**Persistent supply constraints from TSMC CoWoS packaging and HBM3e memory bottlenecks continue to support B200 rental rates in the mid-$4 to low-$7 per GPU-hour range across specialist providers as of mid-2026, even as NVIDIA ramps Blackwell production toward one million units monthly.** Hyperscaler capex surges and frontier-model training demand have kept on-demand capacity tight, with neocloud spot pricing sometimes dipping lower while enterprise allocations remain prioritized. The closely matched market odds around $4.50–$6.50 reflect trader uncertainty over whether expanded 2026 output—offset by custom ASIC competition from AMD MI350X and in-house silicon—will ease pricing or if sustained AI infrastructure buildout will maintain premiums. Key near-term catalysts include TSMC capacity additions and hyperscaler deployment updates that could shift availability before year-end.