You already know: availability arithmetic and failure domains (D14 Reliability); object storage and replication (D14 Distributed storage); GPU memory and interconnects (D13); cost per token and utilization (D12 Inference economics); checkpointing against failures (D15 Training infrastructure); the D16 bridge's idle-capacity bill.
Why this matters for Atlas. The cloud turns hardware into an API: machines, storage, networks and permissions on demand, billed by the second. That flexibility is also the trap: costs scale with what you provision, not what you use; GPUs are scarce and region-specific; moving data across regions or out of the cloud costs money; and a single over-permissive credential can expose everything. This lesson covers the primitives AI systems use (compute including GPU instances, storage tiers, networking), the pricing models and how to combine them, identity and access management, and the cost engineering habits that fix bills like Atlas's.
1. Compute: instance families and GPUs
The question. Which machine should each Atlas workload run on?
Start from nothing. Clouds sell instances (virtual or bare-metal machines) in families tuned for different ratios of CPU, memory, network and accelerators: