01 The short version
Best option when the rest of the platform is already on AWS and the data gravity is real. Managed container services for the workload itself, with the platform's own metered AI service kept as a capped fallback rather than as the primary path.
- g6e (L40S) is the sweet spot for 7B–32B models — considerably cheaper per token than H100 capacity you cannot keep busy.
- On-demand H100 is expensive enough that reserved or savings-plan capacity pays back fast once utilisation is above roughly 40%.
- Bedrock behind the gateway gives you a fallback with no new vendor relationship and no new key on anyone's laptop.
- Inferentia is worth benchmarking if your model fits — the economics change materially when it does.
02 What we do here
BUILD
Inference on AWS
Open-weight models running on AWS capacity, gateway routing in front, full cost and usage records underneath, the networking and IAM to hold it together. Deployed in your account with Terraform you keep.
RUN
Operate it monthly
Capacity planning against AWS quota and pricing, model rotations, incident response, and a monthly report that reconciles against your actual AWS invoice.
FOUNDATION
Accounts, network, IAM
If the AWS footprint does not exist yet, or exists as one account somebody made in 2021, we build the landing zone underneath before anything else goes on top.
03 The others
We are not loyal to any of them. The right platform is the one where your data already is, your latency is acceptable and the GPUs are actually available this quarter — and that answer changes.