Skip to main content
guanjiawei.ai
Blog
DeepSeek Valued at 300 Billion, Liang Wenfeng Invests 20 Billion
AIDeepSeekLiang Wenfenggeek spiritthoughts

DeepSeek Valued at 300 Billion, Liang Wenfeng Invests 20 Billion

After reading about DeepSeek's fundraising at a 300 billion valuation, I'm not surprised. What truly shocked me was something else — Liang Wenfeng pulled out 20 billion, entering the same round alongside external investors at the high valuation he set himself. I've never seen anything like this in my career.

Jiawei GuanJiawei Guan9 min read
Share:

Yesterday when the DeepSeek funding news broke, I was taken aback scrolling through my social media feed.

The outside world is discussing the 300 billion RMB valuation, the lineup led by China's major funds with Tencent and Alibaba following, and the fact that this is the first time the company has ever raised money since its founding. All of this matters. But after reading the news several times, what truly shocked me was another set of numbers: DeepSeek contributed 20 billion internally in this round, entering alongside external investors at this same valuation. And Liang Wenfeng holds nearly 90% of the shares.

In all my years working in the AI industry, I've never seen a move like this.

1. Buying Your Own Cake

An analogy makes clear how abnormal this is.

Imagine you have a cake. You baked it for several years, and your employees baked it with you. One day the market says it wants to put a price on this cake, and the estimate comes out to 300 billion. Everyone is thrilled, because the small slice each person holds is suddenly worth money.

Normally, at this point the founder breathes a sigh of relief: the money previously spent is water under the bridge, external capital is coming in at a high valuation, the company has cash, and employee stock options have an anchor price.

But Liang Wenfeng did one more thing. At the 300 billion valuation he himself set, he pulled out another 20 billion to buy in alongside external investors.

He himself accepts this 300 billion price tag, and is willing to double down. Translated into the polite language of a shareholders' meeting, that's "I'm very optimistic about the company." But polite words don't require 20 billion in real cash.

The source of the funds is also intriguing. Liang had never raised external capital before; all the company's money in its first few years came out of his own pocket. Founders like this are usually wary of outside investors, and they only open up for the first time because they're running out of cash. But DeepSeek isn't short on money. High-Flyer Quant alone had assets under management exceeding 70 billion RMB in 2025, and the dividends from this single private equity firm to him personally are enough to fund research for many more years.

The real purpose of this round is to give employee stock options a market-recognized price anchor. The model industry is currently in a fierce war for talent; without a clear valuation, you can't retain people with equity. The fundraising was for this purpose, but the founder himself following on with 20 billion far exceeds that purpose. This is a statement made with money: I believe in this price myself, and I'm willing to keep adding to my position at this price.

2. The First Time I Heard About This Company

Rewind to late 2023.

Back then, the hottest labels for large models were the "Four Little Tigers" and "Six Little Tigers": Zhipu AI, Moonshot AI, Baichuan, MiniMax, 01.AI, and StepFun — the rankings shifted depending on who you asked. Investors and media were all circling this list. I was at Zhipu AI at the time, and no one in the entire circle had heard of DeepSeek.

Until one day a friend asked me on social media: Have you heard of DeepSeek? I heard their cost advantage is pretty insane.

I went to look up this company's background. It was spun off from a private equity fund called High-Flyer Quant, formally established to work on large models only in May 2023. The earliest versions weren't particularly impressive in terms of capability. They had originally aimed for performance, but when performance fell short of expectations, the MoE architecture unexpectedly reduced costs by an order of magnitude. Then they did something quite surprising: since costs were low, they simply lowered API prices.

This was the first time the industry heard the name DeepSeek. Not because the model was stunning, but because the price was cheap, and the training cost data was written in their paper — anyone in the tech community could see at a glance that they really had something.

3. At That Time, Only a Few Nationwide Could Field a Ten-Thousand-Card Cluster

2023 was the height of the US-China chip war. The A100 had long been placed on the restriction list, and available stock was dwindling. I was at Zhipu AI at the time and could directly feel that the A100 was a scarce resource. There weren't many players domestically who could truly field a ten-thousand-card A100 cluster. SenseTime was one; by its 2022 annual report disclosure it had 27,000 GPUs, with A100 reserves exceeding ten thousand.

Then DeepSeek's numbers surfaced: High-Flyer had already invested 1 billion yuan in 2021 on "Firefly No. 2," building a cluster of 10,000 A100s. Meaning before large models had become a mainstream topic, they had already prepared the hardware foundation for today's large model battles.

The industry began taking this company seriously starting from this point. A quant private equity firm going to hoard ten thousand A100 cards and building a self-developed deep learning platform. This isn't something ordinary secondary-market players would do.

4. The Price Butcher

May 2024, V2 was released.

The moment the price list came out, everyone realized something was different. Input at 1 yuan per million tokens, output at 2 yuan per million tokens — roughly 1% of GPT-4 Turbo's price at the time. This wasn't a price cut; it was moving prices onto an entirely different track.

The chain reaction followed quickly. Zhipu AI slashed GLM-4-Plus by 90%, from 50 yuan all the way down to 5 yuan; ByteDance's Doubao directly quoted 0.0008 yuan per thousand tokens, claiming the industry's lowest; Alibaba's Tongyi and Baidu's Ernie also followed with cuts. For a whole month, the industry's pricing baseline was rewritten.

This price war was different from any seen before. Previously, model companies cut prices following OpenAI; however OpenAI cut, everyone followed. This time, a previously unknown Chinese company set an anchor two orders of magnitude lower than OpenAI, dragging the entire domestic market onto a completely different cost curve.

5. So Low-Key It Was Abnormal

DeepSeek doesn't resemble a typical Chinese AI company in many ways.

