Sources: Google in talks with Marvell on memory-processing unit to pair with TPUs — and a new TPU is coming

Deal and the chips

It has been reported that Google is in discussions with Marvell Technology to develop a memory processing unit (MPU) designed to work alongside the company’s Tensor Processing Units (TPUs). The talks allegedly center on hardware that would move memory closer to compute — a response to the growing bottleneck between vast model weights and the accelerators that run them. It has also been reported that Google is planning a new TPU iteration aimed specifically at running AI models, not just training them.

MPUs are meant to ease the pain of feeding huge models with data fast enough. Pair one with a TPU and, in theory, you get lower latency and better efficiency for inference workloads. Marvell, known for networking and custom silicon, would bring chip-building experience; Google would bring system-level design and AI software stacks. Neither company has confirmed details publicly, and it has been reported that discussions are ongoing.

Why it matters

Why should you care? Because this is where the cloud wars get tangible. Nvidia still dominates inference and training today, but hyperscalers from Amazon to Google are betting on bespoke silicon to cut costs and wrestle control back from third-party vendors. Is Google trying to chip away at Nvidia’s lead? Absolutely. And there’s emotion here — engineers racing to squeeze milliseconds out of inference, product teams promising faster, cheaper AI features. It’s the classic tech scramble: innovate or be left behind.

If the reports hold up, expect this move to accelerate the industry trend toward vertical integration — custom processors, tailored memory hierarchies, the whole shebang. It has been reported that talks are at a formative stage, and concrete timelines remain unclear. For now, the smartest part of the AI boom might be the hardware under the hood — not the flashy models everyone quotes in headlines.

Sources: theinformation.com