Outside the doors of large model companies, there are hordes of investors eager to pour in money.

One of the biggest variables for large model companies this year is that DeepSeek has also started to seek financing. This means that even the idealist Liang Wenfeng has reached a new cognitive turning point.

However, different from other model companies, even when raising funds, Liang Wenfeng is not aiming for rapid commercialization. Recently, Bloomberg cited people familiar with the matter as saying that Liang Wenfeng promised at least once in an investor meeting that the company would prioritize technological breakthroughs, pursue greater goals, and achieve AGI rather than focus on making profits and short - term commercialization.

It's not just DeepSeek that is highly sought after by capital. Kimi, Jieyue Xingchen, and even Zhipu and MiniMax in the secondary market have all seen their valuations soar several times in the capital market in the first half of this year.

In the past three years, the investment wave in large models has at least reached two capital peaks. Judging from the participating institutions, the amount of funds raised, and the decision - making time, the capital wave in 2026 is undoubtedly more intense and more concentrated. When it comes to investing in model companies, investors have their own criteria, but there is an increasingly clear consensus that AI brings a paradigm revolution, and top - tier model companies will become the largest leading enterprises in the industry chain.

01

The Last Few Tickets on the Ship

People queued up to invest in DeepSeek. "Recently, many LPs, both known and unknown to me, approached me, wanting to invest in DeepSeek. There is at least 4 billion yuan in funds looking for an investment opportunity," said Li Siyang (a pseudonym), an investor from a first - tier investment institution. "It's really crazy. Many people want to get in."

In April, as soon as the news that DeepSeek was seeking financing spread, it caused a stir in the investment circle. First - tier investors tried every means to get close to those at the core of DeepSeek's financing to gather information. Even most DeepSeek employees who didn't have detailed information about the financing received a lot of invitations from investors to meet, not to mention Liang Wenfeng himself. A DeepSeek employee told us that he didn't even dare to go out and meet people.

Cao Bin, the founding partner and chairman of Gaoxin Capital, didn't hide his feelings during an external interview: "If there is an opportunity to invest in DeepSeek, we are more than willing to do so."

But in fact, most financial investors are not within DeepSeek's consideration. On April 23, we exclusively reported that DeepSeek's valuation in this round of financing was as high as 300 billion yuan, equivalent to 44 billion US dollars. This round of financing is not targeted at ordinary financial investors. It mainly involves state - directed funds, and there are requirements for the identity of LPs.

The investment institution where Li Siyang works once learned about DeepSeek's financing situation through very high - level channels, but was rejected due to the strict financing conditions.

"If DeepSeek lowers its threshold, no matter how much money is needed, it can basically be found. Everyone knows that DeepSeek's valuation of 300 billion yuan is definitely worth investing in," said Li Siyang. Graduated from a top overseas university's computer major, Li Siyang has a keen technical perspective and has participated in the financing of several top - tier model companies at home and abroad in the past few years. When Kimi announced its financing last year, she strongly recommended that her institution buy shares in Kimi in multiple rounds.

The competition in large models is essentially a comprehensive competition in terms of talent, money, and computing power. Training a model with trillions of parameters involves hardware, algorithms, data, and electricity. The investment in GPUs for hardware alone starts at hundreds of millions of dollars. After more than three years of competition among model companies, only a few independent general - purpose model manufacturers can still stay in the game. Besides DeepSeek and Kimi, they also include Jieyue Xingchen among the six major AI startups, as well as the listed Zhipu and MiniMax.

It's difficult for new model companies to enter the market in the future. For investors with FOMO (fear of missing out) emotions, the scarcity of top - tier model companies means that no matter which one they manage to invest in, it may be the last chance to get on board in this round of the AI era.

In March, Kimi launched its latest round of financing, originally planning to raise 1 billion US dollars. As a result, it raised 2 billion US dollars, exceeding the target by 1 billion US dollars. Its valuation also more than quadrupled compared to three months ago. In April, a partner of an industrial investment institution focusing on the general automotive technology field was looking around for Yin Qi's phone number, wanting to directly invest 100 million yuan. He didn't conduct due diligence and didn't know much about AI, but he believed that "once Jieyue Xingchen goes public, its valuation will rise to 200 billion yuan."

After Zhipu and MiniMax went public, the capital market's pursuit of "core large - model assets" was further amplified. The valuations of the two companies in the secondary market once increased by 5 to 7 times, which also put the entire industry back into a strong FOMO mood. Among the several leading domestic model companies that are not yet listed, except for DeepSeek, both Kimi and Jieyue have released IPO plans or capitalization signals to varying degrees.

As the secondary market continues to give high - premium feedback, the valuation logic of large - model companies in the primary market has also been further pushed up. Following Kimi's financing pace, Jieyue Xingchen announced on May 8 that it had completed a new round of financing of 2.5 billion US dollars. Coupled with the 5 billion yuan raised in the Pre - IPO round in January this year, Jieyue's valuation also tripled in three months.

