Cloud Infrastructure Costs: A CFO Explains Where the Money Goes
Archi4 min read
Contents
- Cloud Infrastructure Costs: A CFO Explains Where the Money Goes
- The Three Line Items That Dominate
- Egress Fees and the Lock-In Tax
- Idle Capacity and the 3 A.M. Bill
- The Headcount Nobody Budgets For
- What Finance Wants From Engineering
- The Board Conversation
- Advice for Companies Signing Their First Big Commitment
- Where the Money Is Going Next
Cloud Infrastructure Costs: A CFO Explains Where the Money Goes#
Cloud infrastructure spending reached levels that would have seemed absurd a decade ago. Public cloud bills now rival payroll at many software companies, and finance teams are asking harder questions about what those invoices actually buy. We spoke with a chief financial officer at a mid-sized SaaS company about the economics of running workloads on someone else's computers. The conversation has been edited for length and clarity.
The Three Line Items That Dominate#
What surprises boards most when they see a cloud bill for the first time?
Compute gets the most attention, but it is rarely the problem. The three that dominate are committed contracts, egress, and what I call the orphan tier: storage and services nobody owns anymore. Committed use discounts are the big one. You sign a multi-year agreement to get a lower rate, and then your usage shifts. Now you are paying for capacity you reserved but no longer need. That is a fixed cost dressed up as a variable one.
How big a share of spend is that?
In our case, committed contracts account for roughly 40 percent of the annual bill. Egress runs 12 to 15 percent. The orphan tier, storage buckets and services with no clear owner, is about 8 percent. That last bucket is the one that makes auditors unhappy, because nobody can explain it.
Egress Fees and the Lock-In Tax#
Egress gets called a lock-in tax. Fair?
Fair enough. Moving data out costs money, and the price is set by the provider. That changes the math on any migration. If moving a petabyte out costs six figures, the business case for switching clouds gets harder to make. Finance people call that a switching cost. Engineers call it a trap. Both are right.
Does that show up in planning?
It should, and often doesn't. We now require every architecture proposal to include an exit estimate. What would it cost to move this workload to another provider in three years? If the number is large, the design gets a second look. That single question has changed more decisions than any cost optimization tool we bought.
Idle Capacity and the 3 A.M. Bill#
What about idle capacity?
Idle capacity is the quiet cost. Instances that run around the clock but do work for a few hours. Auto-scaling helps, but it only works if the application is built for it. Retrofitting autoscaling onto a monolith is expensive engineering work, and the savings show up 18 months later. Finance has to be patient, and finance is rarely patient.
How do you measure it?
We track utilization by team and by service. Anything under 20 percent average utilization gets flagged. That number is not a rule, just a trigger for a conversation. The conversation usually ends with someone admitting the service was a proof of concept that shipped.
The Headcount Nobody Budgets For#
Is there a people cost to cloud spending?
Yes, and it is the most underbudgeted line. We have four full-time engineers on cost and capacity work. Their fully loaded cost is roughly $700,000 a year. That is before any tools. The cloud bill is the visible number. The team managing it is the invisible one.
Does that team pay for itself?
It did in year one, mostly through commitment planning and cleaning up the orphan tier. After that, the returns flatten. The honest answer is that cost engineering is a permanent function now, not a one-time project.
What Finance Wants From Engineering#
What do you ask of engineering leaders?
Three things. First, a forecast that is within 15 percent of actuals, quarterly. Second, an owner for every service, with a name attached. Third, an exit estimate for anything that touches more than one provider. None of these are technical requests. They are accounting requests, and they change how infrastructure gets designed.
The Board Conversation#
How has the board conversation changed?
Five years ago, cloud spend was a growth story. Now it is a margin story. Boards ask what percentage of revenue goes to infrastructure and how that compares with peers. The answer is usually 15 to 25 percent for a SaaS company at our stage. Investors want to see that number trend down as you scale.
Is that realistic?
It can be, but not through discounts alone. The gains come from architecture: fewer services, better utilization, less data movement. Every one of those is an engineering decision. Finance can set the target. Finance cannot hit it.
Advice for Companies Signing Their First Big Commitment#
What would you tell a company about to sign its first multi-year cloud agreement?
Model two scenarios: usage grows 50 percent and usage grows 10 percent. If the second scenario makes the contract painful, you are signing too big. Start smaller than the sales team suggests. You can always commit more later, usually at a similar rate. The opposite is much harder to undo.
Anything else?
The invoice is a design document. Every line item reflects a decision somebody made, often years ago. Reading it carefully is the cheapest audit available.
Where the Money Is Going Next#
What is the next cost pressure?
AI workloads. Training and inference are expensive, and the billing models are less predictable than traditional compute. We are building usage caps and per-team budgets before the spend arrives, not after. The companies that get this right will treat AI capacity like any other line item with an owner and a forecast. The ones that don't will discover the cost at the same time as their investors.
More articles