Talk about the 8 AI companies in Liang Wenfeng's vision

The 34,000-word full transcript of Liang Wenfeng's internal investor exchange meeting went viral online. For certain reasons, the full text is no longer accessible, but the more mysterious it becomes, the more the industry grows interested in Liang Wenfeng and his company DeepSeek, including its highly anticipated upcoming V4 model. This stands in stark contrast to the practices of some peers who go to great lengths to hype their own large models, echoing a line from Liang Wenfeng:

"There's a strange paradox: the things you crave the most often elude you, while the matters you don't fixate on too hard tend to come surprisingly easily."

What intrigues me are the several AI companies Liang Wenfeng mentioned. During this 4-hour, tens-of-thousands-of-word long conversation, Liang Wenfeng referenced a total of 8 peer AI companies (excluding chip firms, even though NVIDIA was the most frequently mentioned name), namely Zhipu, Moonshot AI, Alibaba, Tencent, ByteDance, and the three Silicon Valley giants.

1. Zhipu: Forced to Open Source, Not by Original Intention

Liang Wenfeng mentioned Zhipu multiple times, listing it as one of the competitors DeepSeek is willing to assist. Meanwhile, MiniMax, the other company often grouped with Zhipu as the "two Hong Kong-listed large model stars", did not even enter his field of vision — it was not referenced a single time throughout the tens of thousands of words of conversation.

"For example, Zhipu also does open source, but their open source effort is different from ours. There's a sense that Zhipu went open source reluctantly, that it wasn't their original intention. But for us, this is exactly what we intended to do from the start."

"We also hope these AI technologies can be applied to all kinds of production environments, to improve social productivity and help all industries boost their operational efficiency. We have very strong motivation to make this happen."

For some companies, "open sourcing" is a choice of business strategy, especially for those that were forced to shift from an originally closed-source model to open source. There is also a common practice of "open sourcing the previous generation model while keeping the latest one for internal use". But DeepSeek uses exactly the same model internally as the one it open sources:

"We would never release a slightly inferior model as open source, then use a better model for our own deployment. The open-sourced model and the one we deploy ourselves are completely identical."

Liang Wenfeng also added,

"When it comes to open source, we had everything very clearly thought out from the very beginning. First, it's about our vision; second, we believe open source is beneficial for making AI a commercially successful business."

"When we cut our prices back then, many people in the company group cheered, everyone was very happy. Because this is the very purpose we put so much effort and care into perfecting the model: to make AI services extremely affordable, with excellent performance, so that everyone can make full use of them. That's what makes us happy. This is our motivation, our vision, and the shared consensus that unites our entire company to work toward this goal."

Reading these words, I can truly feel the genuine open source spirit coming from Liang Wenfeng and the entire DeepSeek team.

2. Alibaba: Several Times Higher Cost, We Are Willing to Lend a Hand

Alibaba is a Tier 1 player in China's AI value chain, especially in the open-source model and computing power ecosystem. Liang Wenfeng referenced Alibaba in topics including cost comparison, competition, and collaboration.

"But when it comes to cost recovery within ten months, we can achieve that on our own, while other companies cannot. For players like Alibaba or Tencent, who don't have the same level of optimization capabilities as us, their costs should be several times higher than ours. There's still a lot of optimization work behind our results."

"We are very willing to assist and help anyone, even our competitors — including Alibaba, Zhipu, and Moonshot AI — to do better."

Large tech giants have abundant capital and relatively more sufficient resources, especially computing power resources. For example, Alibaba positions itself as an "AI factory" responsible for Token production and integrating and leasing computing power resources, placing companies like Zhipu, DeepSeek, and Moonshot AI in its downstream ecosystem. That's why Alibaba has invested in multiple leading large model vendors, some of which were even acquired using computing power resources as equity contribution.

DeepSeek has relatively more limited resources, yet it insists on taking the open source and inclusive path, having refined its engineering optimization capabilities to the extreme. Companies like Alibaba can currently tolerate several times higher costs, using subsidies to boost C-end applications and B/D-end model and cloud service adoption, but they will definitely become more cost-conscious in the future. Judging from the financial reports of major tech giants like Google, Meta, and Oracle, the continuous rapid growth in capital expenditure is leading to increasingly tight free cash flow, which is even turning negative. When even wealthy landlords no longer have surplus grain, no matter how rich a tech giant is, it will have to "cut unnecessary costs". Therefore, there is still considerable room for cooperation between DeepSeek and companies like Alibaba in the future.

