The cloud spend patterns quietly eating mid-market margins

If you run a software business, your cloud bill is not an IT line item. It is cost of goods sold. Every dollar of waste in it comes straight out of gross margin, and at scale a few points of waste is real EBITDA. The uncomfortable part is how normal that waste is.

None of this is a few careless teams overspending. The waste is structural, and it is quiet — which is exactly why it is so recoverable once someone goes looking. I once took a single enterprise environment from chronic overspend to more than 40 percent leaner, roughly $360,000 a year, without disrupting environment performance. Here is where it tends to hide.

01 — Always on

You are paying for peak, 24 hours a day

Most workloads are provisioned for the busiest moment they will ever see, and then left running at that size around the clock. Dev and staging environments hum along at full capacity over nights and weekends when nobody is using them. Compute that could scale to demand instead sits flat at peak. This is usually the single largest line of recoverable spend. In my own work, redesigning compute and autoscaling on a steady-state workload cut its compute cost by roughly two thirds. The money was never buying performance. It was buying idle headroom.

02 — Zombies

The resources nobody turns off

Unattached storage volumes from instances deleted months ago. Idle load balancers with no healthy targets. Snapshots and old machine images retained forever. Orphaned IP addresses billing by the hour. Test environments someone spun up for a demo in 2023 and never tore down. None of it serves a customer, all of it bills monthly, and because each item is small, nobody flags it. Together they are rarely small.

03 — On-demand

You are renting at on-demand prices for things that never turn off

Cloud providers give steep discounts — Reserved Instances, Savings Plans, committed-use discounts — for committing to capacity you are obviously going to use anyway. Plenty of mid-market companies run their entire steady-state baseline at full on-demand rates because nobody owns the commitment strategy. That is frequently a 30 to 60 percent discount left on the table for workloads that demonstrably run every hour of every day.

04 — Data transfer

The data-transfer bill nobody can read

Egress charges, cross-availability-zone chatter, and NAT gateway data-processing fees are close to invisible until you go looking. A chatty architecture that moves data across zones or out to the internet more than it needs to can quietly add a meaningful percentage to the bill, and it never appears as a service you could just turn off. You have to find it in the traffic patterns.

05 — The root cause

Nobody owns the bill

The reason all of the above persists is the same in almost every case: spend is untagged and unallocated, so no team can see what their own product or feature actually costs. When cost cannot be attributed, it cannot be owned, and what is owned by everyone is owned by no one. This is also why "we'll have an engineer look at it" rarely works. Cloud economics sits permanently outside the team's core competency, it is intermittent, and pulling your best engineer off the product to learn it is a bad trade.