Filing: Broadcom agrees to produce future versions of Google's TPUs and expands its Anthropic deal to give the startup access to ~3.5 GW of computing capacity
- Broadcom said it agreed to produce future versions of Google's artificial intelligence chips,
Context & Ripple Effects
This deepens an existing Google–Broadcom–Anthropic supply chain rather than creating a new one: Anthropic had already placed substantial orders for Google's Ironwood TPU racks through Broadcom in late 2025, and now has a larger, longer-lived claim on next-generation capacity. The accompanying disclosure that Anthropic's revenue run rate rose sharply gives the infrastructure commitment a clearer demand-side rationale.
For Google, retaining Broadcom for future TPU production resolves a supplier-continuity question that had remained open after reports it was also planning to work with MediaTek on a next-generation design. The arrangement turns Google's TPU program into a more explicit source of capacity for a major external AI developer, not solely an internal platform asset.
First-order effects
- Broadcom gains continued manufacturing work on future Google TPUs and a broadened commercial relationship tied to Anthropic's access to roughly 3.5 GW of compute capacity.
- Anthropic secures a defined path to large-scale next-generation TPU capacity, while Google gains a committed external customer for that infrastructure.
Second-order effects
- The deal reinforces demand visibility for the TPU production stack after Anthropic's earlier Ironwood rack orders, making capacity planning more consequential for Google and Broadcom than a one-off procurement arrangement.
- It also raises the strategic value of Google's alternative chip supply relationships: reported plans to involve MediaTek while retaining Broadcom pointed to a multi-supplier TPU strategy, which can reduce dependence on any single manufacturing partner.
Third-order effects
- If similar arrangements proliferate, frontier-model companies may increasingly obtain compute through long-duration, provider-backed capacity agreements rather than building every layer of infrastructure themselves.
- Custom accelerators could become a more contestable route to serving external AI workloads, with cloud platforms combining chip design, manufacturing partners, and committed capacity customers into a single infrastructure offering.
The trend: This is one data point in the platformization of AI infrastructure, where proprietary accelerator programs are being paired with contracted external capacity to finance and secure scale.