Cloud savings have an order, and getting it wrong costs three years

Almost every cloud cost conversation starts in the wrong place. Somebody has been told that Savings Plans or committed-use discounts will take 30% off the bill, the finance team has heard the number, and the question arrives as "how quickly can we commit?"

The number is real. The sequence is what makes it expensive.

Three places, in this order

• Waste comes first. Idle instances, volumes nobody detached, oversized databases, public IPv4 addresses that have been chargeable since February 2024, snapshot policies with no expiry. On a typical mid-market estate this is 10–20%, and none of it requires an architecture change, a maintenance window, or a risk conversation.
• Commitments come second, and only second. A Savings Plan or a CUD is a promise to keep spending at a level for one or three years. If you make that promise while 20% of the estate is waste, you have promised to keep paying for the waste. The discount is real and the saving is real, and you have still made yourself worse off than if you had spent three weeks cleaning up first.
• Architecture comes third, because it takes actual design work: transfer paths, storage classes, scheduling, managed-service choices, and occasionally the honest conclusion that a workload was never suited to cloud.

What the order costs when it is reversed

Take a straightforward case. An estate spends ₹10 lakh a month. A provider proposes a three-year commitment covering 80% of it, at a 30% discount. On paper that is ₹24 lakh a year.

Now suppose a fifth of that estate is waste, the ordinary kind, not anything exotic. The commitment has been sized against ₹10 lakh of consumption, and the ₹2 lakh a month of waste is now load-bearing: switch it off and utilisation drops, and you pay for the unused commitment anyway. You have converted a fixable problem into a contractual one, for three years.

Clean up first and the same commitment gets sized against ₹8 lakh. The discount is smaller in absolute terms and you are genuinely better off, because the ₹2 lakh has gone rather than been locked in.

The unglamorous three

Almost every first pass we do turns up the same three items, and none of them is interesting:

• NAT gateways. In the Mumbai region one costs about ₹3,100 a month before a byte crosses it. The recommended one-per-zone across three zones is ₹9,400 a month, idle: ₹1.1 lakh a year. Sometimes all three are load-bearing. Often two are there because a template put them there.
• gp2 volumes never moved to gp3. Around 20% cheaper for the same disk, with a 3,000 IOPS baseline, no downtime, and one command.
• Cross-zone replication nobody designed. 500 GB a day between availability zones is roughly ₹28,600 a month, usually because a subnet was picked in a hurry two years ago.

Above roughly ₹1 lakh a month of spend, data transfer alone is usually 10–30% of the bill. It almost never appears on anyone's dashboard, because the console files it under headings nobody opens.

What to ask a provider

If someone quotes you a percentage in the first week, ask what they are committing you to and over what term. A number that arrives before anyone has looked at utilisation is a commercial proposal, not an engineering finding.

And ask what they would not change. A cost review that produces only a list of savings has not been read carefully enough.