Broadcom-Google Deepen Custom AI Chip Pact, Anthropic Gets 3.5GW

Broadcom and Google Extend Custom Silicon Alliance

Broadcom and Google have extended their long-running partnership on custom AI silicon, committing to joint development of next-generation TPU-based accelerators and the networking fabric that ties them together. The updated agreement locks in Broadcom as the co-design partner for Google's TPU v7 and v8 families, and includes custom Ethernet switching, optical interconnects, and HBM memory integration.

Just as significantly, Anthropic has committed to 3.5 gigawatts of TPU-based compute beginning in 2027, in what is one of the largest single compute deals announced for Google Cloud to date. The commitment nearly triples Anthropic's previously disclosed TPU footprint and places Claude training squarely on Google silicon for the foreseeable future.

"Custom silicon is no longer a science project for hyperscalers — it is the primary axis of competition." — Broadcom executive, speaking to analysts

Why Custom Silicon Is Eating the Accelerator Market

The Broadcom-Google-Anthropic triangle is the clearest evidence yet that hyperscalers are serious about reducing their dependence on Nvidia. Custom TPUs offer Google and its customers better perf-per-watt at a lower long-term cost, and — crucially — guaranteed supply. Anthropic's 3.5 GW commitment alone represents compute that would otherwise have flowed through Nvidia's H100/H200/B100 ecosystem.

For Broadcom, the deal solidifies its role as the go-to ASIC partner for frontier AI. The company is already working with Meta on MTIA, ByteDance on in-house accelerators, and Apple on server-class silicon for its private cloud compute stack.

Anthropic's Multi-Cloud Balancing Act

Anthropic has historically distributed its training workloads across AWS Trainium, Google TPUs, and Nvidia GPUs, a deliberate hedge against single-vendor lock-in and capacity crunches. The new 3.5 GW TPU commitment does not replace its AWS deal — rather, it stacks on top, reflecting the company's need for every watt of compute it can secure as it races to train Claude's next-generation models.

What This Means for AI Engineers

For engineers, the custom-silicon wave is reshaping the AI stack. XLA, JAX, and Pallas — the compiler and kernel layers that make TPU-native training work — are becoming some of the most valuable skills on the market. Expect hiring at Google, Anthropic, and Broadcom to intensify around compiler engineering, distributed training, and hardware-software co-design. For job seekers, the shift away from Nvidia-only stacks opens doors that simply didn't exist 18 months ago.