Google is paying SpaceX $920 million per month for compute capacity. The number is $11 billion per year. That scale is not supplementary cloud spend but a structural declaration that Alphabet, which announced an $85 billion AI capital raise in the same week, has chosen to route its training capex through a private aerospace operator rather than back into the Google Cloud infrastructure it already owns.

The shape on the US side is now legible. Anthropic runs on SpaceX Colossus. OpenAI runs on Microsoft Azure. Google, despite operating its own TPU cluster infrastructure across Singapore and Tokyo GCP regions, is contracting SpaceX for B200 NVL72 rack-density sourced from TSMC's Hsinchu fabs, hardware that BIS October 2023 controls bar from PRC datacentre import queues at any volume. On the PRC side, Baidu, Zhipu, and DeepSeek run on Huawei Ascend 910B clusters with no equivalent commercial compute partner to externalise the next training run's capex onto. The Ascend 910B delivers approximately 800 GB/s of HBM2e memory bandwidth per chip against the B200's 8 TB/s; BIS October 2023 controls have converted that gap from a procurement lag into a structural condition of the bilateral.