The US software sector has been buzzing lately.

When Snowflake’s earnings dropped, markets seemed to exhale. Not the “AI will disrupt everything” kind of excitement — something more grounded: turns out SaaS isn’t dead, software companies can still grow in the AI era, and investors still care about predictable revenue, margins, and cash flow.

For the past six months, one question has been hanging over SaaS stocks: if AI can do the work directly, does traditional software still have value? More bluntly — if Agents become the new interface, does SaaS degrade from an operating system to a database? Snowflake gave the market a lifeline. Product revenue kept growing fast, full-year guidance went up, and AI demand didn’t gut the business model. It actually made markets believe again that data infrastructure is still a core asset in the AI era.

But the real takeaway here isn’t “AI is bullish for SaaS again.” The real signal is: markets are willing to value SaaS again — but only if you can prove AI isn’t a margin-destroying black hole.

SaaS stocks are undergoing a value reset. Not back to the pre-AI era — back to the most basic business logic: the value of a software company isn’t determined by how many tokens it consumed. It’s determined by how much cash it made.

What SaaS Used to Sell Was Near-Zero Marginal Cost

The most attractive thing about SaaS was never a polished interface or the subscription model itself.

It was this: selling the same software to your 100,000th customer costs almost nothing more than selling it to your first.

That’s why traditional SaaS commanded high valuations.

Write the code once, deploy the service once, then sell it to more and more people. There are server costs, support costs, sales costs — but the core product’s marginal cost is tiny. The bigger you scale, the better the gross margin, the more comfortable the cash flow.

Capital markets love this model.

Because it prints money.

Early losses can be explained as customer acquisition. Sales expenses can be framed as growth investment. R&D can be called a moat. As long as ARR keeps climbing and NRR doesn’t collapse, investors are willing to believe profits will show up at some point in the future.

That logic worked in the past.

Then AI arrived.

AI didn’t just add a feature to SaaS. AI stabbed the most attractive part of software companies — low marginal cost — right in the chest.

Tokens Are the New COGS

In the old SaaS model, serving one more customer barely moved your cost structure.

Now every Agent run, every generated report, every code scan, every data analysis request burns tokens.

And tokens aren’t magic.

Tokens are bills.

Before, with traditional software: the more users engaged, the healthier the margins. Now, with AI features: the more users engage, the more real the cost becomes.

That’s the awkward corner AI SaaS has backed itself into.

Users think they’re buying intelligence. Vendors are paying for inference. Users feel “wow, it can actually do things.” Finance sees “holy hell, the API bill exploded again this month.”

This is why I keep saying: a lot of companies right now aren’t doing AI transformation — they’re rewriting their own P&L into a model provider’s revenue statement.

The CEO thinks they’re buying productivity.

The CFO sees a new cost center.

The Story Model Providers Sold Is Becoming Enterprise Invoices

Over the past year, the best storytellers weren’t SaaS companies. They were model companies.

Anthropic, OpenAI, Google — each outdoing the last.

They talked about general intelligence, Agent workforces, AI employees, software engineers being replaced, white-collar work being automated. Every phrase was designed to make executives’ blood run hot.

But inside actual companies, the story takes a different shape.

Picture this: a CEO comes back from a product launch and decides: “We’re going all-in on Agents.”

The team starts hooking up APIs, buying seats, integrating internal systems, letting Agents write code, test code, write docs, run tests, do analysis. A month later, the invoice arrives.

Business outcomes? About the same.

Productivity? About the same.

Processes? About the same.

The only thing that grew for certain: token consumption.

Then comes the awkward moment.

Say AI isn’t working? The CEO wonders if you don’t know how to use it. Say it’s working? Finance asks where the ROI is. Say wait longer? The invoice doesn’t wait.

Model companies love this world.

Because whether or not you improve efficiency, the moment you start experimenting, they’re already billing you.

Enterprises pay for certainty. Model companies sell possibility.

In the short term, this looks like an innovation budget. In the long term, it’s vendor margin transfer.

What SaaS companies fear most isn’t that AI is too weak. It’s that AI is strong enough but not cheap enough.

Markets Have Stopped Listening to Stories

In 2021, fast growth was all a SaaS company needed. Losses were fine.

The logic was simple: grab land now, figure out profits later.

By 2026, that pitch doesn’t land anymore.

Investors are back to scrutinizing Rule of 40, free cash flow, gross margin. Not because they suddenly got conservative — but because AI made “profits will come eventually” sound a lot less inevitable.

Before, losses meant you were expanding.

Now, losses might mean every user click is costing you money.

