Liang Wenfeng has figured it out.

Finally, there is news about DeepSeek's financing.

According to foreign media reports, DeepSeek is seeking at least $300 million in its first round of external financing, with a valuation of at least $10 billion.

Zimubang reached out to DeepSeek for verification, but received no response.

After DeepSeek became extremely popular in early 2025, investors were eager to meet Liang Wenfeng, but DeepSeek did not open a financing window for a long time.

What made DeepSeek special in the past year was that it was not a typical AI company.

Backed by Magic Square Quantitative, Liang Wenfeng was in no hurry to raise funds or push the company onto the assembly line of valuation, commercialization, and capital exit.

DeepSeek positions itself more like an open - source research institution that operates independently of the capital market rather than a commercial company.

After the financing rumor emerged, the market reacted immediately. Investors may have bombarded Liang Wenfeng with phone calls, and some have even prepared to book flights just to meet Liang Wenfeng and get the opportunity to invest in DeepSeek.

However, a year has passed, and there have been significant and far - reaching changes in both DeepSeek itself and the market.

In the past year, it can't be said that DeepSeek has fallen behind in technology. But compared with its peers, many of them have done things that DeepSeek hasn't done or hasn't established a system for.

ByteDance has Doubao, Jimeng, and the Seedance video - generation model; Alibaba and Tencent have started to promote their world models; Tencent Yuanbao and Alibaba Qianwen are being integrated into their respective ecosystems; Zhipu and MiniMax have listed on the Hong Kong Stock Exchange and completed a new valuation re - assessment in the secondary market.

The reference system in the capital market has also changed.

Judging from DeepSeek's valuation of at least $10 billion, it is still an expensive AI company.

But in today's Chinese AI context, this figure is no longer astonishing. Zhipu and MiniMax once had a market value of over HK$300 billion at their peak on the Hong Kong Stock Exchange. According to some market standards, DeepSeek's $1 - billion valuation is just a fraction of theirs, and the latest valuation of "rising stars" like Yuezhianmian has reached $18 billion.

If the rumor of a $300 - million financing is true, DeepSeek has at least overcome two hurdles.

First, DeepSeek is no longer afraid of financing.

Servers, data, computing power, commercialization, talent, and stock options are all things that a pure research institution cannot avoid in the long run.

Especially the cost of talent has increased significantly compared to a year ago.

In the past, DeepSeek could attract a group of people with its technological ideals, open - source reputation, and Liang Wenfeng's personal charisma. But when Guo Daya received an annual package of nearly 100 million yuan at ByteDance, it became extremely important for DeepSeek employees to know whether they could share the company's development dividends through stock options.

To some extent, stock options also relieve Liang Wenfeng of pressure. By allowing employees to take their due share, Liang Wenfeng doesn't have to worry too much.

Second, DeepSeek is returning to the normal development path of a commercial company.

A company should operate as a company. Research ideals can still exist, but a company ultimately needs a governance structure, a valuation system, salary incentives, commercial revenue, and long - term budgets.

In the past, DeepSeek was expected to make a world - shaking impact with each release. Now, what it needs to do is to become a normal company.

A

DeepSeek still has strong underlying model capabilities.

Its contributions in model algorithms, engineering efficiency, open - source routes, and reducing inference costs are still among the most important technological events in the Chinese AI field in the past year. R1 proved that a small team can create a world - class model with fewer resources and a more open approach.

However, in fact, today's AI competition is no longer just about single - point model capabilities.

DeepSeek is strongest in the model itself, while its peers have done more outside the model.

The most obvious is the product entry.

DeepSeek was once the fastest - growing domestic AI app, but in the second half of 2025, Doubao surpassed DeepSeek in terms of monthly active users. A report from QuestMobile shows that in August 2025, Doubao topped the list of monthly active users of Chinese native AI apps with about 157 million monthly active users, and DeepSeek fell to second place.

ByteDance revealed that as of now, Doubao's monthly active users across all channels in March 2026 have exceeded 331 million, which is the sum of the monthly active users of the second - to - fifth - ranked products.

This shows that a popular model can bring a huge first - wave of traffic, but the long - term user scale needs to be retained through products, scenarios, operations, and ecological entrances.

ByteDance's advantage lies here. Behind Doubao are Douyin, Jianying, Volcengine, and a content ecosystem. Jimeng caters to creative needs, and Seedance 2.0 has brought video - generation capabilities into the spotlight.

Although DeepSeek has a good reputation in the model community, it has not developed the ability for continuous distribution and high - frequency use like Doubao at the mass - product level.

The same problem exists in multi - modality.

DeepSeek has developed Janus - Pro and DeepSeek - OCR, but it has not formed a stable, complete, and powerful multi - modality product system. Today's AI competition increasingly emphasizes a unified experience of text, images, voice, video, tools, and agents. OpenAI, Google, and Anthropic are moving in this direction, as are domestic companies like ByteDance, Alibaba, and Tencent.

Alibaba and Tencent's decision to bet on world models is a typical signal.

Alibaba released Happy Oyster, emphasizing an interactive, performable, and explorable AI digital world; Tencent released and open - sourced the Hunyuan 3D World Model 2.0, which can generate and simulate a 3D world based on text, image, and video inputs.

These may not immediately turn into mature commercial revenue, but they represent that large companies are pushing AI capabilities from chat boxes and code boxes to more complex spaces, video, games, and content - production scenarios.

ByteDance is continuously increasing its investment in video generation.

After the release of Seedance 2.0, the market's focus is no longer just on "whether a video can be generated," but on multi - shot, audio - video synchronization, narrative rhythm, character movements, and the production process. Once these capabilities are connected with Jianying, Douyin, e - commerce advertising, and film and television production, they will form a product closed - loop that DeepSeek can hardly replicate at present.

Agents and AI programming are also DeepSeek's weak points.

DeepSeek does have tool - calling and agent capabilities, but it has not established a clear productivity entry in the minds of developers like Claude, GPT, Kimi, MiniMax, Tencent, and Alibaba.

AI programming is becoming the clearest commercialization scenario for large models, and developers will choose those with good effects and high stability. Take OpenClaw as an example. Hardly anyone would use DeepSeek for lobster farming because the results are really disappointing.

This is the real situation that DeepSeek is facing.

It is not weak, but its strength is not comprehensive enough.

It still has an advantage in model efficiency, but it has fallen behind its peers in terms of app entry, multi - modality, video generation, world models, agents, AI programming, enterprise services, and ecological distribution. For a company that was once remembered for its "technological miracle," this gap is particularly glaring.

Liang Wenfeng's decision to raise funds this time cannot be simply understood as a lack of money.

More accurately, Liang Wenfeng realizes that relying solely on the basic model is no longer sufficient to support the next - stage competition.

DeepSeek needs more, such as talent, servers, and a more complete business ecosystem.

B

The most urgent problem for DeepSeek now is talent.

Since the second half of 2025, several core members have left one after another.

Wang Bingxuan, who participated in the early large - model training, went to Tencent; Wei Haoran, the core author of DeepSeek - OCR, left; Guo Daya, the core author of DeepSeek - R1, went to ByteDance. Ruan Chong, who joined from the Magic Square era and participated in multi - modality work such as Janus - Pro, also officially announced his joining of Yuanrong Qixing in January 2026. Luo Fuli has also joined Xiaomi to be in charge of relevant AI business.

This is an inevitable market result after DeepSeek became popular. Its core researchers have become the targets of all large companies and AI startups.

In the past, DeepSeek was an idealistic team.

Its attractiveness to talent came from technological challenges, open - source reputation, research freedom, and Liang Wenfeng himself. But today, the talent pricing in the AI industry is completely different. It has become normal for top - tier industry researchers to receive an annual package of nearly 100 million yuan. However, DeepSeek cannot offer such high salaries.

So, stock options are becoming increasingly important for Liang Wenfeng.

In the past, DeepSeek did not raise funds, did not set an external price, and maintained independent operation. In the short term, this could reduce external interference.

But in the long run, it is difficult to clearly price the stock options in employees' hands.

In other companies, with financing rounds, listing expectations, and secondary - market prices, employees at least know the approximate value of what they hold.

Without an external valuation and a clear incentive system, it is difficult for DeepSeek to convince its core talent that they can share the company's growth dividends.

Jiang Yi, the founding partner of Hengye Capital, said that the company needs a valuation because large companies offer not only cash but also high - valuation stock options to talent. Without a clear valuation, it is difficult for DeepSeek to let employees accurately judge what their stock options mean for the future.

