## The Fear of Cloud Costs I work at a cloud company. I understand that the cloud is not as cheap as advertised, but it is also not rocket science to understand how you are spending. And yet, for the longest time, I have been hesitant to spin up resources on my personal cloud account. The fear of accidentally raking up huge costs kept me from experimenting freely. I suspect there could be others like me. People who want to learn, want to try things, but stay on the fence because cloud billing feels like a black box that might suddenly demand hundreds of pounds. If you are amongst this group, this article is for you :) ## The Context Last week, I was building an agent for an internal hackathon using Google's `agent-starter-pack`. Before touching the company-provided sandbox, I wanted to get a feel for how everything worked on my personal GCP project first. Call it "saving my allowance from company" if you want. At the end of the development, I wanted to understand the costs in my personal sandbox and get rid of the unnecessary costs. Since experimentation with AI-tooling is a default of my evenings these days, I decided to use Claude Code to get a foundational crash course in cloud cost management. ## Using AI to Understand Cloud Costs Having an AI walk me through the billing reports, suggest the right `gcloud` commands, and explain what each cost line item meant made the whole exercise faster and less tedious. Not to mention, lesser friction to get started - We all love a good conversational partner :D Claude helped me - Parse the billing breakdown by service and region - Generate a script to scan for resources across all GCP regions - Understand why my Vertex AI costs were spread across 10 regions (API load balancing, not deployed infrastructure) - Execute the cleanup systematically with verification steps Cloud cost management is not difficult. But it is tedious enough to put it off. Having an AI assistant removes that friction. Total spend for 3 days of hackathon work: **less than £2**[^1]. | Service | Cost | |---------|------| | Vertex AI (API calls) | £1.02 | | Cloud Storage | £0.03 | | **Total** | **£1.05** | Most of the cost came from API calls to Gemini 3 Flash. I chose the cheapest model deliberately because in my humble view, a*n overpowered LLM cannot solve the problems of a poorly designed agent workflow*. Better to get the system's design right with a cheap model first. ## Simple Things That Keep Costs Low ### 1. Read the billing reports The GCP billing console has "Group by Service" and "Group by Location" views. ~5 minutes of reading tells you exactly where your money is going. ### 2. Set up a budget at the start I set a budget alert for £5-10 per month[^2]. The budget does not stop you from spending. It just emails you when you are approaching the limit. That is enough to make you pay attention. ### 3. Clean up when you are done I finished the agent work during the week. On the weekend, I shut everything down. The temptation is to think "it is not a lot of money, let's leave it there." I was guilty of this for a long time. My project had buckets and BigQuery datasets sitting around from old experiments. They were costing me maybe £0.01-0.05 per month. Absolutely nothing. But the habit of cleaning up matters more than the amount saved. It builds the muscle memory you will rely on when the stakes are bigger and there is a VM running or a Vertex AI endpoint you forgot about. This cleanup was not about saving £1. It was about building that habit. ## What I Found | Resource | Count | Action | |----------|-------|--------| | Storage Buckets | 2 | Deleted | | BigQuery Datasets | 5 | Deleted | | Service Accounts | 1 custom | Deleted | | Vertex AI API | Enabled | Disabled | | VMs | 0 | Nothing to do | | GKE Clusters | 0 | Nothing to do | The Vertex AI costs were confusing at first. Charges appeared across 10 different regions, but I had not deployed anything to those regions. Claude helped me (the trainee AI guy) understand - these were API inference costs. Google load-balances requests globally. When you stop calling the API, the costs stop. There is nothing to "delete." ## Final points To be fair, this was API-only work with no deployed infrastructure. If you are training models or running persistent endpoints, costs can escalate quickly. But the same principles apply. Read your billing reports. Set a budget. Clean up when you are done. Do not let the fear of an unknown bill stop you from learning on the cloud. [^1]: Yeah, that's not much. So, I am a happy guy. But I still went ahead with understanding this and deleting what I did not need because that's a habit I would like to get into! [^2]: Ironic, really. I spend way more than that on a night out. But hey, priorities.