$47 billion in annualized API revenue, per Simon Willison's sourcing of Anthropic's internal figure, is the number that explains what the $965 billion Series H is financing. Anthropic is not raising equity to prove a valuation; it is pre-committing to NVL72 cluster capacity at NVIDIA's current spot pricing before an IPO locks the balance sheet into public-market quarterly reporting cycles. Alphabet's concurrent $80 billion raise, reported by TechCrunch this week, runs the same logic: Google's Gemini training runs on internal GCP infrastructure whose capital costs are absorbed by Google Cloud enterprise API contracts, and the raise reserves cluster capacity on the same planning horizon. Both balance sheets are doing the same thing. Enterprise API deployment revenue funds the next cluster reservation.

Baidu's ERNIE 4.5, Alibaba Cloud's Qwen 2.5, and ByteDance's Doubao each run inference on Ascend 910B clusters manufactured at SMIC's N+2 node, where HBM bandwidth runs at approximately 1.6 TB/s per card against the B200 NVL72's 8.0 TB/s, with CoWoS-L packaging for B200 modules supplied through TSMC's Hsinchu fabs under export-compliant supply chains. That is not a software problem. It is a physics constraint on context-window size and batch throughput at inference scale, and it is the constraint that limits what a PRC enterprise API contract can promise a buyer. None of the three labs carries a publicly disclosed annualized API run-rate comparable to Anthropic's $47 billion figure, because the US enterprise market that would generate one is largely inaccessible and the domestic PRC market has not produced equivalent disclosed revenue. BIS October 2023 controls (Advanced Computing Rule, 4,800 TOPS threshold, updated November 2024 to close parametric workarounds) ensure the hardware gap does not narrow through import. Anthropic's IPO, now in preparation according to Wired, will price that asymmetry into public equity markets before the end of 2026 Q3.