Where has AI investment reached?

There has never been a shortage of dream chasers in the AI race.

Over the years, the list of entrepreneurs has been constantly refreshed. From Tang Xiao'ou, Yin Qi, Zhu Long, and Zhou Xi in the era of the "Four AI Dragons," to Wang Huiwen, Li Kaifu, Wang Xiaochuan, Yang Zhilin, Yan Junjie, and Tang Jie after the explosion of large models, and now the popular post - 90s entrepreneurs like Wang Xingxing, Peng Zhihui, Xiao Hong, and Guo Wenjing... Numerous familiar or unfamiliar names have flocked into this AI wave.

Another force driving this wave forward at an accelerated pace is capital. Investment institutions such as Sequoia Capital, Hillhouse Capital, Inno Angel Fund, Yida Capital, and BlueRun Ventures are active on the front line of AI, accompanying entrepreneurs.

Behind this is a huge industrial chain worth trillions, attempting to reshape the entire economic ecosystem. If we break down the AI industrial chain, it can be roughly divided into three layers:

Foundation layer: computing power, algorithms, and data;

Technology layer: large models, platform tools, and general technologies;

Application layer: robots, mobile/wearable devices, drones, and implementation scenarios in various industries.

These three layers are interlocked, yet present distinct development opportunities and investment logics.

According to statistics from Xiniu Data, in August this year, there were a total of 163 investment and financing events in the AI field, with a disclosed total financing amount of 7.68 billion yuan. Compared with the same period last year, the number increased by 66, but the financing amount decreased by 43%. Among them, the intelligent robot track had the highest number of investment and financing events, with 33 in total, followed by AI healthcare, chips, and computing power.

On the surface, investors are making more frequent moves, but in reality, they are becoming more cautious in "digging into their pockets." In the AI industrial chain, some are selling shovels while others are digging for gold. So, who has struck real gold? And who is just accompanying the race?

After communicating with multiple investors, we found that the technology layer has become a game for large - scale enterprises, leaving few opportunities for investment institutions; the foundation layer requires in - depth industrial accumulation and patience, and is now mostly dominated by state - owned capital and RMB funds; the application layer is generally favored, and its core lies in the insight into the essence of business.

For investors, simply "staying in the game" is far from enough: either run fast enough to seize the window of opportunity or dig deep enough to build an ecosystem. Staying in the middle ground actually poses the greatest risk.

Technology layer: A game for large - scale enterprises, a narrow space for venture capital

The first wave of AI fever in the public's impression started in 2016 when AlphaGo defeated Lee Sedol. This event pushed the popularity of AI from the industrial circle to the general public. Investment institutions were extremely enthusiastic about all projects related to "AI," and later the Four AI Dragons, SenseTime, Megvii, Yitu, and CloudWalk, gradually became well - known.

The second wave of AI venture capital fever originated from the explosion of large models. Large models are also the most representative track in the technology layer.

In late November 2022, ChatGPT emerged out of nowhere and reached over 100 million users within two months of its launch. In March 2023, GPT - 4 was launched at an astonishing speed, which directly ignited the entrepreneurial enthusiasm for large models in China.

Within less than a year, the domestic market quickly entered the "Hundred - Model Battle."

Large - scale enterprises such as Baidu's "ERNIE Bot," Tencent's "Hunyuan," and Alibaba's "Tongyi Qianwen" have successively released their large models.

After Wang Huiwen, the co - founder of Meituan, entered the market to establish Beyond Light Years, a group of startups including Baichuan Intelligence, Lingyi Wanwu, Dark Side of the Moon, Minimax, Jieyue Xingchen, and Zhipu AI emerged, known as the "Six AI Tigers." These projects once became the focus of Internet giants and investment institutions.

However, some investors predicted at an early stage that large models are more like a "game for large - scale enterprises."

One reason is the huge investment amount and high risk. The valuation of Lingyi Wanwu in its angel round was as high as $1 billion; Dark Side of the Moon raised 2 billion yuan in its angel round, and more than $1 billion in its Series A+ round eight months later. Investments of such a scale can usually only be borne by Internet giants like Tencent, Alibaba, and Meituan, except for super - top funds like Sequoia Capital China.

Another reason is that the investment window period is very short. The valuation of some projects doubled within just a few months, and institutions often faced the situation of being "unable to afford" before they could even make a decision.

Take Minimax as an example. Its valuation in the Pre - A round in July 2022 was $500 million, $1.2 billion in the Series A round in June 2023, and soared to $2.5 billion in the Series B round a few months later.

The third reason is more realistic: the return period of large models is long, and the monetization path is unclear. Funds need to consider the exit issue. In contrast, large - scale enterprises have massive data, abundant capital, and a large user base. Investing in large models is a crucial and inevitable choice for them - they are all afraid of becoming the "Nokia of the AI era." Seizing the opportunity of large models means holding the ticket to the future.

Based on these three points, some early - stage institutions choose to observe cautiously.

Wang Sheng, a partner at Inno Angel Fund, once said bluntly, "Since OpenAI launched ChatGPT, we have judged that the ultimate winners of this war will be large - scale enterprises. Entrepreneurs' opportunities are either to sell to large - scale enterprises or focus on some vertical fields, both of which determine that the investment value will not be particularly high."

Now, the domestic general large - model market has gradually converged into the "Top Five Base Models": ByteDance, Alibaba, Jieyue Xingchen, Zhipu AI, and DeepSeek.

DeepSeek, relying on the capital advantage of Magic Square Quant, takes the path of open - source and engineering optimization; ByteDance and Alibaba adhere to self - research; in addition to receiving support from state - owned capital, Zhipu has received investment from Tencent, Xiaomi, Meituan, and Alibaba, and Tencent is also among the investors of Jieyue Xingchen.

Interestingly, in the battle for large - model investment shares, the alliance of large - scale enterprises has become a rare scenario in investment history. Tencent has bet on Minimax, Zhipu, Baichuan Intelligence, Jieyue Xingchen, and Dark Side of the Moon; after acquiring Beyond Light Years, Meituan has also invested in Zhipu and Dark Side of the Moon; Alibaba has invested in Dark Side of the Moon, Minimax, Zhipu, Baichuan Intelligence, and Lingyi Wanwu.

Currently, the few investment opportunities left for institutions in the technology layer are vertical models.

Compared with general large models, the commercialization path of vertical industry models is relatively clear.

Zhang Yu, the founding partner of Qingzhi Capital, said, "As long as it can truly reduce costs, improve efficiency, and solve problems for customers, industry models can make money." According to him, all five or six projects that Qingzhi has invested in the industry - model track have generated revenue, and some have achieved profitability.

When the story of general large models gradually gives way to vertical models, the real opportunities for venture capital institutions emerge.

Foundation layer: The water sellers are making huge profits

If the end - game of large models is a game for large - scale enterprises, then what about the more fundamental foundation layer?

The AI foundation layer specifically includes the following four aspects:

Computing power: AI chip/hardware manufacturers, cloud computing platforms;

Data: data services and processing, data providers;

Model toolchain: AI development frameworks, MLOps platforms, vector databases;

Security and compliance: AI security, ethics, and compliance.

These seemingly "heavy - asset" tracks generally have the characteristics of definite demand, high technological barriers, and long return periods. Investment institutions choose the most classic "water - selling logic."

During the "Gold Rush" in the United States in the 19th century, few people actually struck gold, but those who sold shovels and water made a fortune. The AI industrial chain is very similar: it is unknown whether a project can succeed, but the "water sellers" in the foundation layer are almost guaranteed to make a profit.

Take two representative companies as examples.

In fiscal year 2025, NVIDIA, which provides computing power for many AI manufacturers, had an annual revenue of $130.5 billion, a net profit of $72.8 billion, and a gross profit margin as high as 75%. In the recently released Q2 report for fiscal year 2026, its single - quarter revenue was $46.7 billion, and the net profit was $26.4 billion.

Cambricon, known as the "first domestic AI chip stock," also achieved a performance explosion. In the first half of 2025, its revenue was 2.881 billion yuan, the net profit attributable to the parent company was 1.038 billion yuan, and the gross profit margin was 55.93%. Its stock price briefly exceeded that of Kweichow Moutai at the end of August, becoming the "king of stocks" in the A - share market.

Of course, the valuations of these leading enterprises are high, and the capital threshold is extremely high. However, investors can still share the dividends of industrial growth by deploying in more segmented upstream and downstream sectors, such as high - speed interconnection, optoelectronic chips, and advanced packaging.

