SaaS and the hidden cost of AI infrastructure

Your software bill went up again this year and nobody on the team can really explain why. Same plan. Same number of users. Except now there is an "AI credits" line in the middle of it. That new line did not come out of nowhere: it came from warehouses full of graphics cards burning power somewhere in the United States. And the recent news is that those warehouses just got a faster path to being built. The EPA, the American environmental agency, published permitting guidance to speed up approval of power generation dedicated to data centers. Translated into your SaaS: more capacity, faster, with less regulatory friction.

This sounds far from your company. It is not. The cost of your AI SaaS is born there.

Why an American environmental rule shows up on your bill

An AI product's cost has three layers. The first is the software itself, which is cheap to copy. The second is the graphics card, which is expensive and contested. The third, the one almost nobody looks at, is power and cooling.

A modern AI data center consumes somewhere between tens and hundreds of megawatts. A large shopping mall consumes about 3 MW. In other words: each one of those buildings pulls the equivalent of dozens of malls running 24 hours a day. This is not an operational detail, it is the main bottleneck in the sector today. There is no shortage of money to build. There is a shortage of grid connections and permits.

When permitting loosens, construction accelerates. When construction accelerates, capacity supply grows. And then two things happen at the same time, in opposite directions:

• The cost per AI token falls, because there are more machines available.
• The pressure for you to consume more grows, because whoever built it needs to fill the warehouse.

The second part is what bites your budget.

The pricing model changed and almost nobody noticed

For fifteen years SaaS had a simple mental contract: you pay per seat, per month, and use it freely. Predictable. Easy to get approved by finance.

AI broke that. Now there is consumption in the mix. Credits, tokens, "runs", "tasks", "agent actions". Every vendor invented their own name. The practical effect is the same: part of your bill became variable and depends on how much your team uses.

I have seen a company with 12 seats on a tool get a bill spike because an intern found the "generate report with AI" button and generated 400 reports in a week. Nobody did anything wrong. The contract changed in nature and the internal process did not keep up.

A fixed subscription you approve once. Variable consumption you have to watch every month.

How much this really weighs

Numbers I typically see in 20 to 200 person companies in Brazil:

• 40 to 90 active subscriptions, of which leadership knows about 15.
• 25% to 40% of paid seats unused in the last 60 days.
• AI consumption lines that doubled between 2025 and 2026 on the same tools.

In a 60 person company spending R$ 40k per month on software, a serious cleanup usually gives back R$ 8k to R$ 14k monthly. It is not magic or aggressive negotiation. It is just stopping payment for things nobody opens.

The cruel detail: the variable AI portion grows exactly where the tool is useful. So cutting blindly is dumb. You cut what works and keep the CRM three people use out of habit.

What data center expansion means for your next contracts

The easy read is "everything gets cheaper, great". The correct read is more annoying.

The unit price of inference has been falling consistently. A task that cost X in 2024 costs a fraction of that today. But company bills went up in the same period. The reason is the usual one in technology: when it gets cheap, you use much more. It is the Jevons paradox applied to your corporate card.

So here is what changes with more capacity built:

• Bargaining power improves. With more supply, an annual contract with a consumption cap becomes negotiable. Ask for it.
• Lock-in gets more dangerous. A vendor with its own infrastructure will push you deeper into its ecosystem. Migrating later is expensive.
• Portability becomes a requirement, not a luxury. If your data and prompts are trapped in the vendor's format, you have no alternative when the price goes up.

One question to bring to your next renewal: "what does it cost to leave in 12 months?". If nobody on your team can answer, you are already exposed.

A practical way to sort this out in two weeks

You do not need a six month consulting project. You need discipline for ten business days.

Week 1, inventory. Pull the last 12 months of corporate card statements and filter for recurring charges. Tools you swore you cancelled will show up. Build a spreadsheet with name, internal owner, monthly cost, billing type (fixed or consumption) and renewal date. That alone scares most managers.

Week 1, real usage. In each tool, export the list of users active in the last 60 days. Compare against paid seats. The difference is money in the trash.

Week 2, decision. Three piles: keep, cut, renegotiate. Cut dead seats the same day. For consumption tools, turn on spend alerts if the tool offers them, and set a monthly cap per team.

Week 2, light governance. One rule is enough: every new subscription needs a named owner and a review date. No owner, no purchase. That keeps the problem from coming back in six months.

Automating this watch is simple and worth it. A script that reads the statement, cross references the spreadsheet and sends a summary on WhatsApp or email on the 5th of every month does the job. It takes about two hours to build and saves you the same argument every quarter.

The most common mistake: treating AI SaaS as an IT expense

SaaS with AI is not an IT line. It is an operations line. The difference matters when you decide.

An IT expense you compare to last year and try to reduce. An operations expense you compare to the result it produces. If a R$ 3k per month tool saves the sales team 40 hours, it is cheap. If a R$ 300 one has not been opened in two months, it is expensive.

Ask the right question: how much does this tool give me back in hours or in revenue? Whoever does not measure that ends up cutting what worked and keeping what only makes noise.

And there is the environmental point, which nobody wants to touch but is on the table. If permitting rules loosen to speed up construction, part of the real cost of AI leaves your bill and lands somewhere else: the local power grid, water for cooling, emissions. You do not pay that on the invoice. Someone pays. If your company has a public sustainability commitment, it is worth at least knowing where what you buy actually runs. Some vendors publish that information. Most do not. Asking already creates pressure.

What I would do if it were your company

I would pick the three most expensive tools on the list and look calmly at what each one delivers. Cancel whatever is idle. Negotiate a consumption cap on the AI ones before renewal, taking advantage of growing capacity supply. And set up an automatic monthly report so you never have to do card statement archaeology again.

The rest is noise. New data center in the United States, environmental rule change, vendor fights: that is context, not a task. The task is knowing what you pay, to whom, and what you get back.

If you looked at your software bill this month and felt that discomfort of not understanding half the lines, that is the signal. This can be solved with organization and a bit of automation, without switching systems and without a giant project.

See how I work and reach out for a conversation.

Your software bill went up again and nobody on the team can explain why.

Same plan. Same number of users. But there is a new line item: "AI credits".

For 15 years SaaS was predictable: you pay per seat and use it freely. AI broke that contract. Now part of the bill is variable and depends on how much your team uses.

I have seen a company with 12 seats get hit with a bill spike because someone found the generate-report-with-AI button and ran it 400 times in a week. Nobody did anything wrong. The billing model changed and the internal process did not keep up.

In the 20 to 200 person companies I work with, 25% to 40% of paid seats have not been opened in 60 days. In an operation spending R$ 40k per month on software, a serious cleanup gives back R$ 8k to R$ 14k monthly.

This is not a six month consulting project. It is two weeks of inventory, cuts, and an automatic report that lands on the 5th of every month.

If you opened your bill this month and did not understand half the lines, that is the signal. Reach out and I will show you where to start.

#SaaS #ArtificialIntelligence #CostManagement #Automation #Technology