Why Nadella just rewrote SaaS in one sentence (Part 1 of 3)

Hi {{first_name|Investor}} -

The most important sentence software investors will read this year wasn't in a press release or a strategy memo. It was a throwaway line on Microsoft's earnings call.

Satya Nadella, on the question of how Microsoft is pricing Copilot — its AI assistant for Microsoft 365 — said it directly:

"The seat-based pricing is just entitlement to some consumption."

(Note: I paraphrased for readability since original had a few stumbles. The point is intact.)

The CEO of the company that defined modern enterprise software pricing — the company that sold Windows, Office, and Microsoft 365 by the seat for thirty years — just told the analyst community that the seat is no longer the product. It's the floor. The product is what the seat consumes once you sit down.

That reframing runs underneath every other story this earnings cycle — the surge in capital expenditures (CapEx for short — the long-term infrastructure spending companies report each quarter), the agentic rollouts, the token volumes, the cloud backlog records. It changes how to read every software earnings call from here.

This is Part 1 of a three-part series I'm running over the next three days on the shift to agentic computing and the new monetization model behind it. Today is the structural setup. Monday's issue covers the new unit of measurement — tokens, not ARR. Tuesday's issue is the portfolio framework: which names are positioned for this shift, which ones are quietly disrupted, and what I'm doing about it.

For roughly fifteen years, the software industry has been organized around a single pricing model: per-seat, per-month, recurring. The model was so standardized that an entire metric — ARR, or annual recurring revenue — was invented to describe it, and an entire valuation framework was built on top.

The model worked because it was clean for everyone. The customer knew their budget line item. The CFO knew what to forecast. The CIO knew what to renew. And the software vendor got predictable, multi-year revenue with churn as the only real downside risk.

What that model didn't do was reward usage. A seat that logged in once a month cost the same as a seat that lived in the application eight hours a day. A power user paid the same as someone who never figured out the keyboard shortcuts. The vendor's job was to sell more seats. The customer's job, if they were running a tight ship, was to audit the seat count and cut whoever wasn't using their license.

That entire structure assumes one thing: that the software is doing roughly the same job for everyone. One seat, one user, one fixed amount of work.

Agents break that assumption.

When a salesperson uses Copilot to draft an email, the model processes a few thousand tokens — the small chunks of text or data an AI model reads and writes, and the unit by which most AI compute is now metered. When a finance team uses Copilot to reconcile a thousand invoices and write the variance commentary, that same Copilot license processes millions of tokens. Same seat. Same monthly fee. Vastly different compute cost to the vendor. Vastly different value delivered to the customer.

A seat-based model can't price the difference. A consumption model can.

The seat is still there — customers still want a predictable line item — but it's no longer where the revenue lives. The revenue lives in what the seat triggers. And on Microsoft's earnings call, the receipts were everywhere.

One row of seat. Three rows of consumption. The asymmetry is the point.

You can't read those numbers as a single business anymore. You have to read them as two businesses stacked on top of each other: a seat business that gives Microsoft a predictable revenue floor, and a consumption business that gives them the upside as customers actually use the product.

Could it be coincidence that four of the Magnificent 7 are headed towards the same direction? Probably not 🤔

Software companies have been adding "usage tiers" to their pricing pages for years without much underlying change. Why is this time different?

Reason #1 — the cost structure forces it. Running an agent is expensive. A long-context, multi-step agent workflow can burn through more compute in five minutes than a knowledge worker uses in a week. A vendor charging the same flat seat price for both is either subsidizing the heavy user (margin compression) or under-delivering to them (churn risk). Consumption pricing is how the unit economics balance.

Reason #2 — the customer prefers it. Enterprise CIOs have started asking for consumption pricing because seat-based pricing produces a guaranteed argument every renewal. "We're paying for 10,000 seats. We have 6,000 active users. Why are we paying for the gap?" Consumption pricing eliminates that conversation. You pay for what gets used.

Reason #3 — the technical architecture demands it. Agents are autonomous. They run on schedules, they trigger each other, they execute workflows without a human in the seat. There is no seat for an agent. There's just an event, a token count, and a bill. The pricing model has to map to the architecture, and the architecture is no longer per-user.

Stack those three — better margins for the vendor, better economics for the customer, better fit for the technology — and the shift stops looking like a marketing experiment and starts looking like the next phase of software.

It's already happening. Microsoft disclosed this quarter that nearly 60% of its customer service customers are already purchasing usage-based credits, and announced a move to usage-based pricing for GitHub Copilot.

