DeepSeek has finally broken its silence on fundraising. According to foreign media reports, DeepSeek is seeking at least US$300 million in its first external funding round, at a valuation of at least US$10 billion.
After DeepSeek went viral in early 2025, investors scrambled for a chance to meet its CEO Liang Wenfeng, but the company had long kept its funding window shut.
What made DeepSeek most unusual over the past year is precisely that it never acted like a typical AI company. Backed by High-Flyer, Liang was in no rush to raise money, nor to push the company onto the assembly line of valuation, commercialization, and capital exit. DeepSeek has positioned itself less as a commercial enterprise and more as an open-source research institute operating entirely independently of capital markets.
As soon as the fundraising rumor emerged, the market reacted swiftly. Yet a full year has passed, and both DeepSeek and the market have undergone profound changes that are impossible to ignore.
Over the past year, it would be unfair to say DeepSeek has fallen behind technologically. But compared with its peers, many have done things DeepSeek either hasn't done or systematized.
The benchmark in capital markets has also shifted. By itself, a valuation of at least US$10 billion still makes DeepSeek an expensive AI company.
But in today's Chinese AI landscape, that number is no longer stunning. Zhipu and MiniMax both reached market caps above HK$300 billion at their Hong Kong peaks; converted at certain market rates, DeepSeek's US$10 billion valuation is only a fraction of theirs, while Moonshot AI's latest valuation has reached US$18 billion.
If the US$300 million fundraising rumor is true, then DeepSeek has already crossed at least two major thresholds.
First, DeepSeek no longer treats fundraising as "something to fear." Servers, data, compute, commercialization, talent, and stock options—none of these can be long avoided by a pure research institute. Talent costs in particular have risen far above what they were a year ago.
DeepSeek once attracted people through technical idealism, open-source prestige, and Liang's personal appeal. But when Guo Daya lands a huge annual package at ByteDance, it becomes critically important for a DeepSeek employee to know whether they can share in the company's growth through stock options.
To some extent, options also relieve Liang of pressure: employees take what they deserve, and he no longer has to worry excessively.
Second, DeepSeek is returning to the normal development path of a commercial company. The research idealism can continue, but the company ultimately needs a governance structure, a valuation system, incentive compensation, commercial revenue, and a long-term budget. In the past, DeepSeek was expected to astonish the world with every release. What it needs to do now is simply become a normal company.
DeepSeek's underlying model capabilities remain strong. Its contributions to model algorithms, engineering efficiency, open-source approaches, and inference cost reduction are still among the most important technical events in Chinese AI over the past year. R1 proved that a small team can also build a world-class model with fewer resources and a more open path.
Yet in reality, today's AI competition is no longer about single-point model capability. DeepSeek's greatest strength is the model itself, while its peers have done much more beyond the model.
The most obvious area is product entry points.
DeepSeek was once the fastest-growing AI app in China. But by the second half of 2025, Doubao had overtaken DeepSeek in monthly active users. According to QuestMobile, in August 2025, Doubao became China's No. 1 native AI app with about 157 million MAUs, pushing DeepSeek to second place.
ByteDance has disclosed that as of March 2026, Doubao's cross-channel MAUs exceeded 331 million—more than the sum of the second through fifth place products.
This shows one thing: a viral model can generate a huge first wave of traffic, but retaining users in the long run depends on product, scenario, operations, and ecosystem entry points.
That's where ByteDance has the advantage. Behind Doubao are Douyin, CapCut, Volcano Engine, and a rich content ecosystem; Jimeng captures creative needs, and Seedance 2.0 has pushed video generation capabilities into the market spotlight.
While DeepSeek has prestige in the model community, at the consumer-product level, it has not achieved the sustained distribution and high-frequency usage that Doubao enjoys.
Multimodality is a similar issue.
DeepSeek has built Janus-Pro and DeepSeek-OCR, but it has yet to form a stable, complete, and strong multimodal product system. Today's AI competition increasingly demands a unified experience across text, image, voice, video, tools, and agents. OpenAI, Google, and Anthropic are moving in this direction, as are domestic players like ByteDance, Alibaba, and Tencent.
The fact that Alibaba and Tencent have begun betting on world models is a telling signal.
Alibaba released Happy Oyster, emphasizing an interactive, performable, and explorable AI digital world. Tencent released and open-sourced Hunyuan3D-2.0, a 3D world model that generates and simulates 3D environments from text, images, and video inputs.
These may not yet translate into mature commercial revenue, but they show that the big players are pushing AI capabilities beyond chat boxes and code editors into more complex spaces—video, gaming, and content production.
ByteDance, meanwhile, is doubling down on video generation.
After Seedance 2.0, market attention has shifted from generating a video to multi-shot composition, audio-video synchronization, narrative pacing, character motion, and production workflows. Once these capabilities connect with CapCut, Douyin, e-commerce ads, and film/TV production, they will form a system that DeepSeek will find very hard to replicate.
