How should one value DeepSeek?

Why did AI · DeepSeek’s valuation multiply fivefold in three weeks? What are the underlying reasons?

Author | Wang Zhaoyang
Email | wangzhaoyang@pingwest.com

Within three weeks, DeepSeek’s valuation was revised four times in public reports:

In early April, approximately $10 billion “initial funding,” on April 22nd, reports said its valuation exceeded $20 billion, on May 6th, it was rumored that a “big fund” was negotiating to lead the investment, with a valuation around $45 billion, and last week, the latest reported valuation cap was over $50 billion—if true, this would be the largest single-round funding ever for a Chinese AI company.

According to reports, the largest check in this round of funding did not come from VC or internet giants, but from DeepSeek founder Liang Wengfeng himself: he personally might invest up to 20 billion RMB, accounting for 40% of the total funding. It is said that he increased his stake from 1% to 34% through capital injection, and with indirect holdings, he controls approximately 84.29% of the equity.

The “Big Fund,” officially the National Integrated Circuit Industry Investment Fund, which has previously invested in semiconductor core capital for SMIC and Yangtze Memory Technologies, has never publicly invested in large language model companies before. If this round closes, it will be the first time.

One of DeepSeek’s most distinctive labels in the past was not raising funds, not commercializing, and not roadshowing, because its parent company, Fantasia Quantitative, always had sufficient capital reserves. Its proactive commercialization efforts are almost zero, which makes this high valuation, compared to other model companies’ already exaggerated “market-to-model ratio,” even more astonishing.

When you put these facts together, you’ll find that explaining this highly-watched financing with the usual model company framework doesn’t work—why did the valuation increase fivefold in three weeks? Why did the big fund come in? Why did the founder himself put in the largest amount?

This round of financing requires a different framework to understand.

DeepSeek is not just another “better model company,” it’s more like an infrastructure company that looks like a model company. This positioning is fully reflected in this funding.

1

1

Liang Wengfeng once said in 2024 about financing: “Our problem has never been money, but the ban on high-end chips.”

The core propositions of model companies are data, algorithms, and talent; chips are a cost item, not a strategic one. Saying chips are the core issue implies that DeepSeek has been thinking from day one not only about “how to make better models,” but more about “how to rebuild a workable system under compute constraints controlled by others.”

If we trace back all of DeepSeek’s most important technological innovations—from MLA to MoE, from FP8 training to extreme inference efficiency—they are essentially different solutions to the same problem: how to produce top-tier models with less, more restricted compute. This shifts the thinking from a model company’s perspective to the R&D logic of an infrastructure company.

If DeepSeek is not a model company, what should its valuation be? Who should it reference? Not Kimi, not Zhipu, not MiniMax. The disclosed financing structure provides a way to judge.

1

2

On April 24th, DeepSeek-V4 was released, and the technical report for the first time included Huawei Ascend NPU and Nvidia GPU in the same hardware verification list—“We verified the fine-grained EP scheme on both Nvidia GPUs and Huawei Ascend NPUs,” for trillion-parameter-level models, this was the first official recognition of domestic AI chips in formal documentation. On the release day, eight domestic chips including Huawei Ascend and Cambrian completed adaptation, breaking the previous norm of months of debugging. For this line, the DeepSeek team worked closely with Huawei and Cambrian for months, and some reports suggest this was also a key reason for delaying the release. They made extensive adjustments and rewrites at the model’s foundational level.

And the funding was initiated during this period: the timing is very interesting. Other companies raise funds to achieve certain goals; DeepSeek’s funding was because something had already been achieved—and in a verifiable, public way—equivalent to Liang Wengfeng releasing a model that completed what would normally require multiple meetings of due diligence.

In the subsequent funding, there are actually three separate funds, with three different logics, not overlapping.

The Big Fund—full name the National Integrated Circuit Industry Investment Fund, the most important national-level industry capital in semiconductors—has historically invested in SMIC and Yangtze Memory, focusing on “semiconductor infrastructure that China cannot do without.” Its mission is to fill the hardware gaps in semiconductor manufacturing, equipment, and materials. Model companies had never been within its scope before.

Its previous hesitation to invest in model companies was because the commercial return cycle for models is uncertain, and investing in models with semiconductor logic doesn’t make sense. Now, DeepSeek has broken this barrier with one move: V4’s adaptation to domestic chips proves it is not just a model company. The Big Fund can now describe this investment differently: it’s about whether the domestic compute ecosystem can support this strategic node—using top-tier models to drive the application iteration of domestic compute, building an autonomous loop of “domestic chips + domestic models”—this is viewed as the same type of issue as investing in SMIC within its logic.

The industry capital that is rumored to be involved, such as Tencent, represents another motivation: ensuring they are not excluded from China’s AI infrastructure. If DeepSeek ultimately becomes that underlying layer, not holding equity means they can only be paying users, but what they bring is not just money—they are an ecosystem gateway, enterprise clients, cloud resources—precisely where DeepSeek’s commercialization is most weak.

This also explains why several other major tech firms with infrastructure ambitions might ultimately choose not to invest. DeepSeek is seen as a potential infrastructure competitor, not just a model company that can be invested in.

1

3

This creates a sense of “split”: locking in national strategic capital to confirm its infrastructure identity; meanwhile, bringing in industry ecosystem capital to open up commercialization—these two often conflict in many companies, but from this financing structure, DeepSeek tries to keep them running completely independently, without mutual drag.

In other words, after this funding, the pace of commercialization and whether the grand narrative can be realized are unrelated.