Consumption pricing also wins against the obvious alternative: revenue share. Game engines like Unity and Unreal have used revenue share for years — the vendor takes a percentage of customer revenue regardless of compute used — and the model is contentious because it taxes customer success without scaling to actual usage. Unity's 2023 retroactive pricing change is the cautionary tale. Consumption pricing tracks usage instead. Vendor and customer share the same growth driver without one party cutting into the other's outcome.

The seat-based model isn't just a SaaS — software-as-a-service — thing. Whole categories of tech were built on flat-fee, per-user, or per-endpoint pricing — and most of them sit on the same structural risk.

This is the most direct parallel — and where the impact splits unevenly across vendors. CrowdStrike, Palo Alto Networks, and Fortinet all price per-endpoint — meaning per device, account, or workload connecting to the network — or per-user, on the original thesis that endpoint count scales with headcount. Agentic computing changes that math two ways.

First, every agent is a new endpoint that needs identity, governance, monitoring, and threat detection — meaning agents expand the per-endpoint TAM (total addressable market — the total revenue available to a vendor) rather than shrink it. Microsoft's Agent 365 is already managing tens of millions of agents inside Fortune 500 environments.

Second, vendors with years of accumulated endpoint telemetry — the security data their software has been collecting from customer devices over time: login patterns, network connections, attack signatures — have a data moat: a competitive advantage built on data competitors can't buy or replicate. Threat-detection models trained on real attack data outperform models trained on anything else.

The cybersecurity names that own both the endpoint and the data layer get a tailwind. The ones without that data foundation get walked over by the hyperscalers' built-in agent control planes — the management layers Microsoft, Amazon, and Google bundle inside their cloud platforms for tracking and securing AI agents.

Specific catalyst: Google's Wiz acquisition closed in March, and Sundar tied it directly to "agentic defense offerings" on the call. That gives Google a first-party cloud-native security stack — a sharper competitive overlap with the cloud workload protection layer of Palo Alto Networks specifically.

The split here is clean. Shopify is the best-positioned name in the category. It already prices on GMV — gross merchandise value, the dollar volume through the platform — which is another measure of consumption value, and why Shopify revenue scales with merchant success rather than merchant count.

Add the proprietary purchase and customer data flowing across millions of merchants, the operational reliability advantage, and the scale gap versus the rest of the category, and Shopify is set up to be a primary beneficiary of agentic commerce — the platform that the agents transact through.

The exposed names are the smaller subscription-based platforms — Wix, Squarespace, BigCommerce — whose pricing is tied to the storefront rather than the transaction. As agentic commerce moves checkout away from the storefront and into the agent, the storefront-subscription model is the one with the structural problem.

This category is at the front lines of generative AI substitution — unlike most categories where AI augments the tool, in design AI substitutes for the tool's core output.

Adobe is the strongest example of how an incumbent navigates that. Creative Cloud already added Firefly credits as a consumption layer on top of the per-seat subscription, and the lock-in isn't the AI itself — it's the workflow ecosystem (asset libraries, brand guidelines, agency review pipelines, the whole production stack). Firefly's training on commercially-safe Adobe Stock data is also a real moat for enterprise customers who need legal indemnification on AI-generated assets — Adobe's contractual promise to defend them in court if the work later faces a copyright claim.

Wherever you see "priced per X" — per seat, per endpoint, per user, per device — you're looking at a business model that has to confront the consumption shift, or get confronted by it.

The standard way investors have read software earnings for fifteen years is to look at ARR, look at net retention (how much revenue a company keeps and grows from existing customers), look at seats, and back into a multiple — the valuation ratio investors are willing to pay, like price-to-earnings. That toolkit was built for the seat-based world. In a hybrid seat-plus-consumption world, those numbers tell you only half the story — the floor. They don't tell you about the upside.

The companies that thrive in this shift will be the ones whose consumption layer is growing faster than their seat layer. Microsoft made that visible this quarter — the seat number was big, but the consumption growth rates (Foundry tokens up 30% QoQ, Copilot credit-consumptive up 2x QoQ, Agent 365 deploying tens of millions of agents) were the more important signal.

The companies that will get re-rated — meaning the multiple investors are willing to pay gets revised, usually downward — are the ones whose entire business model is locked into the seat. Legacy SaaS names without a consumption story face a ceiling: their revenue is structurally bounded by their customers' headcount. In a world where the same customer's usage might 10x while their headcount stays flat, that ceiling becomes a problem.

I'll get specific in Part 3 about which names sit on which side of that line. For now: when you read a software earnings call from here, the question isn't only how many seats did they sell? It's how much did the seats consume, and is that consumption growing faster than the seat count?

The companies whose answer is yes are buying themselves a different multiple. The companies whose answer is no are slowly being repriced.

