Sofia Lindström writes that cloud GPU pricing has diverged sharply in 2026, with AWS offering the lowest single-H100 on-demand rate at $6.88/hr while Azure's eight-GPU node pushes per-GPU costs to $12.29/hr. Azure leads in spot discounts at roughly 80% and object storage at $0.0184/GB-month, but its GPU premium makes it the priciest option for accelerator-heavy workloads.
- Google Cloud enforces an eight-GPU on-demand minimum for H100s, forcing small teams to pay for six unused accelerators
- Specialist GPU clouds like RunPod price H100s at $2.89/hr, a 2.4x discount versus AWS's on-demand rate
- Secondary charges (storage, public IP, egress) add 25-40% on top of advertised VM rates across all three hyperscalers
- AWS P5 GPU quota approval requires written justification and takes 3-7 business days
- Google Cloud enforces an eight-GPU on-demand minimum for H100s, forcing small teams to pay for six unused accelerators
- Specialist GPU clouds like RunPod price H100s at $2.89/hr, a 2.4x discount versus AWS's on-demand rate
- Secondary charges (storage, public IP, egress) add 25-40% on top of advertised VM rates across all three hyperscalers
- AWS P5 GPU quota approval requires written justification and takes 3-7 business days
This blog post provides a guide for optimizing LLM serving performance on Google Kubernetes Engine (GKE) by covering infrastructure decisions, model server optimizations, and best practices for maximizing GPU utilization. It includes recommendations for quantization, GPU selection (G2 vs A3), batching strategies, and leveraging model server features like PagedAttention.
