AI Is Eating Software Margins: How SaaS Companies Now Have to Price In the Token Tax

For many years, software was considered the business model with the sweetest margins in the world: build the product once, sell it as often as you like, collect recurring revenue. With the spread of AI, this logic is breaking down. Every request to a language model causes real costs – for compute, inference, storage, and network. What users experience as a magical feature is, for providers, a variable cost item that grows with every interaction.

Tony Wang, Portfolio Manager at T. Rowe Price, sums up the shift pointedly in a recent analysis: “AI is not just a new product cycle, it is changing the margin structure of the tech industry.” The reason is simple: “Intelligence is no longer free. Every AI interaction requires compute, inference, memory, storage and network resources.”

How High Margins Were Before AI

To understand the scope of this shift, it’s worth looking at the economic foundation the industry has stood on for years. Classic B2B SaaS historically achieved gross margins of 80 to 90 percent – the logic behind it: once the software was built, every additional customer cost almost nothing, because marginal costs were practically limited to hosting and support. At scale, the cost of goods sold for mature SaaS providers typically came in at just 10 to 25 percent of revenue. Seven out of ten publicly listed SaaS companies reported gross margins above 70 percent, with the best in class reaching 80 percent or more (read more).

With AI, this picture is crumbling. Pure AI-first companies currently come in at gross margins of 50 to 60 percent, with inference alone eating up around 23 percent of revenue. Across the Q4 2025 and Q1 2026 earnings seasons, a new operating corridor of 60 to 70 percent gross margin has established itself among publicly listed SaaS providers that openly talk about AI-driven margin pressure – and this compression is not temporary but structural. In other words: those who deeply integrate AI into their product lose, on average, 15 to 30 margin points compared to the old SaaS standard (read more).

From Flat-Rate Subscription to Token Counter

The consequences of this new cost reality are now visible across the SaaS world. Notion, for example, has gradually overhauled its pricing model: classic AI features such as writing assistance or meeting notes are included in the Business and Enterprise plans, but for the new so-called Custom Agents – proactive AI that takes on tasks independently – Notion charges separate credits. 1,000 credits cost 10 US dollars, are shared within the workspace, and reset monthly; once used up, all agents pause until the next billing cycle or until an admin tops up. Consumption depends on how much information an agent processes, how many tools it calls, and which model is used – more powerful models burn more credits.

Notion is not an isolated case, but a symptom. In its early days, GitHub Copilot is said to have caused losses of up to 80 US dollars per user per month among power users, while only charging a flat 10 dollars. Microsoft drew the consequences and introduced usage limits as well as additional fees for heavier use. Zendesk, Intercom, and Salesforce with its Agentforce model have gone down the same path.

Wang sees in this development more than just a technical adjustment. “The industry is shifting from a world of high-margin software products to a world in which every intelligent action carries real costs,” says the portfolio manager. Investors are therefore rephrasing the question: “They are no longer just asking which companies can successfully deploy AI, but which companies can generate attractive returns from these investments.”

Pressure at the Application Layer

For providers at the application layer – the classic SaaS players – the situation is becoming increasingly uncomfortable. Those who don’t integrate AI risk being overtaken by AI-native competitors. Those who do integrate eat into their margins. “Many companies at the application layer are forced to reinvest aggressively to stay competitive,” says Wang. These investments could strengthen products in the long term, “but they can also put pressure on margins and create uncertainty about pricing power and return on capital.”

How much the margin question is weighing on SaaS providers is shown by the back-and-forth at Salesforce: within roughly 18 months, the company has introduced three different pricing models for Agentforce – per conversation, per action via Flex Credits, and most recently classic seat licenses starting at 125 US dollars per user per month. Three pricing models running in parallel for the same product – a sign that the market has not yet found consensus on how AI should actually be sold. More and more often, software providers can be seen moving away from pure per-seat pricing toward an additional consumption-based pricing model, in which – to put it simply – token costs are passed on to customers (read more).

Infrastructure as the New Bottleneck

While the application layer struggles, another layer of the AI stack sits structurally on the sunny side. “The infrastructure side of the AI ecosystem remains structurally advantaged, because every AI workflow ultimately requires more compute, memory, storage and network capacity,” says Wang. This advantage becomes particularly relevant with the transition from chatbots to agents: “A chatbot answers a question. An agent does work, reasons through tasks, retrieves information, updates plans, and acts across workflows. That requires significantly more infrastructure behind the scenes.”

This also shifts the strategic bottlenecks of the industry, according to Wang’s analysis. “In the traditional software era, distribution and ownership of the customer workflow were the scarce resources. In the AI era, the increasingly valuable position may be owning the infrastructure layer on which every intelligent workflow depends.”

Outlook: ROI Still Needs to Be Delivered

The coming quarters will show which application providers manage to translate their AI features into prices that customers accept while at the same time protecting margins. Gartner forecasts that by 2030, at least 40 percent of enterprise SaaS spending will shift to usage-, agent-, or outcome-based models. Until then, what Wang soberly notes applies: “AI creates enormous opportunities, but also introduces a fundamentally different cost structure. The companies integrating AI still have to prove the return on investment, while the infrastructure layer continues to benefit from rising demand across the entire ecosystem.”

For investors, this means: the old rule of thumb that SaaS equals premium margin no longer automatically applies. And for SaaS providers: those who give AI away for free, in case of doubt, also give away their margin.