Broadcom sings landmark deal with Google and Anthropic for custom AI chips

Broadcom’s (AVGO) back-to-back multiyear custom AI chip agreements with Google (GOOGL) and Anthropic put it at the center of a major shift in AI infrastructure. The headline is bigger than rising hardware demand. The deeper story is control: some of the biggest AI buyers are trying to take more influence over the chips that run their most important workloads.

That matters because the current AI boom has been shaped by general-purpose GPUs, especially from Nvidia. Those chips remain essential, but they also come with familiar constraints: high prices, supply bottlenecks, and the harder challenge of turning chips into actual deployed capacity. Broadcom’s role in these new agreements points to a different model. Instead of buying standard parts, hyperscalers are working with a partner to design silicon around their own economics, software stacks, and deployment priorities.

That is the real significance of the Google and Anthropic deals. Google is an established cloud and AI giant with years of infrastructure experience. Anthropic is one of the fastest-growing model developers in the market. They are very different customers, yet both are leaning into the same idea: custom silicon can become a strategic lever, not just a technical upgrade.

Broadcom benefits by moving further up the value chain

It is no longer just supplying components. It is acting as a design and integration partner early in the process, helping customers shape chips around specific bottlenecks such as inference cost, power efficiency, and speed of deployment. For buyers, that creates a path to reduce dependence on the same GPU pool everyone else is chasing. It also gives them more say over performance trade-offs and long-term infrastructure costs.

Broadcom has now won multiyear custom AI chip work from both Google and Anthropic, showing this approach can land with more than one kind of high-end customer. Google has also been open about the cost pressure that AI infrastructure puts on margins, especially as demand for compute rises. In that context, custom silicon is a practical financial tool as much as an engineering one, especially as investors look for which companies stand to benefit from Google’s 2026 capex spending. Anthropic’s involvement adds another real-world signal: this strategy is attractive to companies pushing model performance quickly and operating at a scale where chip choices directly affect competitive speed.

If enough hyperscalers start designing more of their own AI stack, the balance of power in the hardware market shifts. More of the value moves toward the companies defining the architecture and toward the partners, like Broadcom, helping them build it. Nvidia’s position is still strong, and custom chips will not replace general-purpose GPUs across the market overnight. But these deals suggest the old model—in which one dominant chip platform sets the pace for nearly everyone—is starting to loosen at the top end.

What to watch next

The next six to twelve months should show whether this is an early hyperscaler pattern or a narrower set of bespoke wins. A third major validation would matter most. If Microsoft or Amazon pursues a similarly deep design partnership with Broadcom or another custom silicon player, that would strengthen the case that this is becoming standard practice for the largest AI infrastructure operators.

Watch for hard evidence on economics as these deployments mature. Margin improvement, lower total cost of ownership, better power efficiency, or faster rollout times would do more to validate the shift than broad claims about customization. Without that proof, the story remains promising but incomplete. That is especially true in inference-heavy systems where memory constraints such as KV cache pressure can become just as important as raw compute.

Nvidia’s response will also be telling. Any move toward more configurable architectures, deeper integration services, or a stronger push into semi-custom offerings would suggest it sees real pressure from this model. At the same time, investors and operators should pay attention to execution risk. Custom silicon is harder to design, validate, and scale than buying proven chips off the shelf. Technical delays, cost overruns, or disappointing performance would narrow the appeal of this strategy and keep it concentrated among only the biggest buyers. And even if efficiency features improve, techniques like KV cache compression only matter when serving platforms can actually use them.

For now, Broadcom’s wins with Google and Anthropic point to a clear conclusion: the next phase of AI infrastructure may be shaped less by whoever can buy the most chips and more by whoever can define the chips they need. That does not end the GPU era, but it does mark the start of a more contested one.

Google (GOOGL) and Broadcom (AVGO) have a Disruption Score of 4 and 5, respectively. Click here to learn how we calculate the Disruption Score. 

Both Google and Broadcom are a part of the Disruption Aristocrats, our quarterly list of the world’s top disruptive stocks.