DeepSeek still holds an edge in model efficiency, but in app entry points, multimodality, video generation, world models, agents, AI coding, enterprise services, and ecosystem distribution, it has fallen behind its peers. For a company remembered as a technological miracle, this gap seems especially stark.
Liang's move to raise funds should not be understood simply as a lack of money. More accurately, he has realized that relying solely on foundation models is no longer enough to sustain competition in the next phase. DeepSeek needs more talent, servers, and a more complete commercial ecosystem.
DeepSeek's most urgent problem now is talent. From the second half of 2025 onward, several core members have been seen leaving.
Wang Bingxuan, who participated in early LLM training, went to Tencent. Wei Haoran, lead author of DeepSeek-OCR, left. Guo Daya, lead author of DeepSeek-R1, joined ByteDance. Ruan Chong, who joined during the High-Flyer era and worked on Janus-Pro and other multimodal projects, announced in January 2026 that he was joining DeepRoute.ai. Luo Fuli has also joined Xiaomi to lead related AI initiatives.
This is the inevitable market outcome after DeepSeek's explosive rise. Its core researchers have become targets for every major tech company and AI startup.
In the past, DeepSeek was idealistic. Its talent appeal came from technical challenges, open-source prestige, research freedom, and Liang himself. But today's AI talent pricing is completely different. Top industry researchers receiving annual packages of nearly 100 million yuan have become common, and DeepSeek cannot match those offers. That is why stock options have become increasingly important for Liang.
DeepSeek's past refusal to raise funds, set external valuations, or operate under market pressure reduced outside distractions in the short term. But in the long run, it made it hard to price employee options clearly. Without an external valuation and a clear incentive system, DeepSeek cannot convince core talent that they will surely share in the company's growth.
That's why the US$300 million funding round matters. As mentioned earlier, if the valuation is too high, Liang will also bear greater growth pressure. Therefore, this round is likely not just about raising money, but also about pricing the company, the team, and the future incentive system.
DeepSeek's early charm came precisely from not being pushed around by capital markets. For a founder with strong technical ideals, fundraising means new shareholders, new constraints, new communication costs, and an end to the company running entirely at the research team's pace.
But that is the price any normal company must pay.
Another hurdle for DeepSeek to "return to normalcy" is service stability. At the end of March 2026, DeepSeek suffered an 11-hour outage that even trended on social media. No matter how powerful the model, if it serves massive numbers of users and developers, it must pass commercial reliability tests. The simplest, most direct way to fix unstable servers is to spend money on more servers, more compute, redundancy, and a stronger cloud and operations system.
Data and training costs also matter. Early DeepSeek training costs were low because the team pushed model architecture, engineering efficiency, and methods like distillation to their limits. But by the V4 stage, a single training run may already cost more than US$500 million.
At the same time, after companies like Anthropic restricted distillation paths, if DeepSeek continues to compete in the top tier in the second half of the year, it will need to purchase more high-quality datasets, and training costs will rise significantly. The "low-cost miracle" that DeepSeek was famous for will not automatically carry over to future models.
Low-cost training demonstrated the team's capability, but the next-generation foundation model will still face a basic industrial reality: scaling laws.
Stronger models generally require more high-quality data, larger-scale compute, more complex post-training systems, more intensive evaluation, and stricter safety alignment. Foundation models are expensive and compute-hungry. The closer to the top tier, the higher the marginal cost.
The fourth hurdle is commercialization. In the past, DeepSeek's logic was clear: open-source models generate influence, API fees capture developer demand, and strong model capabilities drive distribution and usage.
But this logic is no longer sufficient for DeepSeek today. Liang now wants to build a complete commercial system for DeepSeek, including subscriptions and tiered API pricing.
However, if DeepSeek wants to build a commercial ecosystem from scratch, it must bear the cost of infrastructure. Relying solely on open-source buzz and basic API revenue cannot sustain a foundation-model company with global ambitions. Commercialization is not a betrayal of the technical path; it is the necessary foundation that the technical path must build for long-term competition.
Therefore, DeepSeek's fundraising is not an isolated event, but the inevitable result of servers, data, compute, talent, options, and the commercial ecosystem all converging. For a company like DeepSeek, built on model capabilities, these things matter more than the valuation number itself.
For the same reason, DeepSeek's return to the "normal" path of fundraising, valuation, and commercialization is not a step backward. Fundraising does not necessarily mean Liang is being captured by capital; it can also mean that a company is finally able to sustain long-term competition. For users and the industry, it is time to say goodbye to the "DeepSeek moment." The company has finally stepped down from a story to be admired, and onto the ground where a company should stand.
(Source: Wujicaijing)
Related News:
Deepline | 'Contradictions' of Unitree: Tech giant, marketing silencer
Comment