Google may be turning to an unexpected partner to build its next in house AI chip. According to a report from the analysis firm SemiAnalysis, relayed by Tom’s Hardware, Google is working with AMD to help design its tenth generation TPU, the custom silicon Google has built for years to train and run its own models, and the news matters because it would mark AMD’s first significant role in another company’s custom AI accelerator, a business that has largely belonged to a handful of specialist design houses.
The most interesting detail is what the chip is reportedly meant to do, because the design being explored places general purpose CPU cores directly alongside the tensor compute chiplets that handle the heavy math of AI, all in a single package. That hybrid layout is a departure from the usual approach of building an accelerator that does one thing extremely well, and the reason given is the shift toward reinforcement learning and agentic AI, the kind of work where a model plans, acts, and reacts over many steps rather than simply answering a prompt, which demands more flexible, general purpose computing than standard model training does.
AMD is a logical partner for exactly that kind of chip, since the company brings a deep library of CPU intellectual property, advanced packaging techniques for stitching different chiplets together, and real experience with hybrid designs, most notably its Instinct MI300A, a data center part that already combines x86 processor cores and accelerator chiplets on one package. In other words, the thing Google is said to want is close to something AMD has already shipped, which makes the reported pairing less surprising on a technical level than it first appears.
The strategic read is where it gets more pointed, because Google designing its own chips has always been partly a hedge against depending on Nvidia, whose graphics processors power most of the industry and whose market value has climbed past that of nearly every national economy. Bringing in AMD, Nvidia’s closest rival, to help build the alternative would deepen that hedge and give AMD a foothold in custom silicon that could matter well beyond this one project.
All of that comes with a large caveat, because this is a report rather than an announcement, neither Google nor AMD has commented publicly, and the precise scope of AMD’s involvement remains unclear, so it is possible the arrangement is smaller or different than it sounds. Still, even as an unconfirmed report it is a revealing one, since it fits a broader pattern of the biggest technology companies working to control their own chips and to spread their bets across suppliers, and it suggests the next generation of AI hardware is being shaped as much by the demands of agentic systems as by the race to train ever larger models. Whether Google and AMD confirm it or not, that direction is worth watching.
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →