Everyone's rushing to build "custom" models without asking the first cost question. Saw a guide using BigQuery and Vertex AI for this. No price tags.
Let's break down what that actually costs you per month at medium scale.
* BigQuery: 5 TB processed data, ~$115
* Vertex AI Training: 50 node hours of n1-standard-8, ~$200
* Vertex AI Prediction: 100K predictions, ~$100
* Cloud Storage: 500 GB, ~$10
That's $425/month baseline, before any mistakes. Using always-on endpoints? Multiply that.
The serverless allure is a cost trap for batch workloads. For a static model, you're better with:
* Pre-emptible VMs for training
* Commitments for data storage
* Export the model and run predictions in a cheaper environment (like Cloud Run or even a committed instance)
If you didn't model these costs before building, you already failed.
show me the bill
Your breakdown is solid, but the real trap is assuming you need that scale from day one. Many attribution models are iterative. You can prototype with a subset in BigQuery, maybe 100 GB, for under $5. Train on a single n1-standard-4 for a few hours.
The $425/month becomes real only when you operationalize a full, daily batch pipeline. That's where your preemptible VMs and commitments advice hits home. If you're still validating the model logic, you shouldn't be paying for always-on endpoints at all.
Your bill is too high.