The optimistic mood is almost reflected in the entire fundraising process. The investment decision - making process is accelerating. Traditionally, VC investments require a long - term due diligence process, but now neither LPs nor GPs have enough time for due diligence. Large investment institutions in the market that can directly get a share have to lock in some shares first, conduct due diligence while raising funds, and are even forced to participate in two or more rounds of investment.

At the end of last year, an investment institution got a share in a model company, only short of the last 8 million US dollars. They held a roadshow, and within 24 hours after the roadshow ended, all the money had arrived. The LPs on - site were very enthusiastic and even transferred the money without signing the investment agreement in time.

Investors don't have a chance to hesitate. If they don't have more prior knowledge, they can either invest blindly or miss the opportunity.

Rarely in the market, there is a dual - valuation structure in the same round of financing, and the agreed share price is also changing all the time. Since this year, Jieyue Xingchen has launched its Pre - IPO. Industrial capital and well - connected investors bought in at a valuation of 4 billion US dollars. Financial investors will enter in the next round. The originally agreed valuation price was 6 billion US dollars, but in fact, the financial investors who entered in April bought at a price as high as 8 billion or 9 billion US dollars, a two - fold increase.

Many investors complained that "this is crazy" and blamed Yin Qi, the founder of Jieyue Xingchen, for raising the price.

Zhang Yingjie, the managing director of Dachen Venture Capital, told us that this is determined by the supply - demand relationship. There is a lot of money in the market, but there are only a few model companies worth investing in. Most investment institutions' mentality is that as long as they can get a share, they will invest. He gave the most obvious example that since this year, while the core business of leading companies has not changed substantially, their valuations have increased by 3 to 5 times.

02

The Model is Reaching a Cognitive Turning Point

For many investors, Kimi and Jieyue Xingchen are the only two basic - model investment targets available in the current market. In comparison, DeepSeek is undoubtedly a more attractive option. However, due to its financing rhythm, degree of openness, and the scarcity of shares, ordinary investors have very few opportunities to access it.

When choosing between the only two companies, many investors don't need much hesitation to bet on Kimi. "The logic for a good company is to buy as much as possible," said investor Li Siyang. "When Kimi had a valuation of 10 billion US dollars, those who couldn't invest blindly didn't understand AI."

There is more than just sentiment behind this judgment. Kimi is still one of the few domestic AI products that have truly built a user base. Apart from the AI assistants of large companies and DeepSeek, it still maintains a good user base on the C - end. At the same time, its model performance also has a good reputation at home and abroad. The Composer 2 model base of the AI coding tool Cursor is from Kimi's K2.5 model.

Some investors feel a bit confused about Jieyue Xingchen, which has been very active this year. Whether from the perspective of its C - end products or its all - in approach to the agent terminal, Jieyue Xingchen hasn't left a deep impression on the outside world. A relatively cautious LP, Zheng Ming, told us that since January this year, they have tried to understand the company's model capabilities from actual business scenarios, but after four or five months, it's still difficult to form a clear perception from the outside.

Last year, Jieyue Xingchen actively scaled back its C - end business. Its role - playing product "Maopao Ya" was shut down, leaving only the AI assistant tool Jieyue AI. At the same time, in the B - to - B field, it fully shifted to terminal scenarios, including automobiles, mobile phones, and embodied intelligence. Jieyue has built a strategic narrative of "AI + terminal" and abandoned the currently popular pure API - call commercialization model for large models.

This path has indeed attracted a group of companies focusing on terminals. For example, Huaqin Technology, a global leading mobile phone ODM manufacturer, Longcheer Technology, which focuses on electronic device manufacturing, OmniVision Technologies, a core supplier in the image sensor field, and ZTE, which spans communication equipment and terminals.

But for Zheng Ming, this strategy doesn't have much appeal to ordinary investors. The key for him to judge whether to invest in a company is whether he can directly perceive the model's capabilities in real scenarios. Currently, Jieyue's open - to - the - public products are limited in scale, and user feedback is also limited. He even specifically experienced the intelligent cockpit jointly launched by Jieyue Xingchen and Geely Automobile but still didn't form a deep enough perception. In the end, he didn't choose to invest in Jieyue Xingchen.

The institution where Li Siyang works also didn't invest in Jieyue. Her core logic is that model capabilities are still the primary competitive variable. In her view, currently, it's still a crucial stage where the model is the product, and the model performance is reaching a cognitive turning point. Since the end of last year, she has frequently contacted friends in model companies, and a consensus has formed among a very small circle of scientists, that is, models have started to have a certain "self - research ability." They can not only conduct self - training, find problems on their own, design coding units, put forward scientific hypotheses, and then verify them.

This means that models are evolving from tools into real - sense "research systems."

The most obvious example is the "self - evolution ability" demonstrated by Anthropic's latest model, Claude Opus 4.5, in November last year. During actual tests, the team found that Opus 4.5's understanding ability was significantly better than previous models. It could not only find bugs that previous models couldn't, but also autonomously decide when to act and weigh complex decisions.

