Google’s reported Marvell chip deal signals a wider race to diversify AI silicon supply

Google has reportedly deepened its relationship with Marvell Technology through a custom AI chip agreement that could give the search and cloud giant a multibillion-dollar stake in one of the semiconductor companies competing for hyperscale AI silicon work. The Financial Times reported on August 20 that the deal includes a warrant allowing Google to purchase up to 58.9 million Marvell shares at 206.58 dollars each, a package worth roughly 12.2 billion dollars if fully exercised.

MarketWatch and Investors Business Daily also reported that Marvell shares jumped after the disclosure, while Broadcom fell as investors weighed whether Google is diversifying away from a long-running TPU partner. The agreement was signed on July 29, according to those reports, and centers on custom chips for Google, including AI accelerators and controller silicon used around large data center systems. The full impact on Broadcom remains uncertain, because Google’s chip supply chain can support multiple design partners at once.

The strategic logic is clear. Google’s Tensor Processing Units have become one of the most important alternatives to Nvidia GPUs in large-scale AI, powering internal workloads such as Gemini and increasingly supporting outside cloud customers. As inference demand rises, hyperscalers want custom accelerators, networking, memory and storage controllers tuned to their own workloads. A diversified chip-design ecosystem can help reduce dependency on one supplier, improve negotiating leverage and accelerate specialized designs.

Marvell has spent years positioning itself as a data infrastructure company rather than only a merchant chip supplier. Its portfolio includes networking, storage, CXL memory products, custom compute work and optical interconnect ambitions. The company’s pending acquisition of Celestial AI, announced earlier, also gives it a deeper story around photonic fabric technology for connecting large AI systems. Those capabilities matter because AI infrastructure bottlenecks are moving beyond raw arithmetic into memory movement, rack-scale networking and power-efficient data transfer.

The reported warrant structure shows how financial incentives are becoming intertwined with AI supply chains. Google would not simply buy chips; it would have the option to become a major shareholder if performance and revenue milestones vest over time. Reports say a small portion of shares vests quarterly in the first year, while the remainder is tied to revenue benchmarks through 2033. That aligns Marvell’s upside with Google’s long-term custom silicon needs.

For the broader AI market, the deal reflects an arms race under the surface of consumer chatbots. Model developers and cloud platforms need massive amounts of compute, but the shape of that compute is changing. Training frontier models still demands enormous accelerator clusters, while commercial deployment adds relentless inference demand. The winners will be companies that can combine chips, networking, memory, software and supply agreements into systems that deliver more useful tokens per watt and per dollar.

There are caveats. The reported agreement does not mean Marvell has displaced Broadcom across Google’s TPU roadmap, and investors may be overreading a single disclosure. Broadcom remains deeply embedded in custom silicon and has its own AI partnerships. Still, the news underscores how quickly the AI chip market is expanding beyond a simple Nvidia-versus-everyone narrative. Hyperscalers are building their own supply webs, and chipmakers able to customize around AI infrastructure constraints are becoming strategic partners rather than ordinary component vendors.