I ran two years of production on Azure. Giant Context has run on Google Cloud since its first commit. The industry treats a cloud as a career specialisation that goes on a CV and rarely changes. Most of it is the same set of services under a different set of names.

That Azure stack was twenty-one function apps, a service bus, a Kusto cluster for analytics, a key vault, B2C identity, blob storage and App Insights watching all of it. The whole mapping to Google Cloud came to this:

Cold starts, dead letters, log retention, key rotation, what to do when a queue backs up at two in the morning. All of that is knowledge about queues and processes and secrets, and every bit of it transferred from the left column to the right.

Azure is built for an organization. Subscriptions arranged in management groups, procurement, portals inside portals and governance to configure before anything deploys. All of that exists because its customer has a compliance team and a budget owner who are not the person writing the code.

Google Cloud assumes a project ID and a terminal. The unit of everything is a project, a project is free to create, and a person working alone never has to model an org chart before deploying a container.

A CLI inherits whatever resource model sits under it, so the same job takes a different number of words. Deleting the destination that messages arrive in, on each side:

A Service Bus queue lives inside a namespace inside a resource group inside a subscription, so the command names all three every time. A Pub/Sub topic lives in a project, the project is set once in configuration, and everything after it inherits that. The gap repeats on everything typed in a day.

The overhead Azure asks for is doing real work in a company of two hundred, and it is pure cost in a project of one.

Every service in the stack, at list price, with the free tier attached:

Identity is five and a half times cheaper on Google above an identical free tier. Logging is four and a half times cheaper with ten times the free allowance, which on a platform that writes a log line per request is the difference between a line item and nothing. Service Bus Standard charges $9.86 a month before a single message arrives, where Pub/Sub charges for throughput and nothing else. Data Explorer needs a cluster that exists whether or not anyone queries it, and BigQuery has nothing to provision. Object storage and secrets land within a fraction of a cent of each other.

Cloud Run bills for instance time, and an instance can serve many requests at once. Google publishes two worked examples for the same workload, ten million requests a month at four hundred milliseconds each on one vCPU and 512 MiB of memory. At twenty requests per instance it estimates $13.69 a month. At one request per instance it estimates $81.72.

Six times the bill for identical traffic on an identical machine. Nothing changed except whether the container was allowed to handle more than one request at a time.

This platform is seven services, each with modest traffic, all mostly idle, all thread-safe enough to take concurrent requests. Billed per invocation that is seven meters running at once. Billed as instance time with concurrency, the same seven services are containers asleep most of the day.

With steady load at scale, provisioning is often the cheaper answer and Azure is right. Early, with spiky load and long idle stretches, it is the expensive one. Azure has hosting plans that behave like Cloud Run, and where a service is saturated the two converge.