AI agents decide everything: 10 companies that will benefit from moving to an AaaS model

In March, CEO Jensen Huang spoke about the new stage of the AI revolution: companies will move from selling software (SaaS) to selling the results of AI work (AaaS). So the very logic of their monetization will change. Most importantly, this is already happening. Igor Klyushnev, co-founder of Freedom Holding, named 10 companies that will benefit from this transition.

On March 16 in San Jose at the GTC conference, Nvidia CEO Jensen Huang took the stage in front of 20,000 people and said something like this: every company that sells software on a subscription basis today will, in the coming years, turn into a company that sells the result of the work of AI agents. It will not update the product, it will not add a button with a neural network, but it will change the very logic of what it charges for.

Imagine that instead of buying a lawn mower, you pay your neighbor to make sure your grass is always cut. You are no longer interested in what device he uses, how many hours he spends or when he comes. You pay for results.

This is what Huang calls the transition from SaaS, a service-as-a-service business model, to AaaS, an AI-agent-as-a-service model. It used to be that companies sold you a CRM system and you did the data entry, funnel building and reporting yourself. Now the AI agent does it for you, and you pay not for access to the tool, but for the work done.

Huang's words are not just a beautiful metaphor from the conference. Transformation is already underway, and the numbers prove it.

About 85% of the largest SaaS companies have or are moving to pay-per-consumption models instead of fixed subscriptions.

Microsoft charges extra for each request over the limit in its Copilot products. Salesforce launched Agentforce, an agent who dialogs with a customer instead of a manager, and switched to pricing based on agent actions rather than the number of users.

But the shift to tokenized consumption isn't just a price list change. It's a blow to the very essence of why SaaS companies have always been investors' favorites.

Classic SaaS companies worked almost like a printing press: write code once and sell it to thousands of clients with a gross margin of 70-90%. Each new user cost almost nothing, but brought constant income. That is why the market valued these companies so highly.

AI breaks these mechanics. When a product runs on tokens, each agent response costs money, computing power, electricity, and GPUs. Microsoft sells GitHub Copilot for $10 a month, but at launch the real cost per user was as high as $80. The company was making a loss with each subscription. According to venture capital firm Bessemer, young AI companies today have an average gross margin of about 25% versus the usual 70%+ for classic SaaS.

This does not mean that such a business model is not viable. But investors should realize: the era when the growth of revenue for a SaaS company automatically meant growth in profits is ending.

Cheaper AI models will fix the situation, but not immediately - the pricing power is still at the beginning of the chain, where chips and data centers are located, and only in time will it shift to software and products. As the scale of the AI business grows, we will see its profitability improve over the next couple of years.

And here it is worth remembering an old economic paradox.

In the 19th century, the English economist William Jevons noticed a strange thing: after the invention of a new generation of steam engines that burned coal many times more efficiently, Britain's total coal consumption did not fall, but rose. That's because cheap energy opened up new uses that no one had thought of before.

The same thing is happening with tokens. Gartner predicts that the efficiency of large language models will increase more than 100 times by 2030. This does not mean that less computing power will be needed.

This means that the increasing efficiency and decreasing cost of AI makes these applications viable at scale, for example, in areas such as robotics, autonomous transportation, and medical diagnostics.

The token counter will spin faster, not slower. That is why for Nvidia what is happening is not a threat, but perhaps the best thing that could happen.

Training an AI model is a one-time capital investment. Do it once, spend billions, and you're done. Inference - the generation of AI responses - is a different matter. It is a constant process: every request to an agent, every token, every decision of the system requires computation. And the more agents working in the world, the more GPUs are needed continuously, every second.

Deloitte estimates that inferencing will account for two-thirds of all spending on AI computing as early as this year. Huang at GTC called this "the tipping point of inferencing" and built the entire product logic of the conference around this thesis.

In late 2025, Nvidia acquired the team and technology of startup Groq for $20 billion, the largest purchase in the company's history. Groq specialized specifically in high-speed inferencing, and now this technology is being integrated into Nvidia's product line. In parallel, the company introduced NemoClaw, an enterprise platform for the secure deployment of AI agents within enterprises. Huang made a direct analogy: what Windows has become for the personal computer, NemoClaw should become for agent systems. Partners already include Salesforce, Cisco, Google, Adobe and CrowdStrike.

The logic is simple: if every company in the world will soon be hiring digital agents, Nvidia wants to be both the factory they work in and the operating system they live under.

Who else will benefit from this transition? I would distinguish three groups.

- The first is the owners of the infrastructure.

Agents live in the cloud, which means every new agent is a new account in AWS, Azure or Google Cloud. Amazon, Microsoft and Alphabet in the AaaS world are the ground on which everyone else builds. They are joined by Nvidia with its dominance in GPU and networking hardware, without this company's hardware most agents simply won't launch.

Broadcom is on the list here, too: the company develops AI chips to replace Nvidia's GPUs for Google, Meta, and OpenAI, and makes the networking stuff that data centers don't work without.

- The second is cybersecurity.

Every AI agent that gains access to corporate systems is a new point of vulnerability. The more agents, the wider the attack surface. CrowdStrike and Cloudflare are already positioning themselves in the niche of protecting agent systems, not people.

- The third is SaaS companies that have time to transform.

There is an important caveat here: not everyone will be able to cope, and we should not expect rapid growth in quotations; the market is now looking at this sector with mistrust. But Salesforce, with its Agentforce platform and ServiceNow, which is rebuilding corporate automation for agent-based logic, look like companies that understand where the wind is blowing.

And separately Palantir: while everyone else is discussing a move to AaaS, this company is already doing just that, selling not access to the platform, but ready-made analytics solutions.

The losers are those who sell beautiful interfaces and cheap manual labor. An agent doesn't need a convenient button. It just does the work.

This article was AI-translated and verified by a human editor