← Back to feed

On April 6, 2026, Broadcom filed an 8-K with the Securities and Exchange Commission disclosing two distinct agreements that together define Google’s AI silicon strategy through the end of the decade.

What the 8-K Actually Says

Agreement one: A Long-Term Agreement for Broadcom to develop and supply custom Tensor Processing Units for Google’s future TPU generations, running through at least 2031. This is a design and supply commitment — not just procurement. Broadcom is co-developing the chips.

Agreement two: A Supply Assurance Agreement for Broadcom to provide networking and other components for Google’s next-generation AI racks, also through 2031. High-speed interconnects and rack-level infrastructure are co-designed alongside the compute die.

The second agreement matters as much as the first. Custom accelerators don’t operate in isolation. Google is not buying off-the-shelf components from whoever is cheapest in a given quarter — it is locking Broadcom into the full stack for five years.

Third announcement: Broadcom, Google, and Anthropic expanded their existing collaboration. Starting in 2027, Anthropic will access approximately 3.5 gigawatts as part of multiple gigawatts of next-generation TPU capacity Anthropic has committed to. That consumption is conditional on Anthropic’s continued commercial success.

Broadcom shares rose 4.69% in post-announcement trading.

Why the Duration Is the Story

Five-year chip design contracts are unusual. Typical hyperscaler procurement runs in 18-to-24-month cycles. A 2031 commitment signals that Google has resolved its custom silicon architecture debate internally — TPUs are the answer, Broadcom is the foundry partner, and the roadmap is locked.

The implication for NVIDIA is structural. Google is not a marginal NVIDIA customer. It runs one of the largest AI inference workloads in the world across Search, Ads, Workspace, and Gemini. Every TPU rack that Google deploys is an H100 or Blackwell rack it did not buy. The 2031 date means that calculus is settled for the near-term planning horizon.

The Anthropic Numbers

Anthropic’s 3.5 gigawatts is a useful benchmark for thinking about AI compute scale. For context, 1 gigawatt of AI compute is roughly 50,000 to 80,000 H100-equivalent GPUs at full load, depending on efficiency assumptions. Anthropic is committing to capacity an order of magnitude larger than most enterprise AI deployments will ever require.

The commitment is conditional — Broadcom noted in the filing that Anthropic’s consumption “is dependent on Anthropic’s continued commercial success.” With an annualised run rate exceeding $30 billion as of early 2026, that condition appears to be met.

Key Numbers

• Duration: Through 2031 (both agreements)
• Scope: TPU chip design + supply, plus rack networking components
• Anthropic allocation: 3.5 GW, online from 2027
• AVGO share reaction: +4.69% day after announcement
• Google TPU investment total: Part of a broader $600 billion AI infrastructure buildout Broadcom estimates across hyperscalers through the decade