Who is more valuable, Kimi or DeepSeek?

(Author of this article is Singularity Research Society, authorized by Titanium Media to publish)

Text | Singularity Research Society, Author | Yan Xu, Editor | Meng Wen

Recently, the so-called “Chinese AI Open Source Twin Stars”—Kimi (Dark Side of the Moon) and DeepSeek (Deep Exploration)—have been intensively trending.

First, new models are being released one after another. Kimi just launched K2.6, and DeepSeek quickly followed with V4.

Second, there’s a dramatic shift in the capital market trend.

Two days ago, Kimi announced it had completed approximately $2 billion in funding, led by Meituan Longzhu, with a post-investment valuation of about $20 billion. After three rounds of financing and nearly a year of commercialization, Kimi’s ARR has surpassed $200 million, with paid subscriptions and API revenue accelerating growth.

Almost simultaneously, DeepSeek, which had long refused external funding and relied almost entirely on Fantom Quantitative’s “self-supply” for funding, officially opened up for external investment.

The latest news reveals that DeepSeek plans to raise 50 billion yuan, with founder Liang Wenfeng planning to contribute 20 billion yuan first, accounting for 40% of this round. The post-investment valuation has been pushed from the initial hundred-billion-level to over $51.5 billion, about 2.5 times Kimi’s.

Once realized, this will not only be the largest single funding record in Chinese AI history but also set the valuation ceiling for all Chinese startups’ initial funding rounds.

Both are open-source models, challenging trillion-parameter scales. Why, at the capital table, do these two companies have such a huge difference in chips?

Two Flavors of Money

If we only look at the funding amount, Kimi is currently the most successful large-model startup in China in terms of financing.

Since its founding in 2023, Kimi has accumulated over 37.6 billion yuan in funding.

This number looks dazzling. But if you break it down, you’ll find that what Kimi received isn’t just “money,” but a complete resource system deeply tied to capital, cloud providers, and internet giants.

In early 2024, Alibaba invested about $800 million in Kimi, becoming its largest single shareholder with about 36% ownership. This financing marked a real turning point for Kimi.

However, this $800 million was not all cash. A significant portion was in the form of Alibaba Cloud computing credit quotas, with actual cash investment less than $600 million.

In other words, the “ammunition” Kimi received is essentially an advance on cloud resources; the amount consumed reduces the quota accordingly; Alibaba counts this part as cloud business revenue.

Between cloud providers and large model startups, “you have me, I have you”—here, it takes on another meaning.

Later, Tencent oversubscribed a financing round, making both internet giants competitors and important shareholders of Kimi.

In the latest $2 billion round, Meituan Longzhu, China Mobile, CPE Yuanfeng, and others joined as investors. Alibaba and Tencent also continued to oversubscribe in earlier rounds.

It’s rumored that Tencent is also in talks with DeepSeek. It’s investing in Kimi while also engaging with DeepSeek.

For Tencent, this is more like an “insurance strategy for the AI era”; but for Kimi, Tencent simultaneously balances Alibaba and may also become a behind-the-scenes capital force for competitors.

This is the reality of market-driven AI startups. The more money, the more demands behind each investment.

DeepSeek is backed by Fantom Quantitative. For a long time, DeepSeek’s R&D was almost entirely supported by Fantom’s own funds, with no external VC, no funding timetable, and no binding to cloud providers.

Therefore, in recent years, Liang Wenfeng could take his time. While others competed on commercialization, he focused on training efficiency; while others fought for entry points, he continued open sourcing.

Now, DeepSeek is preparing to introduce external capital for the first time. Even so, Liang Wenfeng still maintains tight control, planning to invest 20 billion yuan himself, accounting for 40% of this round. It’s rumored that the National Integrated Circuit Industry Investment Fund is negotiating to lead the investment, and the appearance of the national team could change DeepSeek’s nature.

These two types of money essentially correspond to two company models: if Kimi is the most typical market-oriented AI startup, DeepSeek is more like an extension of “national strategic capability.”

More interestingly, while the outside world is still comparing which company’s model capability is stronger, their underlying technologies have already quietly “fused.”

In DeepSeek V4’s technical report, it used Kimi’s Muon optimizer; in Kimi K2’s architecture, it incorporated DeepSeek’s MLA.

Their papers cite each other, their tech stacks are intertwined—like two gears meshing, competing while powering each other.

Source: iFanr APPSO

OpenAI even pointed out in a paper that: Kimi and DeepSeek are “the earliest companies to reproduce OpenAI-o1 Long-CoT.”

But now, they are no longer just “followers” of OpenAI. K2.6 brought SWE-Bench Pro’s 58.6% agent cluster parallel programming capability; V4 made contexts of around a million tokens a standard feature, with output length extended to 384K tokens.

Additionally, both companies are advancing domestic chip adaptation.

DeepSeek V4 will support Huawei Ascend 950 in the second half of the year, and Cambrian has already completed Day 0 adaptation; Kimi K2.6 also supports hybrid inference on domestic chips. Agent capabilities, programming ceilings, million-token contexts, domestic chip adaptation, open-source ecosystems… several routes are almost converging.

From “learning to think” to “learning to do,” from “modifying Transformers” to “modifying the computing power base,” this seemingly competitive technological evolution shows that Chinese AI is gradually moving away from merely benchmarking OpenAI, reducing dependence on Nvidia, and forging its own path in the open-source ecosystem.

Why Is It Profitable but Valued Lower?

Kimi already shows the embryonic form of a “mature AI startup.”

It has C-end products, paying users, and a clearer Agent commercialization path. Whether through membership subscriptions or API revenue, it’s already entering an accelerated growth phase.

