Cover Image | Generated by ChatGPT

DeepSeek, a company that is financially robust, is reportedly on the hunt for additional funding.

Foreign media outlets have reported that DeepSeek is seeking funds at a valuation exceeding $1 billion, with a minimum fundraising target of $300 million (approximately RMB 2.04 billion).

The announcement of this news sent ripples through the investment community.

DeepSeek has been known for its reluctance to accept external funding. Since its inception in 2023, it has relied heavily on its parent company, High-Flyer Quant, for financial backing. Even when rumors of funding surfaced in 2025, the company swiftly denied them.

However, considering the meteoric rise of companies like Zhipu and MiniMax, which have both surpassed a market capitalization of RMB 300 billion, and Yuezhi'anmian, which has seen its valuation quadruple, the potential returns for investors in DeepSeek could be substantial.

"This news seems too good to be true. How can the valuation be so modest?" An AI industry veteran shared with Pencil News, expressing skepticism. He believed that if DeepSeek were to raise funds, its valuation should at least be on par with that of publicly listed companies like Zhipu and MiniMax.

So, why is DeepSeek, a self-sufficient AI team that has never been constrained by capital, suddenly rumored to be seeking funding?

- 01 - Why the Sudden Need for Funding?

Talent Incentives

Several AI industry insiders revealed to Pencil News that if DeepSeek is indeed raising funds now, talent incentives could be the primary driver. The funds could be used to introduce market-based pricing for its stock option system, thereby reassuring employees about the value and liquidity of their options.

The AI large model battle is heavily reliant on capital: funds are needed for purchasing hardware and paying competitive salaries.

Pencil News previously reported that the current AI industry resembles professional sports leagues, with major companies spending lavishly to poach top talent. AI researchers are now akin to sports stars, ready to switch teams at any time, driven by the allure of AGI (Artificial General Intelligence) and financial incentives. (Click to read the article: The AI Talent War: Fresh Ph.D. Graduates Earn RMB 5 Million Annually).

"Just last week, rumors surfaced that former DeepSeek researcher Guo Daya had been 'poached' by ByteDance for an annual salary of RMB 100 million (a claim later denied by ByteDance). Earlier, Luo Fuli, who played a key role in developing the V2 model, joined Xiaomi. Of course, DeepSeek has also recruited talent from ByteDance and other major companies," an industry source said.

Mutual poaching among AI large model companies is a monthly occurrence, and the industry's competitive tactics are escalating. The talent war has evolved from a "salary competition" to a "capitalized competition."

In this fierce competition, DeepSeek finds itself at a relative disadvantage, not due to a lack of funds but because of its capital structure. For a long time, DeepSeek has not sought external funding, resulting in a lack of market pricing for employee stock options and illiquidity. In other words, paper wealth and actual wealth are not easily convertible.

DeepSeek's absolute compensation is competitive, but competitors offer clearer paths to financial gains, such as defined option pricing, repurchase mechanisms, or expectations of going public. Talent seeking financial certainty is prone to churn.

Therefore, investors generally believe that one of the key purposes of this funding round is to price and cash out employee stock options.

Changing Cost Structure

The second possibility for funding is that the cost structure of the AI industry has undergone a significant transformation.

Over the past two years, the industry's core competition has centered around training. Model parameter sizes have continued to expand, and the cost of training cutting-edge models has risen at an annual rate of approximately 3.5 times.

However, the situation is now changing, with model costs gradually shifting from training to inference. An AI service processing 10 million daily requests can incur monthly inference costs ranging from $300,000 to $3 million.

Over the entire lifecycle of an AI system, inference costs can be several times to over ten times higher than training costs. By 2026, inference will account for 55% of cloud-based AI spending. Deloitte estimates that inference will constitute about two-thirds of AI computational workloads by 2026.

"At the corporate level, the changes are more pronounced: average large model spending by companies has risen from $2.5 million in 2024 to $7 million in 2025. Some corporate executives even admit that the money spent in the past year is now exhausted in a week," an industry analyst noted.

The growth in AI costs no longer stems from "training a model once" but from continuously operating the model.

DeepSeek's past advantage lay in low training costs and high model efficiency. However, the current challenge is the expanding scale of inference, particularly the higher computational consumption brought about by Agent long-chain reasoning.

Advanced inference models, intelligent agent systems, and underlying infrastructure are all extremely costly. The scale of funding required to develop and operate the most advanced AI models is becoming exceptionally large, even beyond what a successful hedge fund can sustainably cover.

Therefore, fundraising has become a necessity.

Although opening up for funding will undoubtedly trigger excitement in the investment community, regarding the conditions for this funding round, an investor told Yicai Global, "Even if DeepSeek opens up for funding, it's not a game for most. Moreover, according to Liang Wenfeng's vision, the terms will undoubtedly be exceptionally stringent."

- 02 - The Second Half of AI Competition Begins

If we only look at the numbers, a $10 billion valuation for a large model company founded just three years ago is quite impressive.

However, within the current AI industry, this price appears conservative.

Among global leading companies, OpenAI is valued at approximately $850 billion, and Anthropic at around $380 billion. Among domestic leading large model companies, Zhipu and MiniMax have both surpassed a market capitalization of RMB 300 billion at their peaks, while Yuezhi'anmian's latest valuation has reached $18 billion.

In this context, DeepSeek's $10 billion valuation can only be considered "mid-to-low."

Moreover, DeepSeek is not an ordinary startup. In the past year, it has accomplished two remarkable feats: 1) directly challenging the global AI pricing system with low-cost models; and 2) matching or even approaching the performance of leading models.

While the amount DeepSeek is raising in this round is not substantial, the signal it sends is that the competitive logic in the AI industry has shifted from model capabilities to sustained operational capabilities.

This shift is already evident among leading companies in China and the United States.

Although OpenAI's latest valuation has reached $850 billion, the company expects to incur losses of approximately $14 billion in 2026. This indicates that even the industry leader's business model is still built on "continuous burning of money."

If OpenAI represents consumer-oriented products, then Anthropic represents another path—enterprise APIs.

Its data better reflects the future of AI product commercialization: annualized revenue of approximately $1 billion in 2024, growing to $9 billion by the end of 2025, and reaching $30 billion in 2026.

Anthropic's growth is primarily driven by enterprise API calls, developer subscriptions, and tools like Claude Code. It is expected to achieve positive cash flow by 2027.

This shows that AI companies are beginning to form revenue loops through "usage volume." However, conversely, the essence of this model remains that revenue comes from usage, and costs also come from usage. It is still essentially a "war of attrition."

Turning to the Chinese market, over the past year, leading companies like MiniMax, Zhipu, and Yuezhi'anmian have similarly focused on continuously expanding user usage volume and rapidly advancing commercialization. Especially with the emergence of more advanced agents like OpenClaw, the scale of token usage has further expanded, leading to a dramatic increase in annual recurring revenue and a clearer business closed loop.

Against this backdrop, DeepSeek has consistently adhered to a strategy of low-cost training, open-source models, and weak commercialization. This strategy was advantageous during the "training phase": the cost of training V3 was significantly lower than the approximately $100 million spent on training GPT-4.

However, the problem is that, as previously mentioned, the industry has evolved beyond just training. The current competition focuses on products, ecosystems, and commercialization. More importantly, fundraising itself is becoming an integral part of AI competition.

Starting at $300 million, the funds are not just for supporting research and development but also for building the foundations of computational power, teams (stock options), ecosystems, and commercialization. Without fundraising, DeepSeek can continue developing models but may not be able to support a grander AI vision, as a great future is not solely built on technology but also requires other pillars.

This article does not constitute any investment advice.