Persistent supply constraints from HBM3e memory shortages and TSMC CoWoS packaging bottlenecks continue to anchor B200 rental rates near current marketplace medians of $5.66–$6.79 per GPU-hour, with hyperscaler reservations locking most new capacity through late 2026. Trader sentiment reflects uncertainty over whether NVIDIA’s ramp toward one million Blackwell units monthly will meaningfully expand spot and neocloud availability by October, or if sustained AI inference and training demand will keep on-demand pricing elevated above $6.50. New HBM capacity additions expected only in Q4 could begin easing premiums late in the period, yet lead times remain 8–16 weeks for enterprise buyers. The broad distribution across price bins captures this balance between gradual supply growth and resilient demand.