Alphabet's $85 billion capex allocation for 2026, Anthropic's annualized run-rate of $47 billion, and TSMC's CoWoS-L packaging backlog at Hsinchu are the same story read from three ledger lines. API contract revenue funds the next training run; the next training run lands on B200 NVL72 clusters packaged at TSMC Hsinchu; CoWoS-L capacity is at saturation, with Blackwell-generation GPU package lead times extending into Q4 2026. Google, Anthropic, and OpenAI are each compounding enterprise API revenue into the same Hsinchu production queue from the same side of the BIS Entity List.

The constraint for Baidu's ERNIE 4.5, Alibaba's Qwen 2.5-Max, and Tencent's Hunyuan Pro is structural, not strategic. Each runs on Huawei Ascend 910B clusters whose HBM2e memory bandwidth tops at approximately 800 GB/s per chip. A B200 SXM delivers 8 TB/s HBM3e bandwidth per chip, assembled at TSMC Hsinchu under CoWoS-L -- a packaging service that TSMC cannot supply to Huawei or any PRC entity under BIS foreign direct product rules. The training throughput gap is a fab outcome. TSMC's allocation discussions for 2027 Blackwell-cycle volume close with anchor customers Nvidia and Broadcom in Q3 2026, and Huawei holds no eligible quota.