Persistent supply constraints from HBM3e memory packaging bottlenecks and sustained demand for large language model training and inference have anchored current H200 on-demand rental rates around a $4.38 median across neocloud providers, with the tightest bids clustered between $4.00–$5.00. Competitive dynamics hinge on the pace of new Blackwell and additional Hopper capacity reaching the spot market versus hyperscaler commitments, alongside efficiency gains in model inference that could moderate token demand. Traders see limited near-term relief by late October absent major verified deployments, while noting volatility from export rules and weekly capacity fluctuations as key swing factors in this closely contested range.