3. Tencent: No Plans to Build a Super App, No Conflicts of Interest

Liang Wenfeng mentioned Tencent in three scenarios: cost comparison, competitive positioning, and traffic advantage. Besides pointing out that Tencent's costs are several times higher than his company's, just like Alibaba's, he also explicitly stated that DeepSeek has no intention of becoming the next Tencent:

"We are not competing for users, nor are we chasing quick profits. Instead, we are working very hard to find ways to better serve our users. We never had the idea that we want to build the next super app, then compete with someone else, or become the next ByteDance or the next Tencent. We could have chosen that path, but we didn't. In my understanding, this is part of exercising restraint — you don't have to grab every possible profit opportunity."

In contrast, products like ChatGPT, Doubao, and Tongyi are clearly sprinting down the path toward becoming super apps: while continuously driving up their daily active user counts, they are integrating services like ride-hailing, ticket booking, food delivery, and even connecting e-commerce, payment, and wealth management functions, essentially porting WeChat's "nine-square-grid service" model into AI super applications.

By trying to intercept user traffic through AI super applications, companies like Google, Meta, Amazon, and Alibaba are in turn forced to strengthen their consumer-facing AI application offerings, sparking an arms race. DeepSeek has exercised extreme restraint in this area: it focuses on perfecting large language models for AGI to the utmost, and did not touch multimodal capabilities until it added image recognition support in April. It has taken even fewer actions to cooperate with ecosystem partners to enable AI to "handle tasks", "work", or "place orders" on behalf of users.

Liang Wenfeng also summarized his company's cooperation with Tencent:

"When we launched our To B business last year, many people asked: since we are open sourcing our models, wouldn't that create a conflict between Tencent's C-end business and our own C-end business? Since we don't have the traffic advantage, while Tencent has massive user traffic, if Tencent deploys our open-sourced model, it will take over all C-end users and steal our C-end user base. But that won't actually happen, there are many reasons behind that."

When users use Tencent's AI search and other services, they can choose to use either the DeepSeek model or the Hy3 model, but a portion of users will still be accustomed to using DeepSeek's own official channels (website/App).

4. ByteDance: No Competition Against "Doubao", Is Not Focusing on the C-End the Right Choice?

Liang Wenfeng mentioned ByteDance multiple times in topics about competitive strategies and closed-source models. Besides stating that he has no intention of building the next ByteDance, he also indirectly shared his views on ByteDance's Doubao AI assistant.

Why is DeepSeek standing by and watching Doubao catch up to become a mass-market national AI application? Liang Wenfeng's explanation is:

"I shouldn't grab every tiny sesame seed out there. Sure, this 'sesame seed' might be pretty big, but I think compared to the upcoming AGI era, none of these current opportunities are that significant. Looking back now, our decision not to focus on the C-end business last year was probably the right call. Because we can see there are much bigger watermelons waiting for us ahead, while what's in front of us now are just small sesame seeds. If I had raised a huge amount of capital last year and scaled up the C-end business massively, what benefits would that have brought? We wouldn't have gained anything substantial. These are my real thoughts, because I believe the opportunities in the AGI era ahead will be enormous."

The biggest problem currently faced by native AI super applications like Doubao, ChatGPT, and Tongyi is commercial monetization: their user base has grown, their daily active traffic is considerable, but they have no idea how to recover their costs. The more users use their services, the more Tokens are consumed, and the bigger the headache for AI platforms: the general public's willingness to pay is far weaker than programmers' willingness to pay for productivity tools like coding assistants. Claude can make its business model work and operate profitably, but even Doubao's launch of a professional paid version was met with complaints from users. It's also very difficult to insert ads into AI-generated outputs; although OpenAI is experimenting with this, it's unlikely to become a major revenue source. When AI super applications cannot recover their costs, they still have to keep spending heavily on GPU resources to ensure user experience does not degrade — that's why everyone in this space is facing huge operational pressure right now.

Liang Wenfeng also directly criticized the closed-source model adopted by companies like ByteDance:

"Let me elaborate more on open source, because we've received so many questions about this topic. First of all, we will continue to open source our models, and even our most powerful model will likely be open sourced too. Because I can't see the benefits of closed sourcing, there are no inevitable advantages to it. ByteDance keeps its model closed source — what benefits does that bring them? I can't see any real advantages."

