The next track set to produce a batch of ten-bagger stocks: Who is actually making real money from AI?
1
Current AI and Future AI
If we divide the AI industrial chain into upstream equipment, midstream cloud vendors and large model companies, and downstream application companies, at present, the profits of the industrial chain are almost all concentrated in the upstream. NVIDIA alone plus the four major memory vendors take nearly 70% of the industry's profits.
This profit distribution pattern shows the typical characteristics of the early development stage of new technological revolutions. But let's think: the future revenue entry point of the AI industry will be either cloud vendors or downstream applications. Will the profit distribution still remain the same then?
Take the mature Internet industry as an example, the approximate proportion of economic profit distribution is as follows:
Upstream equipment/technology: about 20%–30%
Midstream network/cloud/operators: about 10%–20%
Downstream platforms and applications: about 50%–70%
Obviously, the closer you are to the user entry point, the higher the profit distribution.
The large-scale capital expenditure of cloud vendors in recent years is essentially advancing part of the future profits of cloud vendors as the profits of upstream equipment companies today; while the future revenue of the AI industry will mainly flow into downstream applications and cloud vendors.
As the growth rate of capital expenditure slows down and AI revenue itself increases, the profit distribution of the entire industrial chain will become more and more similar to that of the mature Internet industry, and continue to concentrate in midstream cloud vendors and downstream applications.
At present, the total market value of the US stock AI industrial chain is about 25 trillion US dollars, of which the upstream (including computing power chips, semiconductor equipment, storage, and communication equipment) accounts for 47%, the midstream cloud vendors account for 50%, and downstream applications account for only 3%, leaving huge room for growth.
If you are investing in AI for this year, then of course you should continue to invest in computing power, but if you are investing in the future of AI, there is no doubt that you should focus more on downstream applications. In the past three years, countless 10-fold stocks have emerged in the upstream, but with the end of the peak period of capital expenditure, more profits will remain in the middle and downstream, and the probability of 10-fold stocks appearing will be higher.
In the past, people always felt that the uncertainty of AI applications was too strong. In fact, it was because the scale of native AI revenue in China was too small, but the United States is different. Several clear paths for AI applications have emerged, and a number of companies with great investment value have taken shape. This article takes the listed AI application companies in the US stock market as examples to sort out several key directions.
2
Business Model Matters for Applications
AI applications are not a homogeneous sector. Extremely fierce differentiation has occurred in 2026: Infrastructure-type software driven by AI workloads (including data platforms, observability, and security) has significantly outperformed the market. "Long on infrastructure software and limit SaaS exposure" has become the consensus strategy for current US stock software investment.
However, not all SaaS software deserves a valuation cut. Judging from the interim reports, the AI revenue of many software stocks has grown rapidly, and at the same time, many software infrastructure companies have very high valuations.
Computing power is judged by its prosperity, while applications are judged by their business models. There are three key variables to judge the pros and cons of segmented tracks:
1. Billing model: Usage/consumption-based billing is naturally positively correlated with AI workloads, which can hedge the risk of seat reduction, while the pure seat-based model faces pressure;
2. Position in the enterprise AI execution chain: Platforms that master core data, permissions, and workflow entry points are expected to be revalued from application software to enterprise AI infrastructure;
3. Whether AI revenue is incremental: The market only pays for traceable AI ARR/ACV/workload indicators, not for AI strategies.
Based on these three criteria, AI applications can be divided into 7 sectors across three levels.
3
Category 1: Data Platforms
Suppose there is a large e-commerce company that decides to go all in on AI, plans to hire as few people as possible in the future, and let AI Agents run the company on their own.
This cannot be solved by simply creating several Agent entry points. The first step that all companies need to take is to build an AI data platform. For AI to make decisions and execute actions, it must first know what happened to the company in the past and what is happening now.
In the past, the executors of company business were all employees, so data could be scattered in various places: user data in CRM, orders in e-commerce databases, inventory and financial data in ERP, customer service records in customer service systems, product information in PIM, advertising data on Meta and Google, employee data in HR systems, meeting minutes in everyone's computers and mailboxes, and more data in employees' brains and chat records (if it is a Chinese company)...
