$920 million per month is Google's committed spend with SpaceX for compute access, making SpaceX a principal node in the US AI infrastructure stack alongside Microsoft Azure. Google operates TPU v5p clusters and holds TSMC Hsinchu wafer allocation. It has the chips. Writing an $11 billion annual check to Musk-controlled infrastructure is a choice, not a gap -- and its timing, as the Trump administration negotiates an equity stake in OpenAI and its relationship with Musk's companies tightens, places the contract at an intersection of capital allocation and political alignment that neither Google nor the White House has described plainly.

Baidu's Wenxin group, Alibaba's Qwen team, and Zhipu AI are running their largest training workloads on Huawei Ascend 910B clusters whose peak FP16 throughput, per Huawei's own specification sheets, lands below one-third of the H100 SXM5's 989 TFLOPS, with no sovereign-adjacent infrastructure partner to convert API deployment revenue into the next training run. The ground between the two stacks is India. AirTrunk, owned by Blackstone since its 2024 acquisition, is committing $30 billion to Bengaluru and Hyderabad data centers, with first-facility operations targeted before end-2028. That commissioning date is the first observable signal of whether India holds as the APAC node the bilateral has not locked.