Those are two very different kinds of losses.

The first is investment.

The second is COGS.

Markets will value investment. They won’t indefinitely absorb runaway COGS.

That’s why SaaS pricing logic is shifting: it used to be about growth; now it’s about how much is left after the growth. It used to be about the AI roadmap; now it’s about whether the AI roadmap crushes the margin.

AI can’t just add revenue. It has to prove it won’t eat the software economics model.

The Market Will Split SaaS Into Two Groups

Going forward, the market will split SaaS companies into two groups.

The first group treats AI as a new growth story.

They’ll spend earnings calls raving about Agents, copilots, automation, AI-native workflows. Every sentence sounds right — until you ask the follow-up: “how much has this actually improved retention, ARPU, and margin?” The room goes quiet.

The second group treats AI as a cost structure problem.

They ask first: which requests actually need a frontier model? Which can use a smaller model? What can be cached? What doesn’t need an LLM at all? What should be handled by rules, search indexes, static analysis, and traditional ML first?

The first group is selling stories.

The second group is running the math.

The valuation gap between these two groups will probably start opening up right here.

Because capital markets can absorb AI investment. What they won’t indefinitely absorb is “I can’t explain why it costs this much, but everyone’s using it.”

Especially as users start budget controls, as CFOs start demanding unit economics on every AI feature, as procurement starts comparing token prices across models — the era of burning tokens wildly will start its countdown.

AI is allowed to cost money.

But you need to burn it into something.

Into revenue. Into retention. Into efficiency. Into margins you can explain.

If you can’t — that’s not strategic investment. That’s a financial leak.

Real AI SaaS Doesn’t Treat LLMs Like a Database

A lot of companies right now are using AI with a blunt instrument.

Don’t know how to handle it? Throw it at the LLM.

Code understanding? LLM.

Log analysis? LLM.

Test generation? LLM.

User questions? LLM.

It looks smart on the surface. In practice, it’s replacing databases, search engines, rule systems, compilers, static analyzers, and workflow engines with one giant token incinerator.

That’s not AI-native.

That’s lazy.

Real AI-native SaaS redesigns system boundaries.

If it can be structured, structure it first. If it can be indexed, index it first. If a rule can solve it, don’t use an LLM. If a small model can handle it, don’t use a frontier model. If it can be computed offline, don’t burn online inference. If it can be cached, don’t re-run the inference.

LLMs should be the cognitive layer. Not the trash can.

Dumping everything into an LLM is like routing all computation through database stored procedures. Fast at first. Eventually a swamp.

AI’s value isn’t in letting software companies skip designing systems. It’s in forcing software companies to redesign them.

Value Reset Isn’t the End of AI. It’s AI Growing Up.

When people hear “the era of burning tokens is ending,” they assume it’s a knock on AI.

It isn’t.

It’s the opposite. This is the beginning of AI graduating from toy to industrial tool.

Toys get judged on results.

Industrial tools get judged on costs.

With a toy you say “look how smart it is.”

With an industrial tool you ask “what does each task cost?”

That’s why the SaaS value reset is a good thing. It pushes out companies that can only tell AI stories, and keeps the ones that actually understand software economics.

For the past two years, everyone got too easily fooled by demos.

A demo can make you feel like the future has arrived. But demos don’t run P&L. Demos don’t have SLAs. Demos don’t face the endless edge cases of real users. Demos don’t receive a six-figure invoice at the end of the month.

The real world doesn’t reward demos.

The real world rewards unit economics.

Who Pays for the Tokens

Every SaaS company is about to face the same problem: how do you pass through the AI cost?

Charge by seat — users go wild, vendors hurt.

Charge by usage — users start to hurt.

Bundle it into the enterprise plan — easy to sell, ugly on the financials.

Sell it as a separate AI add-on — customers ask: “is this AI actually worth it?”

That’s the real inflection point.

When AI features shift from “a nice surprise included for users” to “a line item that needs to justify its own ROI,” the market won’t price SaaS the way it did in 2021.

The era of burning tokens wildly won’t end with a crash.

It’ll end the boring way: budget approvals slow down, procurement starts negotiating, CFOs want reports, boards ask about ROI, investors stare at gross margins.

No dramatic pop.

Just the quiet sound of invoices landing.

So the next round of competition in SaaS isn’t about who shouts “AI-native” louder. It’s about who can turn AI into a normal business.

Models can keep improving. Token prices can keep falling. Agents can keep getting stronger.

But the market has already started asking the most basic question:

Every token you burned — whose cash flow did it eventually become?