During execution, employees then search and integrate these data from different channels. For example, if the boss asks "Why did the profit in the Shanghai region drop in the last three months", you can log in to different systems to check, ask colleagues if you don't know, and then you can conduct analysis.
Similarly, an Agent that can really work must know all the above data at the same time. Without data, AI is just a very smart outsourced employee, but it does not know what is happening in the company.
Therefore, the first step of AI transformation is to convert all kinds of company information into data that AI can understand, query, and call, the more the better, which is data migration.
In future companies, every meeting of the project, every communication between colleagues, and even every chat of employees will be synchronously recorded and digitized to become the analysis data and corpus of AI, so as to ensure that AI's decisions are based on real-time information of company operations.
The best performers in the US stock market are Snowflake and MongoDB, and of course there is a less typical but more famous company called Palantir.
SNOW has transformed from a single data warehouse service provider to a comprehensive AI data cloud platform, launching a series of native AI products such as Cortex Code and Snowflake Intelligence, simplifying customer data migration, catalyzing the consumption of core data services, and forming a powerful growth flywheel.
The billing of this type of company is all based on usage/consumption, which is naturally positively correlated with AI workloads. The interim report shows that product revenue growth has accelerated for 3 consecutive quarters to 37% in Q2, and the company raised its full-year product revenue guidance to 6.07 billion US dollars (+36%), verifying that the "AI-driven core platform consumption flywheel" has entered the realization period.
However, the valuations of such companies are also extremely high. SNOW has a nearly 20x PS and a market value of nearly 120 billion US dollars, which includes high expectations for the future. The market's doubt lies in its competitive and cooperative relationship with cloud giants such as AWS, Microsoft, and Google.
4
Category 2: Observability
When more and more of the company's business is undertaken by AI, the company finds a serious problem: it does not know what AI has actually done? Whether the consumed tokens are worth it? Which businesses can be optimized?
In the past, the company had a complete set of records, assessment and evaluation systems for employees' work. The most typical one is KPI management. Although it has received many negative reviews, it can ensure that a company with tens of thousands of people operates automatically according to the process. But AI work is a "black box", AI can also slack off and do "useless work", but its work process may be completely different from that of humans, the traditional KPI will fail, and human employees need to constantly see what AI has actually done.
This is the problem solved by observability applications, which is like the HR of AI, through which humans can see what tools the AI Agent has called? Which API has been called? How many Tokens have been spent? Why did it make this decision? Which step went wrong? Which Agent has higher performance? Which Agent caused the loss?
The representative company in this business segment is Datadog (DDOG), which can monitor the performance indicators of traditional APM work such as the customer's cloud, hybrid or on-premise servers, containers, and networks in real time, and can also provide code-level distributed tracing to help customers deeply understand the performance bottlenecks of applications and microservices. It can also collect, process and analyze massive log data from systems and applications for troubleshooting and performance analysis.
The billing model of this type of application is also consumption-based. DDOG's Q2 revenue increased by 36%, achieving accelerated growth for four consecutive quarters, and significantly raised its full-year performance guidance. However, the valuation of this sector is also extremely high, which makes the stock price easily affected by the negative narrative of the software sector.
5
Category 3: Cybersecurity
Cybersecurity is a category that existed in the traditional software era, but its importance has increased sharply in the AI era. Because AI not only attacks your system, but can itself be used as a hacker's inducement tool.
AI essentially replaces employees, and its permissions are greater than most employees. It can not only access databases, modify data, and send emails, but also perform high-risk actions such as fund transfer, even directly write code and deploy programs, which is like directly invading the brain of your employees and "turning them against you".
Agents have reasoning, data access and action capabilities at the same time, so permissions, identity, auditing, and tool call control will all become new security issues.
There are many such companies, both in the US stock market and A-share market, so I will not give examples here.
Security has the highest certainty among the three sectors, mainly due to strong certainty, plus there is a time lag between AI expenditure and security AI expenditure, leading to strong expectations for future revenue growth.
These three categories, data services, observability and cybersecurity, can be collectively referred to as the digital infrastructure of AI applications, which are the essential three-piece suit for all enterprises that implement AI strategies in the future. The expenditure is both upfront and sustainable, so it will take the lead in breaking out among applications.
