Alphabet’s Google is in discussions with Marvell Technology to co-develop two new chips designed to run artificial intelligence models more efficiently, according to a report by The Information citing two people with direct knowledge of the talks. No contract has been signed yet.
One of the chips is a memory processing unit designed to work alongside Google’s existing Tensor Processing Units (TPUs). The other is a new TPU built specifically for running AI models.
Inference: the process of executing a trained model to answer queries, generate text, or recognise images in real time is distinct from training, which involves building those models from scratch using vast datasets.
The companies aim to finalise the design of the memory-focused chip as early as next year, after which it would move into test production.
Diversifying Away from Broadcom
The talks represent a notable strategic shift for Google. The company has historically relied on Broadcom for its foundational TPU infrastructure.
Just days before the Marvell discussions emerged, Google extended its long-term TPU and networking agreement with Broadcom through 2031.
Bringing Marvell into the picture suggests Google is expanding its chip design ecosystem. Marvell specialises in custom silicon and data centre hardware, which makes it a natural partner for building highly specialised AI chips.
Adding a third design partner alongside Broadcom and MediaTek gives Google more flexibility in contract negotiations and insulates its data centres from supply chain disruptions.
Why Inference, Why Now
This partnership is a direct response to the shifting economics of AI.
The immense costs of training large models are rapidly giving way to the continuous, everyday expenses of running them.
Every search result, chatbot response, and recommendation served by Google’s AI products requires an inference computation. At scale, those costs add up quickly.
Custom inference chips address this directly. Compared to general-purpose Nvidia GPUs, they are optimised for specific tasks.
This translates to faster response times, lower power consumption, and reduced operational costs across data centres handling millions of queries simultaneously.
A Booming Market and Marvell’s Rising Profile
The timing coincides with rapid growth in the custom chip sector. The custom ASIC market is projected to grow 45% in 2026 and reach $118 billion by 2033.
Marvell is well positioned to benefit. The company’s custom ASIC business already contributes roughly $1.5 billion in annualised revenue, with design wins at Amazon, Microsoft, and several other major cloud providers.
In December 2025, Marvell also acquired Celestial AI for up to $5.5 billion, gaining photonic interconnect technology.
More recently, Nvidia invested $2 billion in Marvell and formed a strategic partnership through NVLink Fusion, which integrates Marvell’s custom chips and networking with Nvidia’s interconnect fabric.
Marvell’s stock has rallied approximately 50% year to date, with a 30% gain in April alone following the Nvidia partnership announcement and the Google talks.
Barclays analyst Tom O’Malley upgraded the stock to overweight and raised his price target from $105 to $150.
Pressure on Nvidia
Nvidia has been developing its own new AI inference chips using technology from Groq. Google’s push into custom inference silicon adds further competitive pressure.
If the Marvell partnership proceeds, it would reduce Google’s dependence on third-party GPU suppliers and deepen a broader industry trend: hyperscalers building their own hardware to cut costs and gain performance advantages.
