Google's coverage in our briefs splits between the TPU substitution story (where Google is the most credible large-scale non-Nvidia inference alternative) and direct spot-tape signals from GCP regions where TPU pricing is starting to print at meaningful levels.
The $21bn Anthropic TPU order. Per the May 17 brief: Anthropic's $21bn TPU order (Broadcom design, Google fabrication / platform, 2026–2027 delivery) is the largest single non-GPU inference infrastructure commitment in the publicly disclosed set. Google captures this revenue across both Cloud (TPU rental) and custom-silicon (the underlying TPU silicon being produced via Broadcom). This is structurally distinct from AWS Neuron, where the silicon and the cloud platform are both internal.
The GCP TPU spot signal. Per the May 18 brief: TPU v5LitePod on-demand in ie-dublin printed +381% (from a near-zero baseline), consistent with GCP expanding TPU v5 inference capacity into Europe. This is the cleanest public spot-tape evidence of the TPU substitution arc materialising in pricing data. Comparable to AWS's Inferentia2 Stockholm signal, but for the Google ecosystem.
The four-hyperscaler Broadcom grouping. Google is one of four hyperscalers (alongside Microsoft, Meta, Anthropic+OpenAI) named in Broadcom's $20bn → $100bn AI revenue trajectory. Google's TPU programme is the longest-running and most established of the four — meaning Google captures both more current revenue from existing TPU deployments and longer-tail benefit from the maturing ecosystem.
The DeepSeek V4 / efficiency context. DeepSeek V4 (1.6T MoE, 27% of V3 FLOPs, runs natively on Huawei Ascend, $1.74/M token pricing) creates competitive pressure on Google's frontier Gemini family but also reduces per-token inference cost across the market — a tailwind for TPU as the lower-cost inference substrate of choice for cost-sensitive workloads.
The implicit Nvidia hedge. Google's position is unique: it benefits from both Nvidia's continued growth (GCP is a meaningful Nvidia customer for general-purpose AI compute) and from Nvidia's gradual substitution (TPU captures workloads that move off Nvidia). Few names have this kind of two-sided exposure to the AI infrastructure trajectory.
The composite picture: the cleanest substitution-arc beneficiary in the public-market space, with a structurally-distinctive position vs both Nvidia (as a partial substitute) and AWS Neuron (more mature, with named external customers).