MOUNTAIN VIEW, Calif. — Google is in talks with Marvell Technology to develop two new artificial intelligence chips aimed at running AI models more efficiently, according to a Reuters report citing The Information. The discussions would expand Google’s push to turn its tensor processing units into a stronger alternative to Nvidia’s graphics processors as cloud providers race to cut the cost of serving AI models, April 20, 2026.
The reported plan centers on two chips: a memory processing unit designed to work with Google’s existing TPUs and a new TPU built specifically for inference, the stage where trained AI models generate answers, recommendations, images, code and agentic actions for users. Reuters said it could not immediately verify the report, and Google and Marvell did not immediately respond to requests for comment.
Why Google Marvell AI chips talks matter
The talks matter because inference is becoming one of the most important cost battles in artificial intelligence. Training frontier models is still expensive, but running those models at massive scale can become even more punishing as chatbots, search assistants, coding tools and AI agents respond to millions of users in real time.
A memory processing unit could help Google attack one of the hardest problems in AI computing: moving data fast enough to keep accelerators busy. Modern AI systems depend not only on raw chip speed but also on memory bandwidth, interconnects, software and power efficiency. If Marvell helps Google improve how TPUs access and process memory, the result could be a more efficient rack-scale system rather than just another chip launch.
The second reported chip, an inference-focused TPU, would fit Google’s broader strategy of designing hardware for the workloads it expects to dominate the next phase of AI. That is a direct strategic challenge to Nvidia, whose GPUs remain the industry’s default choice for both training and inference across many AI deployments.
A broader TPU supply chain, not a simple supplier swap
The reported Marvell discussions do not appear to mean Google is walking away from Broadcom. Earlier this month, Broadcom said it had signed a long-term agreement with Google to develop and supply future generations of custom AI chips and other components for Google’s next-generation AI racks through 2031.
That makes the Marvell talks look more like a diversification strategy. Google can keep Broadcom close, explore Marvell for memory and inference-focused designs, and use multiple design partners to scale its custom silicon roadmap. In AI infrastructure, the advantage increasingly goes to companies that can coordinate chips, memory, networking, software and cloud availability as one system.
Google also has signs of customer momentum. In February, Reuters reported that Meta had signed a multibillion-dollar deal to rent Google AI chips to develop new AI models, while also discussing the possible purchase of TPUs for its own data centers. If demand from large AI buyers grows, Google will need a deeper and more flexible silicon supply chain.
Older Google TPU moves show continuity over time
The Marvell talks would not be an abrupt pivot. Google publicly detailed its first TPU in a 2017 technical retrospective, describing a custom accelerator built to run neural networks efficiently inside Google data centers. In 2023, Google introduced Cloud TPU v5p and AI Hypercomputer, widening the pitch from individual chips to a co-designed stack of compute, networking, storage and software.
The strategy accelerated in 2025, when Google unveiled Ironwood as its first TPU for the age of inference. Around the same time, Reuters reported that Google was preparing to partner with Taiwan’s MediaTek on a next AI chip, showing that Google was already broadening its TPU design network beyond a single partner.
The cloud revenue stakes are rising
Google’s hardware push is tied closely to the economics of Google Cloud. Alphabet said Google Cloud revenue rose 48% to $17.7 billion in the fourth quarter of 2025, led by demand for enterprise AI infrastructure, enterprise AI solutions and core cloud products, according to the company’s fourth-quarter earnings release. Alphabet also said 2026 capital expenditures were expected to reach $175 billion to $185 billion as it builds capacity for AI demand.
That spending pressure helps explain why Google wants more control over its AI hardware stack. Nvidia GPUs are powerful and widely supported, but they are also in high demand and expensive to deploy at scale. A successful TPU platform could give Google a way to lower internal AI costs, offer customers an alternative to Nvidia, and improve margins in a cloud market where performance per watt and cost per token matter.
Nvidia remains the benchmark Google must beat
Nvidia still has the clearest lead in AI accelerators. The company reported record quarterly revenue of $68.1 billion and record data center revenue of $62.3 billion for the fourth quarter of fiscal 2026, according to Nvidia’s latest financial results. That scale reflects more than chip performance; it reflects Nvidia’s software ecosystem, developer familiarity, networking, supply relationships and deep integration across the AI industry.
For Google, the challenge is not simply building faster silicon. TPUs must be easy for customers to adopt, competitive on real workloads and available in enough volume to matter. Nvidia’s CUDA ecosystem and broad GPU support remain major barriers for any rival. Google’s advantage is that it can co-design chips around its own models, deploy them through Google Cloud and use internal workloads to test new systems before broader customer rollout.
What happens next
The key question is whether the Marvell talks turn into formal development work and, eventually, production chips. The reported memory processing unit could be finalized as soon as next year before test production, but chip timelines often shift. Designs can change, costs can rise, and supply-chain constraints can alter launch plans.
Even so, the talks send a clear message: Google is not treating TPUs as a side project. It is building a multi-year, multi-partner hardware strategy aimed at the AI inference boom. If Marvell joins Broadcom, MediaTek and Google’s internal silicon teams in that effort, Google’s TPU push could become one of the most serious attempts yet to weaken Nvidia’s grip on AI infrastructure.
For now, Nvidia remains the company to beat. But Google’s reported Marvell discussions show that the battle for AI chips is moving beyond single accelerators and into full-stack systems where memory, networking, software, power efficiency and cloud distribution decide who wins.