5. Moonshot AI (Kimi): Both Open Source Players, But With Different Mindsets

Liang Wenfeng listed Moonshot AI as one of the competitors DeepSeek is willing to assist. Recently, Moonshot AI released its K3 model, which has become extremely popular and drawn widespread attention from the global AI community. Some media even claimed that this model has shaken the trillion-dollar market valuation of US tech stocks, and "credited" it for the sharp plunge in the share prices of Zhipu and MiniMax.

Both Kimi and DeepSeek are players that open their model weights to the public, and they do not strictly follow the OSI definition of "fully open source" (their training data, training code, and RL processes are not fully disclosed). But the two companies have different levels of openness. Kimi opens its model weights to the public, while keeping some of its core technologies as exclusive trade secrets; DeepSeek not only opens its model weights, but also publishes a large number of academic papers and technical details to make it as easy as possible for the entire industry to replicate its results. That's why every time DeepSeek releases a new model, the capabilities of many other models across the AI industry can be upgraded simultaneously.

During this exchange, Liang Wenfeng reaffirmed DeepSeek's open source philosophy, which sets enabling industry "replication" as its core goal:

"We hope you can replicate our results; if you can't, just reach out and we will tell you exactly how to do it. This is an integral part of our open source commitment, it doesn't change just because you are a competitor. Of course, if you are our partner, we will go even further to help you."

6. OpenAI/Anthropic/Google: The Three Giants Will Overtake Each Other in Turn

Liang Wenfeng repeatedly mentioned the three Silicon Valley AI giants: OpenAI, Anthropic, and Google.

Regarding the competition among these three giants, he proposed a concept of "alternating overtaking":

"Anthropic has now surpassed OpenAI — but is this a long-term situation? I don't think so, this is just a temporary extreme state. In the future, OpenAI and Google will most likely overtake each other in turn, that's the pattern we will see. Right now, Anthropic doesn't have that big of an advantage in Code Agent products, it's not completely dominating OpenAI."

"Anthropic had a first-mover advantage, but that advantage will disappear very soon — it's not an advantage that it can hold onto for a long time. All three of these companies are very strong. Among them, Anthropic is the most efficient one, it has spent the least amount of capital, burned the least amount of money."

Chinese AI players are also continuously narrowing the gap with Silicon Valley through relentless "catch-up and competition", but objective gaps still exist, for example in high-quality data labeling. "I think the bottleneck of high-quality data labeling will take time to resolve. Because OpenAI, overseas players, and Anthropic all started much earlier, with more capital and more GPU resources." But when it comes to talent, Liang Wenfeng is far more optimistic than most people expect:

"The shortage of AI talent is a temporary phase, and we have already seen this problem being greatly alleviated. Because there is no real shortage of AI talent out there. Every company can train their own AI professionals very quickly, talent cultivation is a fast process. So across the entire AI industry, no matter it's the ecosystem, model companies, or any other segment, there is no talent shortage. Talent shortage is definitely a short-term phenomenon. In history, there has never been a long-term shortage of a specific type of professional. I remember more than a decade ago, people were saying there was a severe shortage of pilots, and pilot training took a very long cycle — but that problem was resolved very quickly. So no one needs to worry about the lack of AI talent."

We Don't Need That Many Companies Building AI Foundation Models, the Industry Will Definitely Consolidate in the Future

Amid the grand "100 Models Battle" boom, many domestic large AI model companies were not mentioned by Liang Wenfeng. But he shared a thought-provoking statement:

"Right now there are too many companies in China building foundation models, way too many. In the US there are probably only three major players, but in China the number of foundation model developers is excessive. Eventually, we definitely won't need that many companies building foundation models, the industry will definitely consolidate. Right now resources are very scattered, which is a form of waste to some extent. Every company is doing the exact same work, which means each of them ends up with far fewer resources than they would have in a more consolidated market."

In fact, the knockout round for AI foundation large models has already begun. There are very few core players left at the table, beyond the several companies within Liang Wenfeng's field of vision.

This article is from the WeChat official account "Luo Chao Pro", written by Luo Chao, and published by 36Kr with authorization.