Written by: Nicole Ogonowska, IT Growth Manager, Digital Colliers
AWS just reported 19% year-over-year growth in Q3 2026. AI infrastructure demand did most of the heavy lifting. If you run an eCommerce brand on AWS, you should care about that headline. Here is the uncomfortable reason: you probably cannot tell me what percentage of your monthly AWS bill is AI services versus baseline infrastructure. That gap matters more every quarter.
Why the AWS Q3 Numbers Matter for Retail
AWS revenue growth accelerated through 2026 because retailers and other enterprises are spinning up AI experiments at pace. The problem is not the experiments themselves. The problem is that most teams treat AI workloads as part of the general cloud budget. SageMaker inference, Bedrock calls, and Rekognition usage all flow through the same cost center as EC2 instances and S3 storage. Finance sees one number. Engineering sees one dashboard. Nobody can isolate what the AI layer actually costs.
Say your personalization engine runs on Bedrock. Your recommendation service uses SageMaker. You need to know what each costs per session or per conversion. Without that clarity, you are flying blind.
The Cost Attribution Gap
Here is the pattern I keep seeing. A retail brand launches three AI pilots: product recommendation, dynamic bundling, and chatbot support. All three run on AWS. Engineering sets up the infrastructure, tags everything as "production," and moves on. Six months later, the CFO asks which pilot is worth scaling. Nobody can answer. Cost data lives in one line item split across twelve AWS services.
The gap shows up in three places. First, baseline infrastructure costs and AI experiment costs share the same tags. Second, AI workloads often scale non-linearly, so a small traffic spike can triple inference costs without warning. Third, depreciation schedules assume stable capacity usage, but AI experiments spike and die based on test velocity.
Brands that close this gap use separate cost allocation tags for AI services from day one. That means tagging SageMaker endpoints by pilot name. Tag Bedrock usage by feature. Tag any AI-adjacent storage separately from product data. The goal is not accounting precision for its own sake. The goal is to know which experiments earn their keep and which need to die.
What Good Looks Like
Good tagging hygiene starts before the first PoC goes live. Every AI service resource gets three tags: project name, owner, and expected end date. If a pilot has no end date, it is not a pilot. It is production work, and it should live in the baseline budget.
Operators who ship this well also set up a monthly allocation review. Finance and engineering sit down with a spreadsheet. It breaks out AI service costs per project and per feature. Where possible, it ties costs to business outcomes. The spreadsheet answers one question: what would we stop paying for if budget got cut 20% tomorrow?
The FinOps best practice here is to separate capacity charges from workload charges. Your baseline infrastructure keeps the lights on. Your AI experiments test hypotheses. The two should never share a budget line. When they do, you subsidize failed experiments with money that should scale winners.
The Cost of Not Closing the Gap
The industry data here is not encouraging. More than 80% of AI projects fail, and 88% of AI proof-of-concepts never reach widescale deployment. For every 33 AI POCs a company launches, only four graduate to production. Those numbers do not improve when teams cannot see what they are spending per experiment.
The cost of inaction is not just wasted AWS spend. It is opportunity cost. If you cannot tell which AI pilots are worth scaling, you will keep funding the losers. You will starve the winners. You will also miss the moment when a successful pilot needs more budget. The signal is buried in noise.
95% of enterprise GenAI pilots deliver zero measurable P&L impact. That stat should scare any operator who cannot tie AI spending to specific outcomes. If your team cannot quickly answer what AI spending returned in Q3, you are probably in that 95%.
AWS will keep growing on AI infrastructure demand. Retail brands will keep launching experiments. The ones who win will close the cost attribution gap before their CFO forces them to.