[Feature]: Support AWS B200 (p6-b200) Instances

[NetanelK]

Problem

dstack currently does not support provisioning AWS B200 (p6-b200) instances. These next-generation NVIDIA Blackwell GPU (b200) instances are important for deep learning and AI workloads. The code restricts AWS instance families via a hardcoded allowlist and does not include p6/p6-b200 types, so users cannot select them for compute jobs.

Solution

Add support for AWS B200 (p6-b200) instances by including the relevant family in the instance allowlist and ensuring proper image selection and quota handling.

Workaround

Currently, there are no workarounds to use B200s with dstack; users can only provision supported instance families (e.g., p5, g5, etc.). Users who need Blackwell GPUs have to manage infrastructure outside of dstack.

Would you like to help us implement this feature by sending a PR?

Yes

[peterschmidt85]

Thank you, @NetanelK. Indeed, we need to support it!
Few quick questions:

1. Do you plan to submit a PR or want us to do it?
2. Do you have a chance to test it in your AWS account (do you think you have access to p6)?

[NetanelK]

Thanks! I can work on the PR, but I’ll need some guidance on the expected implementation details (e.g., any constraints around instance filtering, image selection, and validation for p6-b200).

I can also test the change on AWS from my side, subject to p6-b200 availability/quota in the target region.

[peterschmidt85]

Thanks! I can work on the PR, but I’ll need some guidance on the expected implementation details (e.g., any constraints around instance filtering, image selection, and validation for p6-b200).

If you know how to implement it, go for it. Of course we'll review it and share feedback if needed.
If you're rather not sure, please also let us know. Not a problem at all to support it on our side!

Regarding the ability to test it, please confirm if you can test it explicitly at least in any region!

[NetanelK]

Thanks! I don't know this part of the codebase well enough to implement it myself. If your team could implement it, I'd be happy to test it and confirm the behavior in .