Liang Wenfeng is still too conservative.

DeepSeek, is out there looking for money.

According to The Information, DeepSeek plans to raise at least $300 million at a valuation of no less than $10 billion, marking its first time opening up to external capital since its founding.

Over the past year, DeepSeek has deliberately maintained a rare stance in the AI industry: no fundraising, no commercialization, no roadshows. The profits from its parent company, Fantasia Quantitative, have continuously fueled it, allowing it to stay outside the logic of capital. This is scarce in the AI circle and has thus accumulated a considerable reputation.

After all, DeepSeek is different from Kimi, Doubao, and Qianwen—its position in public perception and the weight it carries in the larger narrative go beyond mere business topics. It’s more like a proof of China’s AI capabilities, a complex totem, rather than just a company fighting in the market.

For this reason, every event involving DeepSeek carries a weight beyond the company itself.

Why is DeepSeek suddenly starting to raise funds?

The direct trigger for this round of funding is talent outflow.

DeepSeek-R1 exploded in early 2025, and its core researchers immediately became hot targets in the industry, with ByteDance, Xiaomi, Tencent vying to poach them, and several key researchers leaving one after another.

From the overall perspective of China’s AI industry, this kind of talent competition is a normal market signal—talent flowing within the domestic ecosystem is a sign of industry maturity.

Aren’t we always told to respect talent and give top performers higher待遇?

Well, now, that待遇 is finally here.

However, for DeepSeek itself, “being poached” is a big headache: it’s not that things can’t go on without those people, but it’s a matter of principle.

Previously, DeepSeek was a small project cultivated by Fantasia Quantitative, not needing to consider making money, able to close the door and entertain itself, focusing on technological ideals.

The advantage of such a project is obvious—geeks love this environment. Which technical expert can refuse such a free R&D environment? Last year’s brilliance of DeepSeek was a result of this atmosphere.

But this kind of closed, self-entertaining project also has obvious drawbacks—after detaching from external distractions, while gaining freedom in R&D, it loses a reasonable judgment of its own value.

Specifically: due to DeepSeek’s low profile before, it has long lacked a reliable valuation, and because of the absence of a reasonable valuation, it cannot use options to incentivize its technical experts—allowing them to buy company shares at a predetermined price in the future. For example, if the company’s current valuation is $10 billion, you get options that let you buy shares at a price corresponding to that valuation. If three years later, the company’s valuation rises to $50 billion, exercising the options now yields a profit equal to the difference.

This is a common incentive method for companies with a clear valuation and market cap.

Of course, DeepSeek will not neglect its R&D staff. According to previous reports, the average salary of the DeepSeek team is about $10,000 per month, and earning over a million annually is not a problem—definitely a high salary.

But the problem is, the compensation offered in the external market can be tens of millions or even over a hundred million.

To some extent, this flaw has led to DeepSeek being targeted by major tech giants trying to poach its talent.

That’s why DeepSeek is seeking funding—it must first obtain a certain valuation in the market to give its employees options as incentives, otherwise, options are just illusions.

Of course, there’s a deeper, technical pressure:

R1 and V3 demonstrated impressive capabilities in early 2025, but now it’s April 2026.

At that time, DeepSeek’s competitor was GPT, but now, just over a year later, the competitors are Claude, Gemini, Grok. Unfortunately, the performance of DeepSeek R1 and V3 has clearly been surpassed by Claude and Gemini.

AI iteration is rapid, and external judgments of a company’s current model generation position quietly influence top researchers’ decisions to stay or leave. Talent’s willingness to stay depends partly on whether they believe there are enough “tech battles” to fight here.

Therefore, beyond monetary incentives, DeepSeek must also produce products that surpass Claude and Gemini to prove it still has enough “tech battles” to fight—this funding is anchored to a $10 billion market valuation. The core purpose of this figure is twofold: to make employee options real and to demonstrate its technological progress, allowing DeepSeek’s R&D efforts to continue.

What does a $10 billion valuation imply?

So, is $10 billion expensive?

It depends.

One thing we must understand is: over the past year, DeepSeek has been a very laid-back project.

So far, no one has found any systematic revenue from DeepSeek; its API pricing is very low, its B2B product line is still undeveloped, almost like a side hobby for some big shots—completely indifferent to whether it makes money or not, just for fun.

Looking at Anthropic makes DeepSeek more interesting. This company also prioritizes technological research, has launched the epoch-making AI product Claude, but its valuation has already exceeded $380 billion.

Supporting Anthropic’s high valuation are two logics:

The obvious one: Anthropic’s commercialization progress is more aggressive than DeepSeek’s. On the consumer side, various memberships sell like crazy, relying on their strong model capabilities to sell without worry, often employing “big company bullying smaller” tactics—limiting token quotas today, implementing real-name verification tomorrow. It’s a seller’s market, with annualized revenue reaching $14 billion.

The underlying logic: Claude, as a top-tier model, has become a symbol of American technological strength, with international capital like Singapore’s GIC, Blackstone, Goldman Sachs, Temasek, and Nvidia all betting heavily here.

In short: these international capitals favor US AI more.

The huge investments from these international capitals are not just commercial bets but also strategic signals. This “leadership premium” accounts for a significant part of Anthropic’s valuation. Everyone still remembers last year’s events—after DeepSeek-R1 exploded in popularity, it caused quite a stir in the US tech and policy circles. As a top US AI research institution, Anthropic’s strategic value was further amplified in this context.

DeepSeek’s $10 billion valuation follows a similar logic.

DeepSeek’s core technological asset is its highly cost-effective training method, capable of training excellent models at very low costs. More importantly, this methodology demonstrates a highly “dangerous” possibility to the world: it doesn’t rely heavily on Nvidia, nor does it require huge investments, yet it can produce great large models.

This signal is quite serious.

Because if we push this further, a beautiful scenario emerges: perhaps in the near future, DeepSeek will use its cost-effective training methods to train open-source models on domestically produced GPUs that rival Claude in performance.

If this becomes reality, it would be a “hellish” picture for Nvidia, Claude, and the international capital betting on them.

Therefore, from the perspective of the old guard, if we only evaluate DeepSeek’s profitability, a $10 billion valuation is clearly too high. But if we assess its strategic position, this valuation might be “too modest.”

All in all, Liang Wenfeng is still too conservative.