NewsFlash Articles Data Fundraising BlockBeats Pro

DeepSeek raises $100 billion, moving one step closer to an IPO.

动察 Beating

Read this article in 22 Minutes

If all goes well, as early as the end of this year.

The original title: "DeepSeek Raises RMB 100 Billion, One Step Closer to IPO"
The original author: Sleepy

On October 22, Bloomberg reported that DeepSeek's second round of financing was officially completed, with a scale of approximately RMB 100 billion and a pre-money valuation of RMB 500 billion.

Including the first round of approximately RMB 51 billion delivered in June, DeepSeek raised about RMB 150 billion in just half a year. Across both rounds, the largest external shareholders were Tencent and CATL.

From July to October, the second round of financing was paused once. The fundraising target was revised from an initial at least RMB 10 billion to RMB 50 billion, then RMB 80 billion, and finally settled at RMB 100 billion.

Half a year ago, this would have been almost unimaginable. In its first three years, DeepSeek did not take a single cent of external money; the bills for R&D and computing power were entirely supported by Liang Wenfeng's quantitative private fund High-Flyer. In a 2024 interview, Liang Wenfeng also said that DeepSeek had no financing plans in the short term.

This spring, DeepSeek ran into talent problems, which made Liang Wenfeng change his mind.

In the V4 technical report released at the end of April, among nearly 300 authors, 10 names were marked with asterisks, indicating that the people had already left. Wang Bingxuan, a core author of the first-generation large model, went to Tencent; Guo Daya, a core researcher of R1, left in March and was reported in April to have joined ByteDance Seed; Luo Fuli, who left earlier, went to Xiaomi to lead the MiMo large model.

It is said that big tech companies have been poaching DeepSeek's people at high prices. At that time, the options held by DeepSeek employees did not yet have a fair market price. A piece of paper without a price tag cannot stop real gold and silver.

Liang Wenfeng later laid it out plainly to investors, saying that maintaining team stability is the company's only non-negotiable interest. On the other side, DeepSeek was already building its own data centers and planning to manage computing power itself. Relying solely on High-Flyer's resources, it would be difficult to sustain consumption of that magnitude on its own going forward.

According to CNBC, it was precisely after core talent gradually drained away that DeepSeek made up its mind to open up to financing. Going to the primary market to set a sufficiently high fair price for options became Liang Wenfeng's move of paying a thousand gold for a horse's bones.

Ticket

As soon as the news got out, the door was immediately crowded with people eager to hand over money.

An investor, in order to meet Liang Wenfeng, paid RMB 5 million as a meeting fee to an intermediary who claimed to be close to Liang Wenfeng, made three trips to Hangzhou without meeting him, and eventually got the money back. A fund with state-owned background entrusted someone to approach High-Flyer executives with an intent of RMB 3 billion, only received a company email address, and never heard back. On social media, some intermediaries even peddle quotas of unknown origin, with upfront fees as high as 17%.

The final list of those who secured a ticket was handpicked by Liang Wenfeng himself, with most institutions receiving invitations from him proactively. Funds that neither knew him nor had sufficient brand clout could not even get their foot in the door.

The agreement was signed at the end of May, with closing on June 16, and the first round was finalized at approximately RMB 51 billion.

Of this money, the largest portion came from Liang Wenfeng personally, contributing approximately RMB 20 billion. Among external shareholders, Tencent contributed approximately RMB 10 billion, the CATL ecosystem approximately RMB 5 billion; NetEase, JD.com, IDG Capital, and Monolith each took approximately RMB 3 billion; Zhengxin Valley and Shixiang Capital each approximately RMB 1.5 billion, and the National Artificial Intelligence Industry Investment Fund approximately RMB 1 billion.

In just two months, DeepSeek's valuation more than doubled. In late April, the market valuation was still around $20 billion; by the closing of the first round in June, the pre-money valuation was set directly at approximately RMB 350 billion. On this basis, the company carved out a new 5% employee option pool, which, converted at the first-round price, instantly had a book value of approximately RMB 17.5 billion. Including this option pool, the actual pre-money valuation was approximately RMB 367.5 billion, or about $54.3 billion, with the post-money valuation approaching RMB 400 billion.

However, although external capital came in, Liang Wenfeng gave away very little control. The vast majority of external investors could only invest into Hangzhou Chengli, a limited partnership managed by Liang Wenfeng, gaining indirect equity exposure, with shares locked up for five years. The only external institution that directly entered the shareholder register and obtained voting rights was the National Artificial Intelligence Industry Investment Fund.

