AWS Spending Caps Can Delete Your Project
AWS's new project spend limits stop resources at the cap and initially preserve your data.
AWS's new project spend limits stop resources at the cap and initially preserve your data. Its guide says project data is permanently deleted after 90 days paused without action. We unpack the limited rollout, Google's remaining charges, Germany's Kolibri model and COSMIC's contribution rules. Verdict: NEEDS REVIEW.
Read the written edition (English) ↗
What this video covers
- What can a cloud spending cap actually protect?
- How does AWS actually stop a project's spending?
- Who gets AWS's new spend limits?
- What does Google cap, and what keeps billing?
- How small is Germany's new open-weight model?
Transcript
What can a cloud spending cap actually protect?
0:00 A spending cap means your side project is safe. AWS's new limit can pause it, and leaving it paused can eventually erase the data. In this video, what stops the bill? What's the catch? Who actually gets these limits? There's a countdown in AWS's documentation. We'll follow it all the way to the delete button. It's Sunday, October fourth, and this is The Daily Diff.
0:20 On Saturday, Simon Willison called for default spending caps as agents make deploying software easier. Germany released Kolibri, while Cosmic's maintainers drew a line around AI contributions. Willison's argument is painfully familiar. A warning email arrives at midnight, the service keeps spending, and you wake up as the involuntary sponsor of your own infinite loop. AWS introduced project spending limits in its simplified builder experience in
How does AWS actually stop a project's spending?
0:44 September. When usage reaches the limit, AWS pauses the project and stops its resources. Your data is initially preserved. That's a real change in the product contract. The customer can authorize the cloud to interrupt the experiment. A spending alert leaves the decision with whoever is reading their inbox. A coding agent can provision resources while you're making dinner. The bill belongs to the account owner, who has somehow become both the engineering department and the finance department.
1:09 The rollout is limited, and you need a paid plan to set a limit.
Who gets AWS's new spend limits?
1:13 This belongs to the new project experience. Check availability before assuming your existing account has the same controls. The ceiling applies to charges before tax and excludes credits. The minimum is twenty dollars or AWS's conservative spending estimate, whichever is higher. Already running resources can push that minimum up. That matters for a neglected sandbox full of expensive machines. You may have to stop them before setting a smaller limit.
1:36 The cloud's financial advice begins with turning the computer off. AWS also offers optional early controls. About a week before the predicted limit, it can block new resources. Your existing instances stay running, while an autoscaling group may lose permission to expand. Other options pause idle resources, then top cost drivers as the forecast gets closer. A SageMaker endpoint with zero calls for two weeks qualifies as idle. Your forgotten demo finally receives an exit interview.
2:03 Willison wants hard caps switched on by default, with unlimited spending as an explicit choice. That's his proposal. AWS's documentation describes controls you configure, with limited access during this rollout. Google announced its own spend caps in July.
What does Google cap, and what keeps billing?
2:17 In public preview, they target one service inside one project, and Google says AI caps trigger within minutes of reaching the threshold. The listed services include Gemini and Cloud Run. Other services remain unaffected. Fixed contractual commitments keep billing, so pausing new usage still leaves any commitments you've already signed sitting on the invoice. For a business, an unexpected shutdown can be expensive too. AWS explicitly frames these limits around experiments and sandboxes,
2:43 with production use depending on whether a resource pause is acceptable. My developer takeaway is simple. Put experimental agents in a separate project, check which charges the limit covers, and test your recovery process. The guardrail needs an owner who notices when it fires.
How small is Germany's new open-weight model?
2:57 Meanwhile, Aleph Alpha released Kolibri, an open weight model for German and English. Its launch animation says small bird, fast wings. The weights are available under Apache two point oh. Chubby called it a surprisingly strong release from Germany. The useful part is deployment control. A company can hold the weights and run its documents through infrastructure it owns. About three and a half billion parameters work on each token, but the full model has seventy eight billion.
3:22 Its published weights occupy about seventy eight gigabytes, so small bird still needs a big cage. Tejas Kumar independently tested its tokenizer on Germany's constitution. It used about fifteen percent fewer tokens than OpenAI's tokenizer in German. On the English translation, they were effectively tied. That's a narrow, useful measurement of German text efficiency. He says he hasn't run the model itself. Keep that distinction when a friendly launch chart tries to become your
3:46 procurement department. And System seventy six's Cosmic compositor now asks
What exactly does COSMIC reject in pull requests?
3:50 contributors to certify that their pull request contains no language model generated content. The template explicitly includes code, comments and descriptions. Contributors also promise they understand and tested the change. The Flatpak repository has an exception for their own app source and manifest. Read the actual template before treating the whole project as one rule. The machine can produce another patch in seconds. Somebody still has to read it.
4:14 Today the cloud puts limits on the experiment, and maintainers put limits on the review queue. Back to that paused AWS project. Increasing the spending limit can reactivate it, and some resources may need manual restarts. Recovery is a procedure you should understand before an agent spends the allowance. The countdown is ninety days.
When does the pause become permanent data loss?
4:31 AWS says if you take no action within that time after the project is paused, it permanently deletes the project data. Put a backup outside that project's blast radius. A stopped bill and a retained backup are two separate jobs. I want the cap, and I want a recovery plan. My abandoned weekend prototype should retire with dignity. If you'd rather read this than hear me say it, the diff lands in your inbox every morning, free at the daily diff dot dev, link below.
What's my verdict on the new spending controls?
4:55 So today's verdict, NEEDS REVIEW. I welcome enforced spending limits. I'd verify availability, covered charges and recovery before handing an agent the account. A ninety day deletion clock deserves a very visible warning. And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.
Sources
- Willisonsimonwillison.net
- AWS launchaws.amazon.com
- AWS exact rulesdocs.aws.amazon.com
- Googlecloud.google.com
- Kolibrihuggingface.co
- Tokenizer testtej.as
- COSMICgithub.com
- Flatpak exceptiongithub.com
- HNnews.ycombinator.com