The part funded by the Big Fund corresponds to a long-term proposition—compute independence, domestic chip ecosystem, with a cycle of five or ten years, no quarterly returns needed; the industry capital like Tencent corresponds to another track—commercialization can follow the pace of other model companies.

When people discuss its commercialization difficulties, they often overlook some facts: DeepSeek’s API pricing has already broken industry bottom lines—V4-Flash cache hit input as low as 0.02 yuan per million words, with a total cost about 1% of GPT-5.5’s, pushing inference costs down to this level allows DeepSeek to ensure that its API business itself is not unprofitable; according to the latest A16z “Top 100 Gen AI Consumer Apps” report, DeepSeek remains a product crossing US-China user distribution—web traffic is split between China (33.5%) and the US (6.6%). Despite model update “drag,” the mobile user base has not collapsed as outsiders might think. This means that, in theory, other model companies’ profitable methods are also accessible to DeepSeek, with even lower marginal costs.

This split is actually a major characteristic that distinguishes DeepSeek from other companies, and it was intentionally designed from the Fantasia era.

Fantasia Quantitative makes money, DeepSeek does research—this has been its operational structure from day one. Fantasia doesn’t need DeepSeek to generate commercial returns, and DeepSeek doesn’t need to prove its ARR growth to Fantasia. This relationship is hard for outsiders to find a comparable reference; it’s closer to a proprietary trading firm investing profits into fundamental research: research itself is part of its logic, not a project waiting to be monetized.

This split relies on Liang Wengfeng himself holding both sides of the wheel. But the more it develops, the more fragile it becomes—this is personified, not structural—once external capital enters, the risk of breaking the balance becomes real.

So, this round of funding actually aims to turn this dependency on personal will into a fixed part of the equity structure for the first time. The relationship between Fantasia and DeepSeek, once replicated on a larger scale: Fantasia’s first-layer split from DeepSeek, and DeepSeek’s internal commercialization versus compute infrastructure research as a second layer—both layers follow the same logic and are designed by the same person.

Therefore, beyond these two funds, what makes this funding particularly unique is the third: Liang Wengfeng’s own investment.

Reports say Liang Wengfeng personally invested up to 20 billion RMB, about $2.8 billion. In comparison, the recent $2 billion funding round for Moon’s Dark Side Kimi is about the same—meaning Liang Wengfeng’s personal contribution alone is enough to match most single-round funding for model companies.

Thus, the 60% external capital isn’t about keeping DeepSeek “alive” or “training models further,” Fantasia’s funds are sufficient, Liang Wengfeng’s own money is enough. Every external investor entering is because DeepSeek is no longer just a model company.

This personal investment grants a lock on the entire structure—regardless of how much external capital is introduced, core decision-making will not be diluted.

1

4

Additionally, the previously considered biggest reason for this funding was talent loss. But closer observation suggests that talent attrition may have been exaggerated.

Those who left are indeed important—Bingxuan Wang went to Tencent, Daya Guo to ByteDance Seed, Fuli Luo was recruited by Xiaomi with a multi-million annual salary, Chong Ruan joined Yuanrong Qihang, covering core tech lines like foundational models, inference, OCR, and multimodal.

But according to a very detailed recent data compilation by Caijing, among the top 15 most frequently appearing authors in 27 papers, only 2 have left; at the time of the V4 release, DeepSeek-LLM had 86 people, with 71 still listed in the paper, and only 10 out of 300 total staff had left—an attrition rate of 3.3%. Those who follow talent competition among AI companies know how low this number is.

Option pricing has real value for retention, but it’s a secondary issue solved along the way. Doing the right long-term things also naturally retains talent—this is a universal logic behind all DeepSeek’s decisions: it never acts just to solve immediate problems, but once the long-term goals are achieved, the immediate issues tend to resolve themselves.

1

5

Finally, to fully understand its financing and valuation logic, another key topic is: who exactly are DeepSeek’s competitors?

Nvidia has defined the game rules for compute in the AI era; all model companies—both in China and the US—operate within this boundary. DeepSeek is a company trying to redraw this boundary. The moment V4’s Ascend adaptation was implemented, this was the first time this judgment was publicly validated. DeepSeek cares most about Nvidia.

A $50 billion valuation does not price current model capabilities, ARR, or user scale. Globally comparable model companies: Kimi, which just completed a funding round, has a post-investment valuation of about $20 billion, nearly quadrupling from $4.3 billion within half a year, supported by real ARR; Anthropic’s last round valuation was $380 billion, with annualized revenue surpassing $30 billion by early 2026. DeepSeek’s commercial revenue is nearly zero, yet its valuation exceeds Kimi’s by more than twice—and most of this valuation is from the semiconductor national fund, not from growth-stage VCs betting on future ARR.

From the investor structure itself, it’s clear that the premium isn’t about commercialization expectations but about infrastructure premium.

If DeepSeek succeeds in what it aims to do—a frontier AI compute system that doesn’t rely on Nvidia, based on open source, and a Chinese AI infrastructure—its closest comparison would be somewhere between Stargate’s OpenAI and Nvidia, a position not yet defined: it develops top-tier frontier models while trying to solve compute infrastructure with new methods. There is no other company like this globally.

From this perspective, a $50 billion valuation is record-breaking but still undervalued.

This funding prices an event that has not yet happened. Whether it will happen depends on Liang Wengfeng’s ability to accomplish what he has been working on from day one. Looking at the entire path from Fantasia to DeepSeek, he is someone who executes long-termism very thoroughly. Doing the right and long-term thing will naturally solve many specific issues along the way. This round of funding might be the first time the outside world sees a complete outline of this vision.