Broadcom gigawatt deal and the multi-stack frontier — when Google sells shovels to its competitor at gigawatt scale

Anthropic, Google, and Broadcom's multi-gigawatt compute partnership is the most explicit "we sell shovels to everyone" move Google has made in the AI capacity-supply business. Combined with the BIS January 2026 case-by-case-review rule shift on H200 and MI325X chips, the compute layer is now structurally multi-stack at frontier-lab scale and increasingly multi-jurisdiction at policy scale.

The diversification logic is the substantive piece. The Anthropic-Google-Broadcom partnership for multiple gigawatts of next-generation compute adds TPU capacity from Google plus custom-silicon contribution from Broadcom to Anthropic's compute stack — which previously was dominantly NVIDIA with the rumored SpaceX/Colossus 1 arrangement adding capacity at scale. The gigawatt-scale commitment is consistent with Anthropic's compute trajectory toward production Claude 5 and the agentic-platform capacity demands the lab is scaling against.

The Google side of the deal is the strategic signal that distinguishes this from any prior frontier-lab compute arrangement. Google's TPU stack has historically been an internal-DeepMind-and-Google-Cloud asset, with peer-lab access limited or non-existent. Selling TPU capacity at gigawatt scale to Anthropic — a direct competitor in frontier-model markets, a company whose growth threatens Google's own AI revenue trajectory — is the most explicit move Google has made toward treating compute supply as a separate business from model competition. The same arms-dealer logic underlies AWS Trainium 2 deployments and Microsoft Azure's NVIDIA-and-AMD bidirectional capacity sales: the compute layer is monetized across all comers, regardless of whether they compete in the model layer.

The four-way silicon competition that emerges is structurally important. NVIDIA via the GPU stack and the new Groq-acquired inference layer, Google TPU via Cloud and Broadcom-custom-silicon, AMD via the MI400 trajectory, and the custom-Cerebras/Groq/SambaNova standalones — four credible silicon-platform options at the production scale that actually matters for capacity planning. For frontier-lab compute strategy, the implication is that the right answer is no longer "all NVIDIA" — even if NVIDIA remains the largest single supplier. The diversification pattern through 2026 is what every large-scale AI deployer is replicating to manage cost, supply-chain risk, and workload-optimization simultaneously.

The policy layer running underneath the compute-stack diversification is the export-control regime. The BIS January 2026 rule shift from "presumption of denial" to "case-by-case review" on H200 and MI325X chip exports to China represents the prior administration's calibrated answer to the China-edge tradeoff. The shift lets specific exports proceed under exception process rather than being categorically blocked, which loosens the framework meaningfully while keeping case-by-case discretion. The current US House debate on further restrictions tests whether the calibration shifts back toward stricter controls under new political pressure.

The transatlantic-coordination layer is the third dimension. The US-EU coordination talks gaining momentum through May 2026 would knit together the BIS framework, the US AISI evaluation regime, the EU AI Act, and the UK AISI's third-party evaluation operations into a coordinated allied posture. If the coordination produces harmonized export-control requirements, the policy uncertainty in compute-supply decisions reduces materially; if it doesn't, the multi-jurisdiction patchwork persists as the operating reality. Frontier labs make compute-supply commitments inside whichever regulatory environment exists; the partnership announcement signals Anthropic and Google making their commitments under the current ambiguity.

For the standalone-inference category (Cerebras, Groq pre-acquisition, SambaNova), the multi-stack frontier-lab compute environment is competitive pressure. Frontier labs increasingly source inference capacity from multiple suppliers — Cerebras for some workloads, NVIDIA-with-Groq-LPU for others, custom-Broadcom-silicon for still others. No single non-NVIDIA inference supplier becomes a primary anchor; instead, every inference supplier becomes one of several anchors. The standalone players capture meaningful market share but at lower per-customer concentration than the silicon-only investment thesis assumed in 2023-2024.

The line: NVIDIA used to win because there was no alternative. In 2026 Google sells TPU to its competitors at gigawatt scale, and the compete-on-many-stacks frontier is the new normal.

Anthropic — Google-Broadcom compute partnership announcement → · Google Cloud Blog — TPU multi-gigawatt customer relationships → · BIS — January 2026 export-control rule on advanced AI chips →