It's in Hangzhou, not Beijing. It has no government background, nor did it seek local government support. The CEO doesn't do roadshows, doesn't give interviews, doesn't post on social media. Before fundraising, it hadn't cultivated VC relationships; employee hiring relied on word-of-mouth in technical circles, not headhunters.

The most interesting thing was the positions it hired for. At the time, a job title was circulating in the industry called "Data Know-It-All." The requirements were: strong interest in various domains, high self-drive, and passion for achieving AGI. No particular educational background required; knowing how to read files and call APIs in Python was enough. It explicitly stated "no algorithm derivation or handwritten code required." The essence of this role was teaching AI how to be more human, so it actually valued general knowledge and curiosity more. At the time, a product manager friend of mine saw the job description and told me he wanted to apply; in the end, he didn't get in.

I myself also once wanted to go. There was a sense of mystery about this company, stemming from a rare purity: head down doing technology, nothing else. Later there was never quite the right opportunity, but I kept paying attention.

6. From V3 to R1, Pulling the Industry Back to the Main Thread

December 2024, V3 was released.

Performance directly reached top-tier levels among open-source models. Then before long, on January 20, 2025, R1 went live, and they directly released an app. This was DeepSeek's first consumer-facing product, coinciding right before the Spring Festival.

Everyone saw what happened the following week. R1 topped the app stores during the Spring Festival holiday, overseas media reported on it collectively, and people from OpenAI were commenting on X repeatedly. The entire AI agenda for the first half of 2025 was rewritten by this single event.

But R1's real impact wasn't its popularity itself. It was that it pulled a batch of Chinese large model companies that were being forced by capital and markets into "forced commercialization" back to the main thread of "first make the model good."

Before DeepSeek, most Chinese large model companies were closed-source. The logic was: open source gets freeloaded, hurting commercialization. R1 directly proved this logic untenable. A completely open-source model can simultaneously have the best technical influence and the broadest user base; commercialization is something that will follow naturally.

By mid-2025, Kimi open-sourced K2, Zhipu AI open-sourced GLM-4.5, MiniMax open-sourced M2.1, Qwen continued its day-0 open-source strategy, and a whole group of companies previously insisting on closed source all switched. The market no longer discussed the question of "open source or closed source"; it shifted to "is your model good enough?"

DeepSeek held no launch event, made no loud noise, but every time it made a move, the way the industry played its cards changed.

7. xAI Is a Different Script

Someone might say: DeepSeek isn't the only idealistic AI company; isn't xAI one too?

When xAI started, its mission was also to use AI to seek truth, and its narrative had some resemblance to DeepSeek's. It was also founder-funded, team-built from scratch, and challenging big tech. But looking at this company today, the story is completely different.

February 2026, SpaceX announced the acquisition of xAI, merging at a SpaceX valuation of 1 trillion and an xAI valuation of 250 billion dollars, for a combined 1.25 trillion dollars. In March, all 11 of xAI's co-founders from that year resigned. In May, xAI ceased to exist as an independent company, was merged into SpaceX, and renamed SpaceXAI. Musk turned around and rented most of the 220,000 GPUs to Anthropic.

This is another reasonable script: sufficiently large compute, sufficiently strong capital narrative, ultimately absorbed into a larger parent body. If you look at xAI's current product cadence and research intensity, it no longer looks like the original "research for research's sake."

This isn't saying xAI took a wrong turn; it's saying idealism doesn't automatically lead to DeepSeek's path. Whether a company can maintain research priority and product restraint after raising tens of billions is an independent choice, and the further you go, the harder it becomes.

DeepSeek's contrast lies right here. It has reached the level of a 300 billion RMB valuation, with a funding round of roughly 7 billion USD, and market attention on par with the top AI companies globally. But Liang Wenfeng still doesn't come out. In the recent V4 Preview release notes, he quoted a line from Xunzi's Against the Twelve Masters: "Not tempted by praise, not frightened by slander; follow the Way and act, correct yourself with dignity." He didn't say this for PR; his behavior over these years itself has been exactly this.

8. Geek Spirit

China's AI industry isn't short of technical people; what's lacking are people possessed by the obsession of changing the world through technology and willing to get their hands dirty.

Wang Jian was the previous representative of this type. He didn't come from a formal computer science background; his 1990 doctorate was in psychology from Hangzhou University, and from 1993 to 1998 he was even head of the psychology department at Zhejiang University. But in 2009, as CTO of Alibaba Software, he presided over the R&D of the Apsara Cloud operating system from scratch. At the time, few believed China could build a cloud OS from the ground up, but he kept his head down and did it. In the end, Alibaba Cloud grew from lines of code to a business worth tens of billions. In 2019, he was elected as an academician of the Chinese Academy of Engineering — the first from a private enterprise background.

My understanding of geek spirit isn't about how hardcore one's technical skills are; it's about believing technology can make the world better, and being willing to roll up your sleeves and do it.

What I admire about DeepSeek is that it can still preserve this at the scale of a 7 billion USD funding round. At a time when everyone is being pushed to commercialize, pushed to become a hot topic, pushed to exit, there are still people willing to treat research itself as the end goal.

China should have more teams like this. Not all need to work on large models; they can do anything. The key is having this degree of purity in what you do.

That 20 billion isn't about the money. It's a declaration: this company will continue down the research path, and I myself am willing to continue walking that path with it.

Whether this company ultimately succeeds or fails, whether V4 or V5 performs well or not next — as long as this declaration continues to be made, as long as this state is maintained, it deserves everyone's attention.

Respect.

References

Recommended Reading

Subscribe to Updates

Get notified when I publish new posts. No spam, ever.

Only used for blog update notifications. Unsubscribe anytime.

Comments

or comment anonymously
0/2000