After digging into NVIDIA's 84 investments, we discovered a secret.

In this AI revolution, no company has profited more than NVIDIA.

The latest third - quarter report is out. The operating profit increased by 65% year - on - year, reaching $36 billion. The net profit also grew by 65%, reaching $31.9 billion.

Since the launch of ChatGPT 3.5 in November 2022, NVIDIA's operating profit has increased nearly 19 times. The company also once became the first enterprise in history with a market value exceeding $5 trillion.

Another aspect that has shown exponential growth is the number of NVIDIA's investments in startups.

Since 2022, NVIDIA (including direct investments, CVC NVentures, and NV GPU Venture) has made a total of 251 investments in 244 startups.

From the beginning of this year to November 11th, NVIDIA (including direct investments, CVC NVentures, and NV GPU Venture) has made a total of 84 investments in 74 startups, covering North America, Europe, Asia, and the Middle East. This number has exceeded the 76 investments made in the whole of 2024 and is 4.6 times the 18 investments made in 2022.

We found that in these 84 equity investments, almost all were made in enterprises in fields strongly related to AI, such as software applications, computing power, and energy. Among them, the most were AI application companies, with 39 enterprises receiving investments; the large - model field had the highest financing amount, reaching $28.6 billion.

Looking at these layouts, we may be able to see the outline of the next - generation technology landscape through NVIDIA's eyes.

01. Among the 74 companies, what exactly did NVIDIA invest in?

NVIDIA's investments in startups this year cover a variety of technology fields, but almost all the businesses or products of the invested companies are closely related to AI.

In addition to the most common large models and software applications accessing large models, NVIDIA has also completed preliminary layouts in cutting - edge fields such as robotics, computing power, autonomous driving, and quantum computing.

In terms of quantity, among the 74 invested companies, 39 are AI application startups, 10 are model - related companies, and 5 companies each in robotics and data centers received investments.

In terms of transaction scale, 10 large - model companies raised more than $28.6 billion in total, while 40 application - side companies raised less than $6.5 billion in total. Although $20 billion of the $28.6 billion raised by model companies came from the latest round of financing of Musk's xAI, the remaining $8.6 billion raised by the other 9 companies still exceeded the total financing of application - side companies.

At first glance, the gap in financing scale between model companies and application companies may seem a bit surprising, but if you think carefully, such a gap may be reasonable.

Compared with model companies, the cost and threshold for AI application - layer entrepreneurship are much lower. They usually rely on existing large models and cloud computing power for inference, without the need to invest huge amounts of money in model training, underlying algorithm research and development, or building their own computing power infrastructure. Therefore, the early - stage capital demand is smaller, and the financing round amount is naturally lower.

At the same time, currently, application - layer products rely on general large - model interfaces, so their replicability is relatively high, with limited barriers and a lack of core differentiated technologies, which makes investors more cautious about their valuations.

Model companies, on the other hand, master underlying capabilities and occupy a platform - level position. They need to invest a huge amount of computing power and top - notch scientific research talents, with huge capital expenditures. Therefore, the financing scale generally shows the characteristics of "a small number of companies, very large amounts".

In addition, heavy - asset data center and robotics enterprises each raised about $2.5 billion, and two companies in the quantum computing field received a total of $1.8 billion in investments.

02. What signals does NVIDIA's investment release?

NVIDIA's above - mentioned 84 investments release at least four clear signals.

1. Narrowing investment in the model layer

Looking at the few model companies invested by NVIDIA this year, the overseas large - model track may have begun to enter the early stage of "centralized integration" (see the figure below).

In these model companies, we see some highly similar characteristics: either they have established a foothold in the market and are in the late - stage financing rounds, or they are founded by the most influential figures or teams in the industry.

For example, Mistral AI, a French large - model company that completed a $1.5 billion Series C financing in September, is not only a leader in the European AI field but also one of the most crucial enterprises in French President Macron's plan to develop France into a European AI center. Perplexity, which completed a Series E financing in July, is a veteran large - model enterprise with 170 million monthly average visitors and an annual income of $150 million.