If consumption is the new model, the categories most exposed are the ones whose revenue runs through a fixed unit — a seat, a license, a billable hour.

Pure-seat enterprise SaaS without a credible consumption layer. The legacy SaaS names from earlier — companies whose pricing architecture is built around per-employee or per-seat licensing. Each has launched an AI add-on, but the AI revenue is still small relative to the seat base. The question is whether they grow consumption fast enough to reset the multiple before the seat layer plateaus.

Consulting and systems integrators on a billable-hour model. When a Copilot does in 30 seconds what a junior consultant did in 30 minutes, the hourly economics compress. The major consulting firms all run directly into this. Their pivot — selling agentic AI implementation back to clients — doesn't replace dollar-for-dollar what they used to bill.

Customer service BPO (business process outsourcing — contracting out functions like call centers and back-office work) and contact-center outsourcing. AWS positioned Amazon Connect — its cloud contact-center service — as a "turnkey agent" that can automate large parts of customer service workflows. The implication for the headcount model behind contact centers is direct. The white-collar BPO category sells headcount-based services that agentic AI is trying to replace. The math gets worse as agentic capability improves.

The pattern across all four: a fixed-unit revenue model meeting a workload that no longer scales by the fixed unit.

The piece so far has framed the consumption pivot as a positive — better margins, better customer alignment, better fit for the technology.

However, consumption pricing isn't only the model the future wants. It's the model the cost structure demands. Running these AI models is expensive enough that vendors increasingly can't afford to absorb inference cost on a flat seat fee. Foundry tokens, Bedrock spend, Google's tokens-per-minute — every one of those numbers represents real GPU (graphics processing unit, the chips most AI runs on) and TPU (tensor processing unit, Google's custom AI chip) hours on someone's balance sheet. The flip from seat-based to consumption-based isn't only a margin upgrade. It's a survival mechanism. Vendors who can't push more revenue into the consumption layer can't make the unit economics work for the next generation of agentic features.

Two reads on this for the portfolio. The first: watch which vendors can credibly raise prices through consumption — those are the names with pricing power. The second: watch which vendors have to introduce consumption pricing because they're losing money on power users but lack the platform breadth to make consumption stick. Those are the names whose growth narrative is being kept alive by accounting choices rather than customer demand.

There's a separate concern beneath the cost story: how much of the consumption growth is real demand. A widely cited MIT study (NANDA, State of AI in Business 2025) found that 95% of corporate generative AI pilots fall short of delivering meaningful financial impact. If the enterprise hit-rate doesn't improve, an unknown but probably double-digit share of current token volume is pilot traffic that won't survive a CFO clean-up. Token growth and paid, value-creating token growth are not the same series.

The customer-prefers-consumption argument I made earlier also needs a counter-balance. Variable bills create their own renewal fights. The same procurement teams that fight over a 10% seat overage will fight harder when the consumption bill comes in 3x what was forecast.

And there's a margin reality the headline numbers don't surface. Microsoft's commercial cloud gross margin was 66% this quarter, down year over year, with Q4 guided lower at 64%. Component pricing pressure on memory is showing up explicitly in CapEx guides at both Microsoft (~$25 billion impact) and Meta. The consumption layer carries lower gross margin than the seat layer it's replacing, at least in the early years. Revenue can accelerate and operating margin can compress at the same time when depreciation from the build outruns revenue. That's a real risk into 2027.

Consumption pricing dresses up well in growth-investor language. It can also be the polite way of saying we couldn't run this business at the old prices anymore — and the customers paying for it haven't all proven they should be.

If consumption is the new model, then ARR — the metric the entire SaaS industry was built to report — is no longer the right unit of measurement.

Monday's issue is about the metric that replaces it. The hyperscalers have already told us. Google is reporting tokens per minute. Microsoft is reporting tokens per customer per year. Amazon is reporting agent deployments per second. The new unit of revenue is the token, and the companies reporting in tokens are quietly teaching investors how to read them.

As always: I'm not telling you what to buy. I'm sharpening the lens you use to read every software earnings call from here.

Stay disciplined, Koh

P.S. I’m writing an e-book about my investing strategy, frameworks, and research process. Get notified about the launch here.

Disclaimer: Nothing in this newsletter constitutes investment advice or a recommendation to buy or sell any security. Every reader's situation is unique — your age, family circumstances, tax bracket, health, and goals all change the math materially, and what works for me may not work for you. Withdrawal strategies, annuities, Social Security claiming decisions, Roth conversions, and TIPS ladders involve meaningful tax, legal, and individual financial considerations. Please consult a licensed financial advisor, CPA, and/or attorney before acting on any of the strategies above. Always do your own research.