On November 9, 2026, Google added two new chips to its eighth-generation TPU lineup: TPU 8i Max, which doubles the memory of its existing inference chip to 384GB of high-bandwidth memory, and TPU 8tx, a version built with hardware-level encryption so cloud customers can run sensitive AI workloads without exposing data or model weights even to Google itself. It’s the first time Google has shipped four distinct TPU systems in a single year.
From two chips to four, in one generation
Google introduced the base eighth-generation TPUs — TPU 8t for training and TPU 8i for inference — back in April 2026 at Google Cloud Next, splitting what had previously been a single do-everything chip design into two specialized ones. Google claimed TPU 8t delivered up to 2.7 times the training performance of the prior generation, Ironwood, while TPU 8i offered an 80% improvement in inference price-performance and roughly half the inter-chip communication delay. Seven months later, Google is splitting the line again, this time along two different axes: security and memory capacity. In a post on Google’s blog titled “Our newest TPUs: 8tx and 8i Max, built for the agentic era,” SVP and Chief Technologist Amin Vahdat framed both new chips as answers to problems that showed up only after enterprises started running actual agentic workloads on the base chips — not training-bench numbers, but real deployment friction around data isolation and context length.
TPU 8i Max: 384GB of memory for agents that don’t forget
TPU 8i Max physically packages two TPU 8i chips together, and Google says that doubles compute, doubles on-chip SRAM, and doubles HBM bandwidth over a single TPU 8i. According to Mark Lohmeyer, Google Cloud’s VP and general manager for AI and compute infrastructure, that works out to 384GB of high-bandwidth memory per package — up from 288GB on the standalone TPU 8i — delivered at roughly 17.2 terabytes per second of combined memory bandwidth. Lohmeyer’s framing was that the chip adds “more lanes to the highway” rather than just a bigger engine, since memory bandwidth, not raw compute, is increasingly the bottleneck for workloads that hold enormous amounts of context in active memory: an agent working across an entire codebase, or reasoning over hours of video, needs to keep far more state loaded than a single chat response does. Because TPU 8i Max packs twice the silicon into the same rack footprint as a standard TPU 8i deployment, Google is also pitching it as a density play — more compute and memory per square foot and per watt of a data center that’s already built. Google’s own TPU product page lists TPU 8i Max as “coming soon,” with no firm general-availability date yet.
TPU 8tx: the same silicon, locked down
TPU 8tx runs on the same underlying silicon as the training-focused TPU 8t, but pairs it with fifth-generation AMD EPYC host CPUs instead of Google’s own Axion chips, and adds confidential-computing hardware that keeps both customer data and proprietary model weights encrypted and isolated while they’re actively being processed — not just at rest or in transit, which is where most “confidential” cloud offerings stop. Google is aiming it squarely at regulated enterprises and government customers running proprietary models on sensitive data, and says select customers will even be able to deploy TPU 8tx inside their own data centers rather than Google’s. The chip borrows TPU 8i’s interconnect network and claims up to 2x lower latency than TPU 8t with what Google calls a “near-zero performance penalty” for the added encryption — historically, confidential-computing features have come with a measurable speed tax, so a near-zero penalty, if it holds up under independent testing, would be a notable exception. Google also positions TPU 8tx for smaller-scale jobs that don’t need a full training cluster: fine-tuning on private data, model distillation, and reinforcement learning runs that a customer wants to keep entirely in-house.
Why Google is doing this now
None of this happens in a vacuum. Google has spent 2026 pushing TPUs as a genuine alternative to Nvidia GPUs for both its own Gemini training and, increasingly, for external cloud customers who want to avoid Nvidia’s pricing and allocation queues — reporting around the announcement noted Google explicitly framing the new chips as taking direct aim at Nvidia’s AI accelerator business. Splitting a chip generation into four distinct variants in under a year is also a tell about where the actual customer demand is: not more raw FLOPs on a leaderboard, but memory capacity for agents that need to hold long context, and provable data isolation for the banks, hospitals, and government agencies that are the biggest remaining holdouts in enterprise AI adoption. Neither TPU 8tx nor TPU 8i Max is shipping yet — Google has pointed to 2027 for broader availability — so the real test of both chips’ claims, especially that “near-zero” encryption penalty, will come once independent benchmarks and early customers get hands-on access. For now, the announcement is as much a signal about what Google thinks the next phase of agentic AI infrastructure demands as it is a product launch.
Image: Taylor Vick via Unsplash.