Anthropic CEO Dario Amodei also mentioned in an interview with the media in January this year that after the launch of the latest model Claude Opus 4.5, the ability of AI to complete complex tasks end - to - end has reached a turning point. The model's self - improvement ability has also accelerated model iteration. The release rhythm of Anthropic's Claude Opus model has been accelerated from once every three months to once every two months. After Anthropic released Opus 4.5 in November last year, Opus 4.6 was updated and released on February 5, and Claude Opus 4.7 was officially released on April 17.

The desktop agent tool Claude Cowork launched by Anthropic in January this year sparked a domestic boom in desktop agent tools. Amodei admitted that this tool was almost entirely developed autonomously by the Claude Opus model, which only took a week and a half.

The fact that models have autonomous understanding ability not only means an improvement in the code efficiency of R & D personnel and engineers but also means that after the ability boundary of the AI system is broken, unprecedented changes will occur. Li Siyang said, "Two or three decades ago, the largest companies in the world might have been in the tens of billions or hundreds of billions of dollars in scale. In the AI era, I think the growth potential of these model companies will be very large, beyond human imagination."

The capital market is also re - pricing these companies based on their revenue growth rates. Anthropic has soared to become a nearly trillion - dollar company, and its latest announced revenue growth rate in the second quarter is about 127%. The company is expected to achieve an operating profit of 559 million US dollars in the quarter of June. According to the latest analysis by research institution SemiAnalysis, Anthropic's inference gross profit margin has jumped from 38% to over 70%, and its annualized revenue has soared from 9 billion US dollars to over 44 billion US dollars in just a few months.

Similar growth has also started to appear in domestic model companies. After the release of the K2.5 model, its reputation has been continuously improving. Kimi's Annual Recurring Revenue (ARR) exceeded 100 million US dollars in March this year and further increased to over 200 million US dollars in April. Its revenue almost doubled in just one month. It's almost impossible to find such a growth rate in current AI applications. Li Siyang said that model companies will become the largest leading companies in the AI industry, which has become a consensus in the industry.

03

Large Companies Spread Their Nets Widely

In every technological boom, large companies are the most active investors. In recent years, Tencent and Alibaba have almost jointly invested in at least five of the six major AI startups, although some of them have fallen behind.

Alibaba is the large company that has invested in the most large - model companies. Since December 2025, Alibaba has participated in multiple rounds of financing for Kimi and was the leading investor in the C++ round of financing.

The outside world doesn't know how much Alibaba has ultimately invested in Kimi, but Alibaba is always regarded as Kimi's largest external shareholder. As early as the beginning of 2024, Alibaba invested 800 million US dollars in Kimi, setting a record for the largest single - investment shareholder in a large - model company. In March of that year, Alibaba also invested 400 million US dollars and led the B - round financing of MiniMax.

Kimi's long - standing shareholder is Meituan Longzhu. After leading Kimi's A - round financing in 2023, Meituan Longzhu has been participating in its multiple rounds of financing. In the recent D - round financing, it even led an investment of 200 million US dollars. The logic of Meituan Longzhu's investment in Kimi is similar to that of Li Siyang. They still value the technological foresight of the Kimi team. Wang Xinyu, a partner of Meituan Longzhu, has repeatedly told the media about his impression after meeting Yang Zhilin for the first time. Yang Zhilin's understanding of models is very close to the first - principle thinking, and the overall team has strong technological foresight.

In addition to Meituan Longzhu, Kimi has also attracted a new financial investor, the global VC institution Cathay Capital. This company, which manages the largest AI fund in European history, prefers to invest in vertical AI applications or products that combine hardware and software from the perspective of how AI can reconstruct the underlying logic of industries. In the past few years, it has never participated in large - model investments, but in 2026, the first model company it invested in China was Kimi.

Publicly disclosed information shows that Alibaba is the most active large company in investing in large - model companies. It has invested in almost all basic - model companies except Jieyue Xingchen (the equity share of DeepSeek has not been determined yet). Alibaba is an important external shareholder of MiniMax, holding 13.66% of the shares before its listing. Currently, the data disclosed by MiniMax shows that Alibaba's shareholding is 17%. Alibaba also holds about 6% of the shares of Zhipu (before its listing) and is also a shareholder of Baichuan Intelligence and Lingyi Wanwu, which have fallen behind.

It's quite surprising that Jieyue Xingchen, with Yin Qi as its chairman, is not on Alibaba's investment list. You know, in the AI 1.0 era, Alibaba invested in Yin Qi's Megvii Technology multiple times and was one of the important strategic investors in Megvii's early days. Megvii Technology tried to list on the Hong Kong Stock Exchange and the A - share STAR Market but failed to go public due to reasons such as the entity list, regulatory issues, and continuous losses.

Alibaba's substantial investment is largely for computing -