Google starts limited sale of 8th‑gen TPUs to select customers

Google is trying to turn one of its deepest internal advantages into something customers can actually buy. The advantage is TPU — Google’s in-house AI(blog.google)r into public cloud, but not in a wide-open way. Google started limited sales of its eighth-generation TPU systems after showing t(cloud.google.com) (blog.google) ### What is Goog(abc.xyz)s: TPU 8t for training giant models and TPU 8i for inference, reinforcement learning, and fast agent-style workloads. That split matters because AI infrastructure has stopped being one-size-fits-all — training, post-training, and serving now stress hardware in different ways. (blog.google) ### Why split the TPU line in two? Because the bottlenecks changed. Training frontier models wants huge memory pools an(blog.google)swer was to specialize the stack instead of stretching one chip across every job. TPU 8t is the heavy lifter. TPU 8i is the fast responder. (cloud.google.com) ### How big is the training system? Big enough that Google is clearly aiming above “just another accelerator instance.” TPU 8t runs in superpods of up to 9,600 c(blog.google)es the top-end configuration at 121 exaflops and frames the whole thing as part of a vertically integrated AI stack, not a standalone chip launch. (cloud.google.com) ### Why only select customers? Because this is still scarce infrastructure. Google is pitching AI Hypercomputer — its bun(cloud.google.com)hat not every enterprise gets frontier-scale hardware on equal terms. Even bullish analysts are reading this as targeted reach, where Google chooses early customers carefully and keeps the biggest capacity for strategic accounts and internal workloads. That is normal for scarce AI compute, but it is still a gate, not an open shelf. (moorinsightsstrategy.com) of it as Google’s attempt to sell the whole machine room, not just the processor. TPU 8t and 8i sit beside Axion Arm hosts, the Virgo network, Managed Lustre storage, and software layers like JAX and Pathways. The pitch is simple: if Google co-designed the stack end to end, it can deliver better utilization, lower bottlenecks, and better performance per dollar than a pile of loosely connected parts. (cloud.google.com) ### Is this already showing up in Google’s numbers? (moorinsightsstrategy.com)opped $20 billion for the first time, and backlog jumped to more than $460 billion. That does not prove TPU alone is driving the business. But it does show customers are signing up for a lot more Google infrastructure tied to AI. (abc.xyz) ### So what’s the real significance? Google has had custom AI silicon for years. The change is commerc(cloud.google.com) “we will selectively rent this advantage to the market.” If that works, TPUs stop being a hidden ingredient inside Search and Gemini and become a direct weapon against Nvidia-heavy cloud rivals. (blog.google) ### Bottom line? This is less about one chip launch than about Google opening the door — carefully — to its private AI factory. The opportunity is huge. But for now, the keyword is still limited. (blog.google)