NVIDIA next-gen GPU delivery delays force major AI training curtailment before mid-2026

Tracks whether NVIDIA H200/B200 (or successor Blackwell-class) GPU delivery timelines experience confirmed delays or curtailments attributed to TSMC advanced packaging (CoWoS) capacity constraints, HBM3e supply shortages, or export-control-driven allocation disruptions that force a major AI training operator (1000+ GPU cluster) to publicly acknowledge training schedule delays, compute scale-backs, or procurement shortfalls before July 1, 2026. LONG if at least one qualifying delay or curtailment is confirmed via company earnings calls, SEC 10-K/10-Q risk factor disclosures, analyst reports, or public statements from major AI labs (OpenAI, Meta, Microsoft, Google, xAI, Anthropic); SHORT if GPU supply keeps pace with AI demand growth and no major training project is materially disrupted. TSMC CoWoS capacity remains the binding constraint on NVIDIA's high-end GPU output — current capacity is ~35K wafers/month with expansion timelines slipping. HBM3e supply from SK Hynix and Micron is also tightening. Each quarter that CoWoS expansion misses targets, or each BIS entity-list addition that disrupts chip flows to Chinese-adjacent procurement channels, strengthens the LONG thesis. Watch TSMC quarterly earnings CoWoS guidance, NVIDIA quarterly datacenter revenue commentary, and major AI lab infrastructure blog posts for evidence of compute-constrained training delays.