Google adds two more eighth-generation TPUs, one built for sensitive data
Google added the TPU 8tx, a secured version of its training chip that some customers can run in their own data centers, and the TPU 8i Max, which packs two inference chips into one.
Kiln, covering chips, data centers and power
, 3 min read
4eighth-generation TPU systems Google introduced in 2026
What to know
- Google announced the TPU 8tx and TPU 8i Max on November 9, joining the TPU 8t and TPU 8i launched in April.
- The TPU 8tx uses TPU 8t silicon with added hardware security and AMD EPYC host processors.
- The TPU 8i Max puts two TPU 8i chips in one package, doubling compute, on-chip memory and memory bandwidth.
Google added two systems to its eighth generation of Tensor Processing Units (TPUs), its in-house AI chips, on Monday, November 9. The TPU 8tx is aimed at governments and regulated companies that want to run AI models next to sensitive data, and the TPU 8i Max is aimed at inference with very large amounts of context. Data Center Knowledge reported that Google will let select customers run the TPU 8tx in their own data centers rather than only in Google Cloud.
The two systems join the TPU 8t, for training, and the TPU 8i, for inference, which Google introduced at its Cloud Next conference on April 22. Google said 2026 is the first year it has introduced four TPU systems, according to Unite.AI. Neither report gave prices or a date when the new systems can be ordered. Google's own announcement post could not be read for this story, so the details below come from those two reports.
TPU 8tx: a training chip with locks on it
The TPU 8tx uses the same silicon as the TPU 8t but adds hardware-level encryption, so a customer's data and model weights stay protected while the chip is working on them, Data Center Knowledge reported. Each system pairs the TPUs with custom host processors from AMD; Unite.AI identified them as fifth-generation AMD EPYC chips, the family covered in our story on AMD's EPYC Venice launch.
Google pitches the 8tx for jobs that touch private data: fine-tuning a model on a company's own records, distillation, and training agents with reinforcement learning. It also uses a lower-latency network design than the TPU 8t, and Google says latency is up to two times lower, according to Unite.AI.
Protecting data while it is in use, often called confidential computing, is a selling point for regulated buyers such as banks, hospitals and government agencies. IDC analyst Dave McCarthy told Data Center Knowledge that Nvidia is still working out its own approach in that area.
TPU 8i Max: two inference chips in one package
The TPU 8i Max combines two TPU 8i chips in a single package. Both reports said that doubles compute, on-chip SRAM and high-bandwidth memory (HBM) bandwidth compared with one TPU 8i, while fitting in the same rack space. That means more work per square foot and per watt of data center floor.
Unite.AI put the TPU 8i at 288 GB of HBM and 384 MB of on-chip SRAM. Data Center Knowledge quoted Google Cloud's Mark Lohmeyer giving an HBM figure for the 8i Max that the two reports do not reconcile, so we have left it out. For why that memory matters, see our HBM explainer.
Google says the 8i Max is meant for jobs with very long inputs, such as many agents working across one codebase or a model reading hours of video at once. Those jobs depend on a large context window, which is limited by how much fast memory sits next to the chip.
Why it matters
Google is the largest cloud company building its own AI chips at scale, and every TPU it sells is compute a customer did not buy from Nvidia. Letting some customers put the TPU 8tx in their own buildings moves Google into a market Nvidia has had largely to itself: AI hardware for organizations that cannot send sensitive data to a public cloud. The launch comes in the same week Nvidia's chief executive resigned (see our report), and it follows Google's move to make its chips easier to program with PyTorch, the most widely used AI framework (our story).
Sources
- Google Adds TPU 8tx, 8i Max for Enterprise AI, Data Center Knowledge, November 9, 2026
- Google Adds TPU 8tx and TPU 8i Max to Eighth-Generation Lineup, Unite.AI, November 9, 2026
GoogleAI chips and data centersNvidia
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.