On the side of star teams, xAI is a large model founded by Musk himself to compete with OpenAI. Safe Superintelligence is a super - artificial - intelligence model founded by Ilya Sutskever, the former chief scientist of OpenAI, this year with the goal of AGI. Thinking Machines Lab is also founded by the former CTO team of OpenAI.

Overall, NVIDIA did not make "scatter - gun" investments in the model layer but clearly bet on a few global companies with the potential to become model leaders.

2. The energy problem needs to be solved urgently

At the beginning of the month, Microsoft CEO Nadella said that due to insufficient power supply, some AI GPUs can only be idle in data centers and cannot be used. As the largest supplier of AI GPUs, NVIDIA has obviously noticed this problem that must be solved.

Since this year, NVIDIA has invested in 9 energy or resource management application enterprises (see the figure below).

From NVIDIA's layout in the energy field, it is clear that it is actively betting on key technologies that can increase power supply, including the fusion company Commonwealth Fusion Systems and the companies Crusoe and Nscale that build data centers driven by renewable energy.

Commonwealth Fusion Systems is a pure clean - energy enterprise dedicated to building commercial fusion equipment. Currently, the company is building a reactor prototype and plans to start formal testing in 2026. In addition to NVIDIA, the company has also received investments from Google, Bill Gates, and other parties.

Crusoe is a computing - power supplier that uses clean and wasted energy to power data centers. The company captures the natural gas that is flared during oil extraction and converts it into electricity to power modular data centers deployed nearby, thereby increasing power supply and reducing electricity costs.

Similar to Crusoe, Nscale is an enterprise that builds data centers in regions with abundant renewable energy around the world. The company has currently built a 30 - megawatt data center completely powered by hydropower in Norway.

At the same time, another type of companies invested by NVIDIA - Utilidata, PassiveLogic, Phaidra, DeepAware AI, Emerald AI, Yasu - focus on power grid management, energy - consumption optimization, building energy automation, etc. Their goal is not to increase power itself but to improve the operating efficiency of energy in transmission and use.

NVIDIA's bet on these efficiency - oriented companies shows that it has realized that the next - stage growth of AI must rely on more intelligent power dispatching, more efficient energy utilization, and a more flexible power - supply structure.

3. B2B applications dominate

Among the 39 AI software applications invested by NVIDIA this year, 34 focus on enterprise customers (see the table below).

This includes Agent - type applications that improve enterprise software development and programming efficiency, such as Poolside, which just received $2 billion in early - stage financing in November.

There are also platform - type applications that help enterprises analyze and manage data and improve advertising placement, customer expansion, and market research, such as VAST Data, Profound, Exa, etc.

In addition, NVIDIA also invested in many AI - enabled engineering modeling, material research and development, and medical diagnosis software, continuously promoting future technological innovation.

We believe that the main reason why NVIDIA focuses its investments on B2B software enterprises is that, compared with consumer - oriented applications, the business models of enterprise - level software are often more stable and predictable.

Enterprise customers usually purchase software through long - term subscriptions or pay - as - you - go methods. Once deployed, the migration cost is high and replacement is difficult. Therefore, the cooperation relationship has significant stickiness. This makes enterprise - level revenue not only repeatable but also able to grow continuously through renewals, expansions, and cross - sales, forming a relatively clear cash - flow curve.

This can also be reflected in the comparison between OpenAI and Anthropic. Anthropic, which has about 300,000 enterprise customers, is estimated to have an income of $9 billion this year, while OpenAI, which currently has 800 million weekly active users, is estimated to have an income of only $13 billion.

Another point is that enterprise - level applications generally have a relatively greater demand for inference computing power. This means that investing in these startups is not only more likely to bring further growth to NVIDIA's own business but also can establish a deeper binding relationship in the long - term business of customers.

We believe that these two points are the core reasons driving NVIDIA to focus on B2B applications.

4. Robotics

Compared with models and applications, NVIDIA's investments in the robotics direction are relatively cautious, with only 5 enterprises invested this year.

In the robotics direction, NVIDIA's investment focus is slightly inclined to the intelligent layer of robots, such as companies like Field AI, Skild AI, Generalist, which are building robot basic models, operating systems, and general brains.

