HBM prices forecast +30–40% in 2026 as XPU spending growth (22.1%) outpaces GPU spending — the inference-economics story is the memory story
Analyst forecasts have HBM prices rising another 30–40% in 2026 as AI datacenter buildouts consume the constrained advanced-memory supply. Enterprise XPU spending (TPUs, FPGAs, custom ASICs, specialized processors) is projected to grow 22.1% in 2026, outpacing GPU spending growth. The two trends together signal that the inference-economics story is shifting from GPU choice to memory architecture.
The HBM price trajectory is the constraint that shapes 2026-2027 infrastructure planning. Data centers are projected to consume up to 70% of global memory supply in 2026. With HBM3e tight and HBM4 yields uncertain, the marginal cost of inference at the largest deployment scales is no longer compute-bound — it's memory-bound. That changes the architectural decisions hyperscalers are making about which workloads run on which silicon.
The 22.1% XPU growth rate matters because it's faster than GPU spending growth. TPUs, FPGAs, and custom inference ASICs (Inferentia, MTIA, custom Microsoft silicon) are absorbing a growing share of the inference workload precisely because their per-token memory efficiency is better than general-purpose GPUs on workloads where the architecture matches the model. The closed labs are still GPU-bound for training; the inference frontier is moving toward custom silicon faster than the training frontier is.
Clarifai — GPU Shortages: How the AI Compute Crunch Is Reshaping Infrastructure → · The Register — Unpacking AMD's latest datacenter CPU and GPU announcements → · Yahoo Finance — AMD reveals new AI PC chips, details next-gen data center chips at CES 2026 →