Google Adds Two New TPUs: 8tx for Confidential AI and 8i Max for Huge Context Windows

Photo: Hippopx (public domain). Illustrative image.

What happened. Google expanded its eighth-generation TPU portfolio on Monday 9 November 2026, adding two new systems: TPU 8tx for secure and confidential inference, and TPU 8i Max for massive context windows and the highest compute density per rack. Google says it is the first time it has introduced four TPU systems in a single year. The first two — TPU 8t for large-scale training and TPU 8i for real-time reasoning — were introduced at Google Cloud Next in April.

TPU 8tx: confidential AI. The 8tx is aimed at banks, governments and other regulated customers who cannot afford to lose control of their data. It is built on the same silicon as TPU 8t, paired with fifth-generation AMD EPYC host processors, and uses AMD’s Secure Encrypted Virtualization to keep sensitive data and model weights protected and isolated even while they are in active use — the approach known as confidential computing. Google says it is suited to private fine-tuning and secure agent execution as well as small-scale training, and claims up to two times lower latency than TPU 8t with a near-zero performance penalty from the security layer. That latency figure is Google’s own, from its own tests; no independent measurement has been published, and Google’s announcement names no trusted-execution technology of its own and publishes no attestation design a security architect could review.

TPU 8i Max: beating the memory wall. The 8i Max addresses a different problem: the memory wall that stalls large workloads when processors wait on data. It puts two TPU 8i chips into a single package, delivering double the compute, double the on-chip SRAM and double the high-bandwidth-memory bandwidth of TPU 8i, in the same rack footprint. According to Network World, that works out to 768 MB of on-chip SRAM, 17.2 TB/s of memory bandwidth and 384 GB of HBM — 33% more memory than TPU 8i. The target is workloads with very large context windows and multiple cooperating AI agents, where memory capacity, not raw compute, is the bottleneck.

What Google did not say. There is no price, no general-availability date and no named customer for either system. Google’s Amin Vahdat, senior vice-president and chief technologist for AI and infrastructure, told Bloomberg the systems will arrive next year, in 2027. The Cloud TPU product page still labels TPU 8i as coming soon and does not list the two new configurations as broadly available. This is an announcement of capability, not of shipping hardware, and nothing in it should be read as a claim that Google Cloud customers can provision either system today.

Why it matters. Google is the only cloud provider that designs its own AI accelerators at this scale, and the TPU line is its answer to Nvidia’s dominance of AI compute. Splitting the eighth generation into four specialised systems — training, reasoning, confidential inference and huge-context inference — is a bet that the AI market is fragmenting by workload, and that the winner will be whoever has the right silicon for each job rather than one general-purpose chip. The confidential-computing angle is the more interesting of the two launches: if regulated industries are going to run frontier models on their own sensitive data, they need hardware that can prove the data stays private. Whether TPU 8tx can prove that will depend on the attestation design Google has not yet published.

Sources

• Google Blog — “Introducing TPU 8tx and TPU 8i Max” (9 November 2026): https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/?ref=explainx
• ShiftDelete — “Google TPU 8tx and 8i Max: Private AI and Huge Context” (9 November 2026): https://shiftdelete.net/en/google-tpu-8tx
• ReaderSpice — “Google TPU 8tx and 8i Max: Security, Memory and Availability” (9 November 2026): https://www.readerspice.com/2026/11/09/google-tpu-8tx-8i-max-confidential-ai-massive-context/
• Rackline News — “AMD EPYC Powers Google’s Confidential TPU 8tx” (9 November 2026): https://www.rackline.news/story/1yfrsx3?from=models