Anthropic's $965 billion Series H valuation and $47 billion annualized API and enterprise revenue are not separate events. The structure is one loop: Opus 4.8 ships on B200 NVL72 clusters at Anthropic's AWS-colocated Northern Virginia capacity, enterprise contracts accumulate at the $47 billion run-rate, that revenue funds the next training run on Nvidia hardware that BIS October 2023 Export Administration Regulations keep outside PRC purchasing channels, and the pending IPO converts the loop into public equity. Alphabet's concurrent $80 billion capital raise is the same structure at Google DeepMind: GCP deployment density in Singapore and Council Bluffs funds TPU v5p training runs, which ship through the Gemini API surface, which funds the next generation. The loop is self-reinforcing because deployment at this scale is also training signal: enterprise API volume at $47 billion generates fine-tuning data that calibrates successor models in ways that state-directed capital injection cannot replicate.

PRC labs do not run this loop. Baidu's ERNIE 4.5 Turbo, Alibaba's Qwen 2.5 series, and ByteDance's Doubao train on Huawei Ascend 910B clusters at Zhangjiang and Dongguan; the Ascend 910B delivers 256 TFLOPS at FP16 against the B200's 4.5 PFLOPS at BF16, with inter-chip interconnect at 392 GB/s against the NVL72 fabric's 14.4 TB/s. That geometry limits maximum batch size on 200,000-token training sequences. None of the three labs has disclosed API revenue at a scale that funds a training run of the size Anthropic's $47 billion run-rate implies, and state-directed capital from China's AI development budget does not supply the deployment-data feedback that makes the US loop compound. When Anthropic files its S-1, expected Q3 2026, it will disclose training-cost and revenue data that price the bilateral compute gap in public equity markets; Baidu's H-share and Alibaba's New York ADR will be marked against it.