Cloud cost is an architecture outcome, not a procurement one

When cloud bills grow faster than the business, the response is often commercial — negotiate discounts, buy commitments, run a cost-cutting exercise. These help, but they operate on a base that architecture already fixed.

The larger drivers are usually structural: workloads lifted without redesign so they still assume always-on capacity, data transferred across zones and regions unnecessarily, environments that nobody can safely turn off because nobody is certain what depends on them, and storage tiers chosen once and never revisited.

You cannot manage what you cannot attribute. Before optimising, cost needs to be traceable to a team, product or customer — which means tagging discipline enforced in infrastructure code rather than requested politely.

Once attribution exists, the conversation changes. It stops being a central cost-reduction exercise and becomes a set of decisions owned by the teams who can actually change the design.

Environments built by hand are expensive twice: they cost more to run because they are sized defensively, and they cannot be torn down because rebuilding them is risky.

Infrastructure as code changes the economics of switching things off. When an environment can be recreated reliably in minutes, non-production capacity becomes something you use rather than something you keep.

AI workloads introduce a cost line that behaves differently from compute and storage. It scales with usage in ways that are hard to predict and easy to miss until the invoice arrives.

Treat model spend as a first-class cost dimension from the start: attributed per application, monitored for anomalies, with caching, routing and model selection treated as cost controls rather than afterthoughts.