At the hardware level, NVIDIA invested in the current flagship humanoid - robot company Figure AI and the dexterous robotic - arm company Dyna Robotics.

We believe that the hardware - software balance shown in NVIDIA's investments in the robotics field is worthy of attention. While betting on embodied hardware, more attention should perhaps be paid to the "intelligent layer" such as models, perception, and control.

The 4 signals released by NVIDIA's investments, whether it is the centralization of the model layer, the early layout for energy bottlenecks, or the systematic bets on enterprise - level applications and the intelligent layer of robots, all show the way a technology giant allocates resources and builds an ecosystem. It also provides us with a unique window to observe the evolution of the AI industry.

03. Earning $31.9 billion in the third quarter, intensive investments will continue

Why does NVIDIA make investments so frequently?

The first reason is the simplest: NVIDIA has so much cash that it cannot effectively digest it through traditional methods.

Counting the just - released third - quarter financial report, in the recent four quarters, NVIDIA has spent about $52 billion on share buybacks and dividend distributions and $5.8 billion on capital expenditures, both hitting record highs.

NVIDIA also did not save on R & D. It has spent $16.7 billion on R & D in the past twelve months, almost doubling the $8.6 billion in fiscal year 2024.

After these operations, NVIDIA's cash and cash equivalents still accumulated to $60.6 billion. This figure is almost equal to the sum of 4 semiconductor giants: AMD, Broadcom, Qualcomm, and Intel.

Considering NVIDIA's market position and scale, it is no longer realistic to make relatively large - scale mergers and acquisitions in vertical fields such as chips and computing power. For example, in 2020, NVIDIA wanted to acquire ARM Holdings, which was not yet listed at that time, for $40 billion, but the deal was stopped by the Federal Trade Commission on antitrust grounds. After a two - year tug - of - war, NVIDIA finally chose to give up.

As an international giant, mergers and acquisitions also require the approval of regulatory agencies in other related countries. NVIDIA's $69 billion acquisition of Mellanox in 2020 was investigated by the Chinese market regulatory authorities on antitrust and other grounds in 2024.

Therefore, compared with the high time and litigation costs that large - scale mergers and acquisitions may bring, making flexible equity investments across a wide range of startups naturally becomes a better choice.

The second reason is that NVIDIA needs to diversify its revenue sources.

According to the third - quarter financial report, 50% of the revenue in the third quarter of this year came from three customers. In contrast, in the third quarter of 2024, the three leading customers only contributed 36% of the total revenue. This data means that NVIDIA is becoming more and more dependent on a few other technology giants.

The accompanying risk is that once these giants start to reduce capital expenditures or diversify their chip suppliers to reduce their dependence on NVIDIA, NVIDIA's revenue will be greatly affected.

Therefore, with such a concentrated revenue, NVIDIA's frequent investments are mainly to make the invested enterprises become customers, in order to reduce its dependence on giants and prepare for creating new revenue sources.

CEO Jensen Huang also said at the GPU Technology Conference (GTC) last month that "more and more startups are creating more scenarios for using NVIDIA GPUs". Considering that NVIDIA's chips are still very attractive to startups at this stage, establishing cooperative relationships through investments allows NVIDIA to lock in revenue growth during the stage of the explosion of AI applications after the AI construction period is gradually completed.

Although NVIDIA's investments do not require the invested enterprises to only purchase NVIDIA's chips, undoubtedly, NVIDIA is the best choice. And once an enterprise uses NVIDIA's chips and CUDA architecture, it seems that the conversion cost of completely abandoning them will be even higher.

The third reason is that NVIDIA needs to explore more new businesses.

According to NVIDIA's financial report, in the first nine months of fiscal year 2026 (from January 2025 to October 2025), the revenue from data centers accounted for 89% of the total revenue, while the revenue from autonomous driving accounted for only slightly more than 1%.

Through extensive investments in technology - field startups, NVIDIA can quickly understand the market demand, usage scenarios, and data of relevant markets, help the company identify technological trends, complete strategic layouts in advance, and reduce the possibility of being directly