July's business registration change records show that Hangzhou Chengli holds 8.52%, with Tencent accounting for more than 30% of that; the National Artificial Intelligence Industry Investment Fund holds 0.28%. Liang Wenfeng's direct shareholding was diluted from 34% to approximately 31%, and including indirect holdings through Ningbo Cheng'en, the total remains above 75%.

More Volume, Same Price

Right after the first round closed, in mid-July, negotiations for the second round began. The initial signal was a target of at least RMB 10 billion, with a pre-money valuation of no less than RMB 480 billion.

But by the end of July, DeepSeek suddenly verbally notified several investors who were preparing to sign that the agreements originally scheduled to be inked within a few days would be put on hold. It is said that the leak of the content of the first round of closed-door meetings with investors caused too much market buzz, and Liang Wenfeng wanted to keep the second round of talks more low-key.

In early August, the second round quietly restarted below the radar.

This time it proceeded very quietly, and several institutions that had been at the negotiating table in July did not even receive notice of the restart. The fundraising target was directly raised to RMB 50 billion, with a pre-money valuation set at about RMB 500 billion.

Although the quota was opened up, screening of funding backgrounds was not relaxed at all. Some investors tried to enter through SPVs, but the source of funds and the ultimate upstream contributors behind them were scrutinized layer by layer, and they were ultimately blocked at the door.

Lisi, Shixiang, Zhengxin Valley and Shixi, several new and old shareholders of DeepSeek, have all been intensively raising funds for new funds over the past two months, and a string of listed companies followed with subscriptions. Zhihu plans to contribute RMB 1.5 billion, Dahua Co. plans to contribute RMB 500 million, Century Huatong's subsidiary subscribed RMB 650 million, and even Zhongwang Fabrics, which makes furniture fabrics, put in RMB 50 million.

The money kept piling up. By October 6, the quota locked in by DeepSeek was already close to at least RMB 80 billion; on October 13, the final lineup surfaced, with the new shareholders confirmed as Shixi Capital, SMIC Juyuan, CPE Yuanfeng, Legend Capital and Geely; among the old shareholders, Tencent, CATL, IDG Capital, Lisi Capital and Shixiang Capital continued to follow on.

On October 22, the transaction was officially completed. On the same day, Perfect World issued two announcements: the two funds it subscribed to, Lisi Xingchen and Shixiang Shangshi, had completed fundraising, with scales of RMB 7.52 billion and RMB 3.65 billion, respectively. Lisi and Shixiang were both on the follow-on investment list for this round.

The final closing size stopped at about RMB 100 billion, ten times the initial target in July. But the pre-money valuation remained locked at RMB 500 billion, with more volume but no price increase.

Inside and Outside the Card Table

If the first round of financing helped DeepSeek complete option pricing, then the shareholder list in the second round reveals a strong industrial orientation.

Looking at the backgrounds of the new shareholders, two come directly from the heart of the chip industry. Shixi Capital is a semiconductor CVC initiated by GigaDevice, and the other identity of GigaDevice Chairman Zhu Yiming is precisely the founder of ChangXin Memory Technologies.

It is said that Liang Wenfeng set a check size threshold of RMB 5 billion for direct investment in this round, and the amount Shixi Capital plans to contribute in this round is exactly about RMB 5 billion. Another new entrant, SMIC Juyuan, has SMIC behind it. Add Legend Capital under Legend Holdings and Geely, the only automaker on the list, and the weight of industrial capital is quite obvious.

Among the existing shareholders, Tencent contributed approximately RMB 10 billion in the first round, and in this round it again ranks among the largest contributors alongside CATL. As early as February 2025, when R1 went viral, Tencent Yuanbao and WeChat Search successively integrated DeepSeek; by the end of April this year, the news that Wang Bingxuan, a former core author of DeepSeek, joined Tencent was also confirmed. Tencent has both taken in people who came out of DeepSeek and absorbed the largest external share of DeepSeek's first round.

CATL is likewise placing a heavy bet, contributing approximately RMB 5 billion in the first round and doubling that to about RMB 10 billion in this round. In the view of many analysts, CATL's increased stake reflects its recognition of the enormous energy demand behind computing clusters.

Equally intriguing are the names that are not at the table.

Alibaba and ByteDance did not appear in either round; they each have Qwen and Doubao. Hillhouse is likewise absent from the shareholder list. However, in September this year, DeepSeek brought in a new Chief Financial Officer, Yan Wentao, a former partner at Hillhouse Ventures who had previously invested in ByteDance, Zhipu, and MiniMax.

