Strong AI training and inference demand continues to outpace available H200 supply, driving median on-demand rental rates to $4.3–$6.2 per GPU-hour as of late October 2026, with recent 20–30% quarterly gains and spot rates 30–50% lower. Persistent HBM3e memory and power bottlenecks, plus multi-year contracts from hyperscalers and labs, have kept uncommitted capacity scarce even as Blackwell GPUs ramp. Providers including Nebius implemented 20% on-demand hikes in October while hyperscalers list $7–10+ rates; marketplace platforms like Vast.ai show greater volatility but limited relief. Further upward pressure is likely into early 2027 absent major new HBM or power capacity, with traders watching contract renewals and next-generation availability for any reversal.