AWS vs ThunderCompute

AWS employs per-second billing for on-demand instances, enabling precise cost control for short jobs, with spot instances offering up to 90% discounts for interruptible workloads and reserved instances for 1-3 year commitments yielding 40-75% savings. This granularity suits bursty ML training but introduces complexity via data transfer egress fees ($0.09/GB out) and tiered pricing across regions/GPU types (e.g., A100 at ~$3.30/hr on-demand). ThunderCompute uses per-minute billing, simpler for longer sessions but less efficient for sub-minute tasks, with no mention of spot/reserved options. Implications: AWS favors variable, short-duration experiments or large interruptible runs; ThunderCompute suits steady, developer-driven sessions without AWS's fee pitfalls, though total costs remain opaque without public benchmarks.

For small experiments (<1 hour), ThunderCompute's per-minute model may offer better value via simplicity and VS Code integration, avoiding AWS's setup overhead. Large training runs (>24 hours) favor AWS spot instances, slashing costs dramatically for H100/A100 clusters. Production inference benefits AWS's Inferentia for low-latency, cost-effective scaling with per-second billing. Batch jobs align with either, but AWS reserved instances excel for predictable volumes. ThunderCompute potentially wins for solo devs on tight budgets with intermittent use, but AWS delivers superior value for teams leveraging discounts and ecosystem efficiencies, despite higher entry costs—expect 20-50% savings on spots for most ML workloads.

AWS uses virtualized EC2/P2/P3/P4/P5 instances with NVLink for multi-GPU, EBS/EFS storage, VPC networking up to 400Gbps, and EKS for Kubernetes orchestration across 30+ regions. Supports bare-metal via i3en but emphasizes managed services. ThunderCompute details are sparse; focuses on remote dev UX with VS Code, implying virtualized GPUs, but unclear on bare metal, networking speeds, storage (e.g., no S3 equivalent noted), or Kubernetes—likely simpler, dev-centric setup without global redundancy.

AWS offers H100/A100/V100 GPUs with proven multi-node scaling (e.g., 16x H100 via Trainium clusters at 4x NVIDIA speed), low inter-node latency (<10us NVLink), and benchmarks showing 95% P99 uptime. ThunderCompute GPU types/performance unbenchmarked; VS Code integration aids usability, but no data on scaling, interconnects, or multi-GPU efficiency—assume standard for dev workloads, with uncertainty for production demands.