Google Cloud has added two processors to its eighth-generation Tensor Processing Unit lineup, targeting enterprises that need to protect sensitive AI workloads or run inference with large context windows. Announced Nov. 9, 2026, TPU 8tx adds hardware-level encryption to the company’s TPU 8t design, while TPU 8i Max combines two inference chips in a single package to increase compute and memory capacity.
The additions extend a product family Google introduced in April. The company positioned TPU 8t for AI model training and TPU 8i for inference. At the time, Google said the 8t could deliver up to 2.7 times the training performance of its predecessor, and that the 8i offered 80% better inference price-performance. Those figures are Google’s claims about the prior-generation comparison.
TPU 8tx targets protection for data in use
TPU 8tx uses the same silicon as TPU 8t, but adds hardware-level encryption intended to support secure inference and small-scale training. Google Cloud says the processor is paired with custom AMD CPU hosts to help protect and isolate sensitive data and model weights while inference is running.
That focus addresses confidential computing, a model of protection designed to safeguard information while it is being processed, rather than only while it is stored or moving between systems. Mark Lohmeyer, Google Cloud’s vice president and general manager of AI and computing infrastructure, described demand for confidential computing as one of the needs shaping the new products during a media briefing.
Dave McCarthy, group vice president at IDC, said confidential computing is becoming an emerging requirement for AI workloads. He noted that cloud providers have traditionally emphasized protections for data at rest and in transit, while protection during use presents a distinct challenge. McCarthy said a strong confidential-computing capability could help Google Cloud differentiate its offering; his assessment reflects an analyst view, not a guarantee of product advantage.
8i Max increases compute and memory bandwidth
Google Cloud designed TPU 8i Max for inference workloads, including agentic AI applications that need to process large context windows. To make the variant, the company combines two TPU 8i chips in one package. Google says the resulting configuration doubles compute, SRAM and high-bandwidth memory (HBM) bandwidth compared with a single TPU 8i. Lohmeyer said the 8i Max provides 384 GB of HBM.
The additional memory capacity and bandwidth are intended to ease constraints that can arise when AI applications work with extensive context. Lohmeyer compared the approach to adding lanes to a highway. The product is aimed at enterprises seeking to run such workloads while maintaining performance, though the announcement did not provide independent benchmark results or specific performance figures for TPU 8i Max.
Google Cloud also said the 8i Max can place more silicon within an existing rack footprint, potentially increasing computing capacity per square foot and per watt in its data centers. McCarthy said that could help the provider add capacity without immediately adding racks or constructing additional data centers, although he framed that outcome as a potential benefit rather than a stated deployment result.
Availability and market reach
The announcement comes as cloud providers and AI customers seek more specialized processors for different stages of AI work. McCarthy said Google’s release of the 8i and 8t earlier in the year, followed by two related variants within the same calendar year, demonstrates a rapid development pace. He also said custom silicon could help Google Cloud distinguish itself from competitors.
Google’s TPU strategy is not limited to its own cloud infrastructure. Oracle recently announced that it would make Google TPU chips available to Oracle Cloud Infrastructure customers. Lohmeyer also said Google Cloud would allow select customers to deploy TPU 8tx in their own data centers. The announcement did not specify broader availability dates or the terms for those deployments.
Together, the two products address different constraints: 8tx adds an encryption-focused design for workloads requiring protection while in use, while 8i Max expands the compute and memory resources available for inference. Their practical value for enterprises will depend on workload requirements and access, as well as on performance and deployment details beyond the specifications disclosed so far.