Ulanqab and Buying Chips

With RMB 150 billion coming in, DeepSeek has never publicly broken down exactly how this massive sum will be spent, but based on available information, the absolute bulk is still computing power.

As early as the end of April this year, DeepSeek posted job openings for data center operations engineers and delivery managers, with the work location listed directly as the Ulanqab grasslands. By early September, more specific news emerged: DeepSeek plans to build a gigawatt-level data center in Inner Mongolia, housing at least 160,000 Huawei Ascend 950DT chips.

But the problem is that having money does not necessarily mean being able to immediately buy enough chips.

Against that target of 160,000 chips, Huawei can allocate only about 16,000 chips to DeepSeek, while internet giants receive quantities in the hundreds of thousands. Liang Wenfeng has calculated that four Huawei chips are equivalent to one Nvidia chip, while also lagging two generations behind. By September, affected by high-bandwidth memory shortages, Huawei raised the price of the 950DT to more than RMB 250,000 per chip, an increase of fully 20% to 50% compared with two months earlier.

In the primary market, money is easy to raise, but chips are hard to buy.

To solve the chip bottleneck, the two sides are becoming increasingly deeply bound. In September, the market narrative was still that these Huawei chips would only be used for inference, while core training would still have to rely on Nvidia chips obtained by every possible means. But by October, the news was that Huawei engineers had already been stationed long-term in DeepSeek's offices, writing the requirements DeepSeek encountered during training into the design of next-generation chips; DeepSeek would begin running pre-training on the 950DT as early as the end of this year, while Huawei planned to reserve most of the first batch of capacity of the 960DT, scheduled for launch in the first quarter of 2027, for DeepSeek. Of course, DeepSeek has not stopped its efforts to find more Nvidia chips at the same time.

At the end of September, DeepSeek also open-sourced six underlying components for Ascend. On the external investment front, DeepSeek participated in the strategic placement of Unitree Robotics' IPO in August, subscribing approximately RMB 141 million, and also plans to lead the investment in Winverse AI, a robotics company founded by former Tencent AI scientist Hu Han.

2029

Beyond computing power and talent, another hidden thread already points to the secondary market, which is also an important reason why former Hillhouse partner Yan Wentao joined as CFO.

As early as late August, DeepSeek began contacting accounting firms and investment banks, vigorously preparing for a STAR Market IPO, planning to file as early as the end of this year and list in 2027. After Yan Wentao took office, the closing of the second round of financing and the IPO preparations are all matters he is responsible for.

Taking advantage of the window of financing and IPO preparations, the books of this mysterious company were glimpsed by the outside world for the first time. According to financial figures obtained by foreign media, DeepSeek's revenue from January to July this year was approximately RMB 475 million, 10 times that of the entire previous year, with the gross margin of its API business reaching 82.9%, while its net loss during the same period was approximately RMB 715 million. By September, the company's ARR had already reached approximately $1 billion, equivalent to about RMB 6.7 billion.

What underpins this revenue curve is the model and infrastructure iteration that did not stop even during the financing period. On September 10, DeepSeek released V4.1 Flash; the 1017 version updated on October 17 scored 45 on the Artificial Analysis Intelligence Index, surpassing Haiku 5.5, which Anthropic had released just ten days earlier; then on October 20, DeepSeek open-sourced its internal inference framework DeepInfer, compatible with both Nvidia and Huawei Ascend 950.

However, for investors who have just squeezed onto the table, beyond the books there is a more realistic timing issue.

According to the current regulations of the CSRC, new shareholders who make last-minute investments within 12 months before filing must lock up their shares for 36 months from the date of acquisition. If DeepSeek indeed files at the end of this year as planned, the new and existing institutions that poured RMB 100 billion into this round will not be able to cash out their chips until 2029 at the earliest. The portion that came in through limited partnerships in the first round is subject to a five-year lock-up and will have to wait until 2031.

What price the secondary market will offer three years from now—there are currently two completely different reference frameworks in the market.

One is Zhipu on the Hong Kong stock market, whose share price surged to a high of HK$2,980 on June 22, before closing at HK$484.8 by October 20; the other is Changxin Technology on the STAR Market, which soared 465.82% on its debut on July 27, pushing its market value to RMB 3.28 trillion and vaulting it into first place by market capitalization on the A-share market.

Two years ago, Liang Wenfeng said that DeepSeek had no short-term financing plans, a statement that had already become a thing of the past by this spring. During the first funding round, he also told investors that his original intention in founding the company was not to go public. If all goes smoothly, by the end of this year at the earliest, that statement will probably also need to be revised.