Built on top of these three types of digital infrastructure are "execution layer" applications, that is, various enterprise work interfaces, which I divide into two categories: one is various Workflow / Agent platforms, and the other is AI programming.
6
Category 4: Workflow/Agent Platforms
With data and permissions, how exactly does the Agent of this e-commerce company work?
There are two types of current solutions. The first type is based on the original workflow, using AI Agents combined with various original industry software to work, which is called "Workflow / Agent Platform".
For example, when this e-commerce company processes customer returns, the traditional process is reviewed by customer service, queries orders, judges the reason, checks inventory, refunds, processes logistics, notifies the warehouse and other work, and every step is operated by people through software.
In the "Workflow / Agent Platform" mode, the process remains the same, but it is automatically processed by a dedicated Customer Agent.
Under this model, in addition to the large model itself, what is more important is to reasonably embed tool calls, permissions, and manual approvals in the original work process, and the difficulty lies in the collaboration between employees and Agents.
Therefore, this type of application companies are all transformed from original software vendors, such as Salesforce (CRM), ServiceNow (NOW), Microsoft, etc. The billing model also adopts the dual model of "seat + Token billing".
However, this type of non-native AI application vendors are also highly questioned, mainly for two points: first, whether their business growth is driven by AI or non-AI factors? Second, in terms of the billing model, can the incremental revenue from new traffic billing fill the decline in traditional seat revenue?
Relatively radical transformation companies, such as ServiceNow, now even position their platform as an enterprise AI infrastructure that connects data, workflows and AI Agents, and its platform can connect to SAP, Salesforce, and external data such as Snowflake and Databricks.
If the "Workflow / Agent Platform" is the mode of AI driving traditional software, then "AI programming" is to write a new program to complete the work.
7
Category 5: AI Programming
This is actually the first B-end scenario that large models have successfully implemented, triggering a surge in ARR in the first half of the year.
AI Agents are envisioned to call existing software to work, but in practice, a large number of software do not have interfaces. It just so happens that large models are best at writing programs. When encountering such problems, it is better to write a program to solve it on the spot, which is the evolution of large models from code completion to conversational programming, and finally to programming agents.
This type of application includes not only GitHub Copilot and Cursor, but also Claude Code, Codex launched by large model laboratories.
I prefer to distinguish the functions of AI programming from traditional software and large AI models. If you really want to use AI to transform enterprises, you must completely jump out of the original workflow, produce "production tools" by yourself, break the boundary between workers and production tools, so as to truly improve work efficiency and give birth to companies with a market value of trillions of dollars like Anthropic.
This type of software has completely changed from seat fees to Token billing, directly cutting into the application layer, and forming a head-on competition with the previous "Workflow / Agent Platform". Whoever becomes the mainstream AI working mode in the future is likely to give birth to 10-fold stocks.
8
Category 6: Vertical Industries and AI Advertising
The first five types of AI applications mentioned above are infrastructure and production tools in general work scenarios, while the sixth type of application is aimed at some specific business scenarios. General large models are difficult to meet the compliance and accuracy requirements of these specific industries, which gives vertical software a deep moat.
This type of "AI application" is actually built on top of the previous five layers, and the most mature one at present is the AI advertising of Meta / Google / AppLovin.
Advertising is the earliest commercial scenario where AI is implemented. AI drives eCPM/ROAS upward by increasing user duration, targeting conversion rate and creative CTR.
Meta's FY26Q1 ad revenue rose 33%, claiming that impressions increased by +19% and unit price increased by +12%, which was attributed to AI. What is more convincing is Google, whose FY26Q1 search revenue increased by +19% and query volume hit a new high, breaking the perception that AI search weakens ad resilience.
When investing in such application companies, the hardest part is to identify the authenticity of AI's driving effect on business. Many AI application companies whose main revenue comes from advertising have a small proportion of AI-driven revenue, and instead face the risk of being subverted by AI; while native AI advertising applications such as AppLovin, almost every dollar of revenue comes from AI-driven advertising, even if the valuation is on the high side, the actual growth potential may still be underestimated.
In addition to advertising, AI applications in vertical industries such as medicine, finance and taxation, education, and design are also developing rapidly. These industries usually have easily quantifiable ROI and strong willingness to pay for AI