Concrete sizing guidance for LLM workloads in 2026 — covering GPU selection (H100, H200, B200, MI300X, L40S), memory architecture, storage tiers, network requirements, and cost-optimised infrastructure patterns for inference, training, and batch processing.