At the investment level, different types of institutions have different strategies:

Industrial capital such as Alibaba and Tencent values whether the invested enterprises can form synergy with their own businesses. For example, Alibaba has successively invested in companies such as Cambricon, Horizon Robotics, and Deephi Technology because Alibaba is involved in a large number of AI application scenarios in its e - commerce, payment, and cloud computing businesses, with a huge demand for computing power and algorithms. Investing in the foundation layer can not only meet its own business needs but also consolidate and extend the core advantages of Alibaba Cloud.

RMB funds are the mainstream investment institutions in the AI foundation layer. They are deeply involved in the industry and will conduct a combined layout in the early and later stages based on in - depth research.

For example, Tongchuang Weiye focuses on deploying mature projects that have certainty, meet IPO requirements, have stable performance growth, and conform to policy guidance on the one hand; on the other hand, it invests in early - stage and small - scale projects, paying attention to new AI technologies, new architectures, and new talents.

Yida Capital, formed through the internal mixed - ownership reform of Jiangsu High - tech Investment Group, while supporting the self - controllability of key technologies, looks for "essential items" in key links of the AI industrial chain. For example, Yida recently invested in Nanjing Zhixin Materials, a company focusing on large - size lithium niobate (tantalate) materials. Yida values its strategic value as a core material in fields such as AI optoelectronic chips, AR displays, and high - speed communications.

Zhou Zhe, a partner at Yida Capital, introduced that in the computing power layer, Yida has recently focused on deploying enterprises in edge - side/inference - side AI chips, optical interconnection, and heat - dissipation materials. "By inferring the underlying technology from the market - side demand, we can improve the certainty of investment."

This approach of deploying along the industrial chain can not only introduce industrial resources to the invested enterprises and squeeze out valuation bubbles but also capture industrial inflection points through cross - verification, which is expected to achieve higher investment returns.

Overall, currently, investors generally agree on two definite investment lines in the foundation layer:

One is "domestic substitution and self - controllability," covering fields such as optoelectronic chips, high - speed interconnection, and advanced packaging.

The other is that after the improvement of infrastructure, the application layer will create greater value space, which will in turn drive the demand for underlying computing power. Enterprises that are closer to the terminal application market, can rapidly iterate their products, and focus on solving specific problems are expected to rise rapidly.

Smart money sells water, builds roads, and bridges. The "water - selling logic" in the foundation layer provides them with more stable returns.

Application layer: The most bustling arena, from embodied intelligence to low - altitude economy

If the technology - layer investment has become a game for large - scale enterprises and the foundation layer is a long - term layout for state - owned capital and industrial capital, then the application layer is the direction where investors can truly show their prowess at present. The combination of different scenarios and AI is giving rise to a number of new opportunities.

The hottest segment is embodied intelligence.

In August, several robot and embodied - intelligence companies such as Mech-Mind Robotics, Songyan Power, Lingdong General, and Zhipingfang completed a new round of financing one after another. There were also reports about the listing plans of Unitree Robotics and Zheyuan Robotics, pushing the capital enthusiasm to a peak.

In Wang Sheng's view, the core value of robots has now shifted from hardware to "AI intelligence." The "brain" responsible for cognitive decision - making has made rapid progress thanks to the breakthroughs in large models, while the "cerebellum" responsible for motion control and real - time response has lagged far behind due to the lack of a unified technical route. "The existing cerebellum technology path is likely to be completely subverted in the next two or three years."

Based on this judgment, Inno Angel Fund invested in Qianjue Technology (which excels in "software," emphasizing the universality and adaptability of the "brain" and supporting multiple hardware platforms) and also deployed in key links of the industrial chain such as core components to build a collaborative ecosystem.

He predicted that by the second half of 2026, the market evaluation criteria will shift from "telling stories and releasing demos" to commercial implementation. Companies that cannot verify their application scenarios will be eliminated.

Cai Wei, a partner at Lightspeed China Partners, believes that the end - game of embodied intelligence will be "a hundred flowers blooming." "Because it needs to be combined with usage scenarios, even long - tail scenarios."

For example, elderly - care service robots need a gentle touch and emotional interaction ability, while warehouse - handling robots require strong load - bearing and navigation abilities. These fundamental differences in demand put forward completely different requirements for the form, skills, intelligence level, reliability, and cost of robots.

The low - altitude economy is another track that has attracted a lot of capital attention.

In the first half of 2025, there were a total of 52 financing events in the low - altitude economy track, a year - on - year increase of 48.6%, involving an amount of 1.74 billion yuan. Among them, the drone segment accounted