Kimi’s ARR, i.e., annual recurring revenue, has surpassed $200 million. This figure was proactively disclosed by Meituan Longzhu.

In the primary market, investors emphasizing ARR are essentially endorsing the valuation. After all, few domestic AI startups have established stable revenue models.

DeepSeek’s logic is entirely different. Its core strategy is to first build an ecosystem, then pursue commercialization.

DeepSeek’s API pricing has long been maintained at about one-tenth of OpenAI’s. It cares more about model penetration, developer ecosystem, and open-source influence than short-term revenue.

Therefore, to this day, DeepSeek’s actual revenue remains undisclosed. Meanwhile, its user base is rapidly expanding. Currently, DeepSeek’s monthly active users have reached 127 million, 14 times Kimi’s 9 million.

A very subtle situation has emerged: Kimi, with over $200 million ARR and a more mature commercialization path, is valued at about $20 billion; while DeepSeek, whose revenue scale is still undisclosed and emphasizes low-cost open access, has already soared to $51.5 billion, about 2.5 times Kimi’s.

This reflects a shift in the valuation logic of the capital market.

Today’s AI investments reward not just “how much you can earn now,” but also “what you might become in the future.”

Once national capital truly enters, DeepSeek’s narrative could shift to “China’s AI infrastructure,” and its valuation logic would naturally no longer be based solely on traditional corporate P/E ratios.

In the AI investment context of 2026, Kimi’s “profitable” status might instead mean a clearer boundary and limited imagination.

But this valuation paradox won’t last long.

The Information reported that after this round of financing, DeepSeek will “accelerate revenue planning and commercialization,” while speeding up model releases to “align with industry mainstreams.” It’s also known that the V4.1 launched in June will include tools specifically for enterprise users.

This means DeepSeek is also being pushed to tell a commercial story.

In the past, Liang Wenfeng could be unhurried. Because Fantom’s funds had no external LPs and no exit cycle. But once external capital enters, the clock starts ticking.

The issues Kimi faces today—revenue, growth, commercialization efficiency, capital expectations—DeepSeek will likely face in the future too.

To some extent, Kimi’s $200 million ARR is more like a “pioneer map.”

Yáng Zhílín and Liang Wenfeng’s Accounts

Yáng Zhílín and Liang Wenfeng are both from Guangdong. One from Shantou, the other from Zhanjiang.

Kimi and DeepSeek are among China’s first players of open-source trillion-parameter models, sharing a very similar technological faith: both believe in Scaling Laws and are challenging trillion-parameter-level large models.

DeepSeek is more skilled in reasoning models, while Kimi emphasizes Agent capabilities.

Though their technical routes differ, their underlying goals are highly aligned. Especially in the most fundamental architectural innovations, the two companies often “collide” in the same direction.

Kimi published a paper on “Attention Residual,” while DeepSeek developed mHC residual connections;

Kimi explored Kimi Linear in linear attention, DeepSeek advanced DSA in sparse attention. Seemingly different routes, but fundamentally both are challenging the “ancient infrastructure” of Transformers.

However, regarding “how to protect technological ideals,” they have taken completely different paths.

Yáng Zhílín’s approach is through institutional design: dual-class shares, dual-layer equity structures, absolute voting rights for the technical team. Additionally, he brought in Zhang Yutong, who initially appeared as a partner at GSR Ventures during Kimi’s fundraising negotiations and was key in helping Kimi secure nearly $1 billion from Alibaba.

Later, due to conflicts over interests with GSR, she left the fund, experienced a period of public controversy, and by late 2025, publicly appeared as “Dark Side of the Moon President,” fully responsible for strategy, funding, and commercialization.

These are precisely areas where Yáng Zhílín is less comfortable or unwilling to spend long-term energy. He is a typical technical founder. In Nvidia GTC 2026, he spent much time discussing Muon, training efficiency, and stability issues at trillion-parameter scales.

Liang Wenfeng is also a tech geek, but he controls the company more directly: with real money.

In the first round of external funding, he invested 20 billion yuan himself, accounting for 40%, without relying on complex institutional arrangements or special voting rights, aiming to “beat capital with capital.”

It’s hard to say which approach is smarter. Institutional design offers high leverage with less equity, but it’s human-made, and execution can lead to friction, disputes, or unforeseen costs.

According to reports, during Kimi’s early days, Yáng Zhílín took core team members from his previous company, Circul Intelligent, but the old shareholders’ waiver agreement was never fully signed. At that time, the hype around large-model financing led to many issues being implicitly accepted as “just get on board first.”

Later, with Alibaba’s nearly $1 billion investment, disputes surfaced.

Zhang Yutong, who negotiated this round for Kimi, was then a managing partner at GSR Ventures, and her husband, Wang Zhen, was also a co-founder of Kimi. Subsequently, Tiger Zhu Weitong publicly commented that “fiduciary duty is a red line,” and the old shareholders of Circul Intelligent initiated arbitration.

Therefore, a carefully designed dual-class share structure cannot fully resolve the personal relationships and procedural issues left at the company’s founding.

The benefit of real money is clarity—no ambiguity—but it requires having that much money and being willing to invest it. Liang Wenfeng can afford this with Fantom backing.

Different choices also reflect the resource endowments of the two companies. Kimi has been a market-oriented startup from the start, so it must learn to coexist with capital long-term; DeepSeek, having survived the toughest phase with self-funding, can handle control issues more assertively.

Each has its own way of doing things.

Technologically, they are each other’s “infrastructure.” Business-wise, they once followed two different tracks. But as DeepSeek begins to introduce external capital, these tracks are gradually converging.

Money is weighty. Once external capital flows in, all companies will ultimately face the same bill.

Kimi has taken the lead; DeepSeek is just starting.