Strong AI demand for large-scale training and inference continues to anchor B200 rental rates near current on-demand medians of roughly $7–$8 per GPU-hour across major providers, supporting the 76.5% market-implied probability for $7.00+ at year-end. Blackwell supply is expanding through hyperscaler and neocloud deployments, yet capacity remains constrained with many 2025–2026 allocations already committed, limiting downward pressure in the near term. Spot and reserved tiers trade lower, but on-demand benchmarks have held steady or edged higher over the past 90 days amid record Nvidia data-center revenue and persistent power and memory bottlenecks. Key near-term catalysts include additional GB200 rack shipments and any acceleration in alternative chip availability that could ease tightness before December.