Google is adding more muscle to its custom AI chip strategy… and longtime chip partner Broadcom may have more competition for the business.
Google and Marvell have expanded their relationship to develop custom silicon spanning AI inference accelerators, networking, storage controllers, and near-memory computing. The agreement, reported on Aug. 19 by Financial Times, also gives Google warrants that could eventually allow it to acquire up to 58.9 million Marvell shares, which would cost about $12.2 billion to exercise in full at the stated price.
The deal provides another signal that the AI infrastructure race is moving beyond simply securing more GPUs. Hyperscalers increasingly want specialized chips, networking, memory, and tightly integrated systems built around their own workloads, creating new competitive pressure across the semiconductor supply chain.
Google gives Marvell a bigger role in custom silicon
Marvell is expected to work with Google across several categories of custom chips rather than a single AI accelerator.
That breadth matters. AI infrastructure increasingly depends on moving and feeding data as efficiently as processing it, making networking, storage, memory, and specialized accelerators part of the same performance equation.
The agreement could also give Marvell a significantly larger foothold inside Google’s infrastructure. The warrants issued to Google are tied to conditions that include purchases and revenue generated through custom products, and Google would pay about $12.2 billion to exercise all 58.9 million warrants at the stated exercise price.
Investors quickly interpreted the announcement as a competitive development. Marvell shares jumped following the news while Broadcom shares declined, as markets weighed the possibility that Google could distribute more of its custom-chip work across multiple suppliers, according to MarketWatch.
Broadcom remains deeply embedded in Google’s custom silicon efforts, so the Marvell agreement does not mean Google is abandoning its existing partner. Instead, it suggests Google wants more options as AI infrastructure becomes larger, more complex, and more strategically important.
The custom AI chip race is getting crowded
Google’s move fits a broader shift away from an AI infrastructure market defined almost entirely by general-purpose accelerators.
Channel Insider has already tracked how Google sits at the center of a massive financing network supporting Anthropic’s TPU expansion, including infrastructure built around Google’s own Tensor Processing Units rather than relying exclusively on Nvidia GPUs.
AI companies themselves are seeking greater control. Anthropic is developing custom AI chips for future Claude infrastructure, even as it continues to use hardware and cloud capacity from companies including Google, Amazon, Nvidia, and AMD.
That trend puts custom silicon at the center of a growing question for the AI market: how much infrastructure will eventually be built around chips designed for specific workloads rather than hardware sold broadly across the industry?
Google appears determined to keep multiple answers available.
What the shift means for the channel
For channel partners, the immediate opportunity is unlikely to involve selling Google’s custom silicon directly. The larger effect could come from what happens around those chips.
As AI infrastructure becomes more specialized, organizations may face a growing mix of GPUs, TPUs, custom accelerators, networking architectures, storage systems, and cloud platforms. That increases the importance of partners who can help customers determine where workloads belong, how different environments connect, and whether specialized infrastructure delivers sufficient performance or cost savings to justify the added complexity.
Channel Insider has previously examined how Anthropic’s expanding compute strategy is spreading AI workloads across Google, AWS, Nvidia, and other infrastructure providers. Google’s expanded relationship with Marvell pushes that diversification one layer deeper, into the silicon powering those environments. That could make AI infrastructure more competitive, but also less uniform.
For integrators, cloud advisers, and infrastructure specialists, that fragmentation may create a familiar kind of opportunity: customers increasingly need someone who understands how all the pieces fit together.
Related reading: For another sign of how AI infrastructure competition is evolving, see how Groq’s $350 million funding round is fueling an Nvidia-powered expansion of its AI cloud.