Claude on AWS and Google Cloud: what I took from the courses

I finished Anthropic’s courses on Claude for AWS Bedrock and Google Cloud Vertex AI this week. Along the way, I managed to exceed Google’s allowance before getting an answer to my first question.

Paid account. Model switched on. One API call. The error was 429: “too many requests”. I had made one.

Every Claude quota on my Google Cloud project was zero. Global, Europe, every region I checked. I requested an increase at 11:51. It was refused at 11:51, automatically, because there wasn’t enough usage history. You need usage to get quota, and quota to get usage.

I completed the exercises using Bedrock and Anthropic’s direct API instead. The Google quota issue remains unresolved. I didn’t open a support case, so I can’t tell you whether a human would have fixed it. For a course exercise, I could take another route and carry on. On a client project, that decision would need a rather longer conversation.

That was what sent me into the paperwork.

I’ve spent fifteen years in finance operations, much of it as the contractor brought in to build something. Before I could start, there was often a meeting with IT and compliance to explain what it would do, what it would touch, and who would be responsible when something went wrong.

I didn’t blame them for asking. The company had customers’ money, employees’ records and a regulator’s attention, and I was a stranger asking to put software next to all three.

Working through the courses brought me back to those conversations. The model call looks much the same with a different client line. But if I were proposing it to a client, “it works on AWS” would barely begin to answer the questions in that room.

Three things from the reading I did around the exercises will stay with me.

The first is to check who actually runs the service. The logo on the invoice tells you who to pay. It does not necessarily tell you who is running your AI.

Amazon Bedrock and Google’s Vertex AI are cloud-operated routes to Claude. Claude Platform on AWS is a different arrangement: Anthropic operates it, with billing through AWS Marketplace.¹ AWS’s own comparison says that service is outside its standard compliance programmes, whereas Bedrock sits within the programmes listed for it.²

That distinction matters before anyone starts reassuring a client that something is “on AWS”. I’d need to name the service and establish whose controls cover it. The client already buying from a cloud provider doesn’t settle every question about another service on the same bill.

The second is that “does it train on our data?” only answers part of what a client is asking.

AWS says Bedrock content isn’t used to improve the base models. Google says it won’t use customer data to train or fine-tune models without prior permission or instruction.³⁴ Those are useful commitments. They leave separate questions about whether prompts are stored, how long they’re kept, and which company’s people might review them.

The documents I read in September showed why I would check those answers for the specific model. On Bedrock, Fable 5 and 5.1 require a mode that keeps prompts and outputs for up to thirty days for AWS review, without sharing them with the model provider.⁵ Google’s documentation required sharing with Anthropic for abuse monitoring for Fable 5 and Mythos 5, with separate restrictions inside its VPC Service Controls security perimeter.⁶⁷

These were findings from the documentation, beyond my course exercises. They changed the answer I would prepare for a reviewer. If someone asks which companies can see information sent by a reconciliation tool, a sentence about encryption won’t answer them. They need the companies named and the circumstances explained.

I would also put a date on that answer. During this reading, I found that AWS had updated its position on sharing with model providers on 2 September.⁸ Reusing an old questionnaire response could mean giving a client an answer that used to be correct.

The third lesson came directly from that first refused request: establish capacity early.

This isn’t unique to my Google account. AWS documents reduced quotas for new accounts; Google says quotas vary by account and access can be restricted.⁹¹⁰ My experience doesn’t establish that one cloud will always let a new customer in and the other won’t. It does establish that paying and enabling a model weren’t enough for my project.

For future work, I want the intended model accessible in the client’s intended environment before I put dates against delivery. If access needs a support conversation, it belongs near the start of the plan. Discovering that while doing a course was inconvenient. Discovering it after promising a month-end deadline would be a different problem.

I’m pleased to have completed both courses. They gave me a practical way into using Claude through the cloud services a client may already have, and the exercises gave me something concrete to investigate. The further reading connected that learning to the work I’ve spent much of my career doing: building something useful and being able to explain why it belongs near a company’s systems.

For the next client conversation, I’ll be better prepared to name the service, explain what happens to the data, and check that we can actually use it. I’d want those answers ready before asking IT to say yes.

Sources