reproduce the recipe. After a blind test, some test drinkers couldn’t tell the difference, while other old Coke lovers thought there was a difference between the two formulas.
Despite this, Coca-Cola’s stock price has not changed and its sales have increased. Because the real moat of Coke has never been the syrup formula, but the filling and distribution network, the brand mentality, and cost discipline.
Coca-Cola is not the point, drinking Coca-Cola is.
The same goes for DeepSeek. Everything that can be open sourced is basically open sourced. If you want to download it, just download it yourself. Companies that dare to disclose their formulas have something more valuable than the formula in their hands.
03
What does Liang Wenfeng want?
After all, Liang Wenfeng is not a philanthropist. People always have to eat, not to mention that DeepSeek now has hundreds of people.
Liang Wenfeng once said clearly: “We clearly want to be commercialized. We still have to survive in the end.”
“We are still essentially a company. We are just considering what money we make, when we make money, how much money we make, and what we rely on to make money. We have choices.”
But when it comes to making money, what Liang Wenfeng wants more is a “win-win” with users. DeepSeek provides cheap or even free products, and users gradually become accustomed to DeepSeek’s model, expanding DeepSeek’s influence and scale.
The current situation is that the model war has passed, and the gap between each flagship model is narrowing. It is difficult to gain position by relying on parameters and lists alone. The focus of competition is shifting from “whose model is stronger” to “who controls the entrance, data flow and format.”
Just like Coca-Cola, it does not sell soda directly to you. It sells concentrated syrup to authorized filling plants around the world. The filling plants buy their own bottles, build channels, and market them. The Coca-Cola Company collects the syrup money and brand fees.
The reflowed data flow and format locking are the real value. Free period stores include Claude Code and Codex, which means placing refrigerators with Coca-Cola printed on them into rival stores. Open source is a syrup business, not a soda business.
Although models are also important, they are no longer the only battlefield.
The most intriguing thing about Search and Fetch is that it can help DeepSeek seize the entrance.
Every time the Agent needs to look up information, there must be a search tool to answer it. Whoever it calls by default occupies an entrance, and users can use this entrance for any model they want to use.
It’s not over yet. Every query, crawl, and reference selection of every connected Agent may become a training signal for DeepSeek, forming a flywheel of “entrance, data, model, and stronger entrance”.
The recipe directly makes the official processing logic into a public library, and each company uses it accordingly, and the behavior is naturally aligned.
For developers, this is a good thing. They have adjusted the Agent according to DeepSeek’s official behavior, and it is not easy to cause problems if it is replaced by a third-party service provider.
For DeepSeek, its prompt word format, tool calling format, and streaming output rules will become the de facto universal writing method as more and more third-party servers adopt it.
DeepSeek officials also stated: “In the future, the inference code and subsequent performance optimization of DeepSeek’s new model will also continue to be open sourced in DeepInfer.”
This means that the team that has built the infrastructure around DeepInfer can keep up with the new DeepSeek model as soon as it is released. This is also a kind of stickiness.
Related Reading
- I went to the Yushu Conference and Wang Xingxing made an explosive speech.2026-10-24
- Is it possible for the entire population to “quit” red fruits?2026-10-24
- Musk opens the door to a SpaceX-Tesla merger as shareholders weigh a deal2026-10-21
- Someone finally bought the original robot2026-10-09
- This year’s Nobel Prize, both AI and experts guessed wrong2026-10-09
