What Broadcom and Google actually agreed to
Broadcom entered a long-term agreement with Google to develop and supply custom AI chips, with the collaboration expected to run through 2031. The scope covers future generations of specialized processors as well as the supporting components that go into Google's next-generation AI infrastructure, extending a design relationship in which Broadcom has already helped build Google's Tensor Processing Units for close to a decade across multiple generations of the chip. The new agreement formalizes that work on a multi-year horizon rather than renewing it generation by generation as earlier contracts reportedly did.
The stated rationale is straightforward: Google wants to reduce its dependence on external GPU vendors, chiefly Nvidia, and believes custom silicon designed specifically for its own workloads gives it better cost and performance outcomes than general-purpose accelerators bought on the open market at whatever price and allocation Nvidia sets that quarter. That logic has driven TPU development for close to a decade, and this agreement extends the visible horizon on that bet out to 2031, a commitment length unusual even by hyperscaler infrastructure standards.
The Anthropic angle makes this a cloud story, not just a chip story
Alongside the Google agreement, Broadcom signed a separate deal giving Anthropic access to roughly 3.5 gigawatts of AI computing capacity running on Google's custom processors, starting in 2027. That is a substantial compute commitment on its own, and it ties one of the fastest-growing frontier AI labs directly to Google's custom silicon roadmap rather than to merchant GPU supply, at a moment when GPU allocation remains a genuine constraint across the entire AI industry regardless of budget.
For Google Cloud specifically, locking in a major AI lab's compute demand against its own chip roadmap is a stronger signal of TPU viability at scale than any published benchmark. It suggests Anthropic's own capacity planning team concluded that committing to Google's custom silicon pipeline years in advance offered a better risk profile than staying fully dependent on GPU allocation from Nvidia, a calculation other AI labs and large enterprises evaluating their own multi-year compute strategy are almost certainly running right now.
Part of a wider custom silicon trend, not an isolated deal
This agreement sits inside a broader pattern across the hyperscalers: AWS has its own Trainium and Graviton chip lines, Microsoft has been developing its own custom accelerators, and Google holds the longest track record of the three with TPUs stretching back to 2015. Each company is pursuing the same underlying economics, aiming to control the cost curve on AI compute rather than remain fully exposed to GPU pricing and allocation decisions set by a small number of external chip vendors with their own supply constraints and priorities.
Broadcom's position as the design partner behind both Google's program and other hyperscalers' custom silicon efforts makes it a direct beneficiary of this trend regardless of which specific cloud provider ultimately captures the largest share of AI workloads. The market reaction, with shares climbing on the announcement, reflects a growing view that Broadcom functions as a toll collector on the broader custom silicon shift across the industry, rather than as a bet riding on any single cloud provider's individual success.
What the undisclosed terms still tell us
Neither company disclosed the financial terms of either agreement, which is typical for infrastructure partnerships of this scale and duration, where pricing often scales with volume commitments that evolve over the life of the contract rather than locking in a single upfront figure. The absence of disclosed numbers makes it harder to size the deal against Broadcom's existing semiconductor revenue, but the six-year duration and the explicit Anthropic compute commitment both point toward a program measured in tens of billions of dollars across its full term, consistent with comparable custom silicon programs elsewhere in the industry.
What is disclosed, the 2031 end date and the 2027 start for Anthropic's access, gives enterprises a concrete timeline to plan against even without pricing details. A program that takes until 2027 to deliver committed capacity to its first named customer is a reminder that custom silicon partnerships of this scale operate on multi-year lead times, not the quarterly release cadence enterprises are used to seeing from merchant chip vendors competing for the same workloads today.
What this means for your chip and cloud strategy
For enterprises evaluating where to run AI training or inference workloads, custom silicon programs like this one signal that pricing and availability on major clouds will increasingly diverge based on which accelerator architecture a given workload targets. Workloads tuned for TPUs may see better economics on Google Cloud over time, just as Trainium-tuned workloads may see better economics on AWS, a divergence that argues against assuming easy, cost-neutral interchangeability between clouds for AI-heavy workloads going forward. Procurement teams that still evaluate clouds purely on list price per GPU hour are measuring the wrong thing as this shift plays out.
The practical move is to build portability into your AI infrastructure strategy now, before multi-year commitments like Anthropic's lock your organization into a specific silicon roadmap by default through inertia alone. Where possible, keep training and inference code abstracted from the underlying accelerator architecture, and treat a cloud provider's custom silicon roadmap, not just its published pricing sheet, as a genuine input into any multi-year capacity negotiation your team runs over the next several budget cycles. The vendors making six-year chip commitments are telling you how they plan to compete; your sourcing strategy should account for that horizon too.