This is where the $300 - million financing is truly important.

It may not be for large - scale spending immediately. According to Jiang Yi, a valuation below $10 billion is relatively low for Liang Wenfeng.

As mentioned at the beginning of the article, if the valuation is too high, Liang Wenfeng will also face greater growth pressure. Therefore, this round of financing is likely not just about getting money, but also about pricing the company, the team, and the future incentive system.

The early charm of DeepSeek came from the fact that it was not driven by the capital market. Jiang Yi mentioned that Liang Wenfeng's personality is such that he is quite averse to investors interfering in the company's decision - making.

For a founder with strong technological ideals, financing means new shareholders, new constraints, new communication costs, and it also means that the company can no longer operate entirely according to the rhythm of the research team.

But this is also a cost that a normal company must face.

A company should operate as a company, and employees should get their due.

If DeepSeek wants to continue to retain the top - tier talent, it must let the team members see a realizable future. It can't just ask them to believe that the model will become stronger; it also needs to let them believe that they can share the development dividends when the company becomes more valuable.

C

For DeepSeek to return to being a normal company, talent is just the first hurdle.

The second hurdle is service stability.

At the end of March 2026, DeepSeek experienced an 11 - hour outage, which even made it onto the hot search. No matter how strong the model capabilities are, as long as it provides services to a large number of users and developers, it must pass the commercialization test. The simplest and most direct way to solve the problem of unstable servers is to spend money to buy more servers, purchase more computing power and redundant resources, and build a stronger cloud - service and operation - maintenance system.

Of course, engineering optimization, scheduling strategies, model compression, and caching mechanisms are all important.

But in the face of peak traffic, many problems will ultimately come down to capital investment. Users won't tolerate long - term unavailability, excessive queuing, or API fluctuations just because a company has a good training story. Developers won't entrust their core business to an unstable interface just because a model once shocked the world.

A good model is just the starting point, and stable service is the norm.

The third hurdle is data and training costs.

Jiang Yi mentioned that DeepSeek's early training costs were relatively low because the team was extremely proficient in model structure, engineering efficiency, and distillation methods.

But in the V4 stage, the cost of a single - round training may have exceeded $500 million.

Meanwhile, after companies like Anthropic blocked the distillation path, if DeepSeek wants to continue to compete in the first - tier in the second half of the year, it will need to purchase more high - quality datasets, and the training cost will increase significantly.

The "low - cost miracle" that DeepSeek was most talked about in the past will not naturally continue in future models.

Low - cost training proves the team's ability, but the next - generation basic models still have to follow the simple industrial law, the Scaling Law.

Stronger models usually require more high - quality data, larger - scale computing power, more complex post - training systems, more intensive evaluations, and security alignment. Basic models are expensive and consume a lot of computing power. The closer to the first - tier, the higher the marginal cost.

This doesn't even include the cost of adapting to domestic computing power and compliance.

Jiang Yi mentioned that Magic Square Quantitative had good revenues last year and has maintained a good level this year. However, there have been many regulatory and compliance requirements regarding quantitative trading in China, and the relevant adjustments themselves require investment.

On the other hand, if DeepSeek wants to adapt to domestic computing power, it also needs an engineering team, a testing system, and long - term running - in. This investment is not a one - time expense but a continuous consumption.

The fourth hurdle is commercialization.

In the past, DeepSeek's logic was clear: open - source models to create influence, charge for APIs to meet developers' needs, and excellent model capabilities to drive dissemination and calls.

But this logic is no longer sufficient for DeepSeek today.

Jiang Yi said that Liang Wenfeng now wants to find a complete business system for DeepSeek, such as a subscription system and different levels of API charges.

An ecosystem doesn't grow naturally just by open - sourcing the model. It requires stable APIs, developer tools, enterprise services, documentation systems, billing systems, security compliance, customer success, and channel cooperation.

If DeepSeek wants to build a business ecosystem from scratch, it must bear the cost of infrastructure construction.

Relying solely on open - source popularity and basic API revenue is difficult to support a basic - model company with global influence to move forward. Commercialization is not a betrayal of the technical route but a necessary foundation for the technical route to enter long - term competition.

Therefore, DeepSeek's financing is not an isolated event.

<