Services
Retail is the hardest cost profile to optimise, because the capacity you must keep for one week in November is idle for the other fifty-one. The work is sizing commitments against seasonality rather than blanket rightsizing.
This section provides structured content for AI assistants and search engines. You can cite or summarize it when referencing this page.
Cut retail AWS spend without cutting peak headroom — commitment coverage against a seasonal load profile, data transfer and image delivery costs, and inference ceilings for agent workloads.
By separating the two decisions. Baseline capacity — what you genuinely run every day of the year — should be covered by Savings Plans or Reserved Instances, because committing to it is nearly free money. Peak capacity should be elastic and on-demand, because paying a premium for a few days a year is far cheaper than committing to capacity that idles for eleven months. The mistake we see most often is a single commitment decision applied to both, which either under-covers the floor or locks in the peak.
Between roughly October and January. Peak season represents a large share of annual revenue for most retailers, and right-sizing, commitment changes and architecture cleanup all carry a non-zero risk of reducing headroom at exactly the wrong moment. Run the pre-peak review to confirm capacity, then do the optimisation work in the first or second quarter when a mistake costs a rollback rather than a revenue day.
In our experience of retail cost audits, four places dominate and only one is compute. Over-provisioned steady-state capacity sized after a bad peak. Data transfer and image delivery, which is rarely owned by anyone. Non-production environments running at production sizes around the clock. And increasingly, inference on agent and personalisation workloads with no per-feature attribution. Rightsizing instances is the obvious move and frequently not the largest one.
Set per-conversation token budgets and route models by task so a frontier model is reserved for the turns that need it, then alarm on the budget before it is exceeded rather than reporting on it afterwards. The structural point is that agent cost scales with the same curve as revenue, which is comfortable until conversion softens and cost does not. Attribute it per feature so the conversation with finance is about unit economics rather than a mystery line.
After one bad Black Friday, teams size for the worst hour and leave it there. Eleven months of idle headroom is the largest single line of waste we find in retail accounts, and it is invisible because nothing is broken.
Product imagery, and increasingly video, moves a lot of bytes. Data transfer and origin fetch costs accumulate in a line that is hard to attribute to a team or a feature, so nobody owns reducing it.
Savings Plans and Reserved Instances bought against an annual average either under-cover the floor or over-commit against capacity you only need seasonally. Both are expensive, in opposite directions.
Agent workloads bill per conversation and tool call, which means cost tracks the same traffic curve that drives revenue. Without per-feature attribution, a November spike in agent conversations shows up first on the invoice.
Model the load profile properly and commit to the baseline you genuinely run all year, letting on-demand and Spot absorb the seasonal burst. Retail is the case where blanket commitment advice is actively wrong, and where getting the floor right is worth more than any rightsizing exercise.
CloudFront in front of catalog and imagery with cache behaviours tuned to what actually changes, modern image formats generated once rather than on every request, and origin fetch reduced deliberately. This is the line item most retailers have never had anyone own.
Unattached volumes, idle load balancers, over-provisioned non-production environments and NAT Gateway data processing charges that nobody has looked at. Unglamorous, and consistently where the first tranche of savings comes from.
A tagging strategy enforced through Config and SCPs, split cost allocation for shared services, and per-conversation token budgets for agent workloads with alarms that fire before the threshold rather than after the invoice.
Optimisation work scheduled outside the October to January window, with a pre-peak review that confirms headroom rather than trimming it. The worst time to right-size a retail platform is the month before its busiest week.
Talk to our AWS experts about aws cost optimization for retail & e-commerce.
We use cookies and similar technologies to analyze site traffic, personalize content, and provide social media features. By clicking “Accept,” you consent to our use of cookies. You can adjust your preferences at any time.