When a bootstrapped, famously frugal, profitable software company tells its staff there will be no pay rise this year because of AI costs, founders should pay attention. That is what Sridhar Vembu said on 13 October 2026. Zoho has paused its usual annual increments, cut its marketing budget and kept headcount flat for 18 months, and he pinned a large part of it on what it now costs to run AI.
You can read this as Zoho's problem. It is more useful to read it as a preview of a cost structure that every software founder building with AI will have to plan for.
What Vembu actually said
According to the PTI report carried by The Federal, Vembu posted on X that "we cannot have our usual pay increase this year," and apologised for not being able to avoid it. He listed three pressures:
"A massive 4-fold increase in server prices which we need for data centres."
"AI usage increasing rapidly, leading to much higher expenses, which we are trying to control."
A "challenging geopolitical backdrop."
He added that "the outlook for software jobs in general is grim," that Zoho would work hard to avoid mass layoffs, and that "the next 3-5 years will decide how our reinvention has worked."
There was one more detail, reported by People Matters from an internal communication first covered by ET Now: concern that some customers are replacing Zoho products with tools they build themselves using AI-assisted coding. That impact was not quantified, and the server-price figure is Vembu's own; neither has been independently verified. But taken together, the message is clear. Costs are rising on one side, and pricing power is under pressure on the other.
Why AI changes software's cost structure
Classic SaaS had a beautiful property. Once the software was built, serving one more customer cost very little. That is where software's famously high gross margins came from, and it is why investors were happy to fund years of losses: the marginal economics were already great.
AI features break that property. Every time a user asks a model to summarise, draft or analyse something, you pay for compute. The cost scales with usage, not with the number of seats. And heavy users, the ones who love your product most, cost the most to serve.
Bessemer Venture Partners' State of AI 2025 report put numbers on how different the economics can look. Its fastest-growing AI "Supernovas" averaged around 25% gross margins, "often negative," trading profit for distribution. Its steadier "Shooting Stars" averaged about 60%, which it described as "slightly lower than SaaS peers." Those are averages across a sample of US-centric startups, not rules. But even the steadier group sits below its SaaS peers, and 25% barely looks like software at all.
Zoho's case adds a second layer. Vembu's complaint is about the servers it needs for its own data centres. A company that buys its own hardware can control costs well when prices are stable. When the price of the servers themselves jumps, it takes the hit directly, up front.
Three questions to ask about your own startup
None of this means "do not build with AI." It means AI costs need to be designed into the business model from the first pricing page, not discovered in month 18.
1. What is your cost to serve your heaviest 10% of users?
Averages hide the problem. A product priced per seat with unlimited AI usage can be profitable on the median user and lose money on the power user. Pull your usage logs, sort by compute cost per account, and look at the top decile. If that group is unprofitable, you have a pricing problem, not a growth problem.
2. Does your price move when your costs move?
Seat-based pricing was built for a world where cost did not depend on usage. If your costs now do, consider whether some of your price should too: usage tiers, credits, fair-use limits, or a premium plan for the heaviest workflows. The trade-off is real. Usage pricing is harder to sell and makes revenue less predictable. But a flat price on a variable cost is a bet that your users will stay light, and the best ones will not.
3. What happens if your customer can build a rough version themselves?
This is the uncomfortable part of the Zoho signal. If AI coding tools let a customer's internal team stitch together a "good enough" version of a simple tool, then thin products with little data, workflow depth or integration value are exposed. The defence is not cleverer prompts. It is the boring stuff that is hard to replicate: deep integrations, compliance, data the customer cannot easily move, and support they trust.
The hiring lesson hidden in the headline
The detail most people skimmed past is that Zoho has held headcount flat for 18 months. That is how a company avoids layoffs: it stops adding people early, long before the pressure becomes a crisis. A freeze on raises is painful, but it is a smaller, reversible cut than letting people go.
For an early-stage founder, the equivalent is simple. Before you hire on the strength of a good quarter, model what happens to your margins if your AI bill doubles while your prices stay put. If the answer is "we would need to cut people," hire slower, or fix pricing first.
What to take from it
Zoho is one company, with its own choices, including running its own hardware. Its numbers may not match yours. But the direction is worth taking seriously: AI makes software more capable and more expensive to run at the same time, and it lowers the barrier for customers to build simple tools themselves.
The founders who handle this well will be the ones who know their cost to serve by customer, price for usage where it matters, and build products that are hard to replace with a weekend of AI-assisted coding. That is less exciting than a demo. It is also what keeps a company out of a "please pray for us" post.
