Cloud Services Compute Service

Google Expands Eighth-Generation TPU Lineup With Confidential and Dual-Chip Systems

Google has expanded its eighth-generation Tensor Processing Unit lineup with two systems aimed at different AI workloads: TPU 8tx, which adds confidential computing to TPU 8t silicon, and TPU 8i Max, which combines two TPU 8i chips in one package for memory-intensive inference. The additions join TPU 8t for training and TPU 8i for inference, introduced at Cloud Next in April 2026.

Google says this is the first year it has introduced four TPU systems. Rather than building each system around new silicon, the company has adapted existing chips through changes to packaging, memory, networking and security. Google says this modular approach lets it tailor configurations to specific workloads in a matter of months. It also reports that TPUs now have nearly 100 customers, spanning frontier AI labs, trading firms, retailers and universities, and says some customers can deploy the systems in their own data centers.

TPU 8tx adds hardware-based protections

TPU 8tx uses the same silicon as TPU 8t and pairs it with fifth-generation AMD EPYC host CPUs. Google says the system builds confidential computing into the hardware to isolate and protect data and model weights while they are in use. Listed protections include PCIe link encryption, cryptographic attestation and secure memory enclaves.

The configuration is intended for regulated enterprises and government organizations running proprietary or frontier models and agents alongside sensitive information. Google positions it for secure inference and smaller-scale training, including fine-tuning on private data, distilling a large model into a smaller one and training agents through reinforcement learning. The goal, as described by Google, is to let teams perform those tasks without moving sensitive data between separate clusters.

TPU 8tx also uses Boardfly, the low-latency network Google developed for TPU 8i. Google claims Boardfly provides up to twice the latency reduction compared with TPU 8t, with a near-zero performance penalty. The company has not published availability timing or pricing for TPU 8tx.

TPU 8i Max doubles key package resources

TPU 8i Max targets workloads that need to handle extensive context, such as multiple agents examining a codebase or processing hours of video. It places two TPU 8i chips on a single package, doubling compute, on-chip SRAM and HBM bandwidth relative to TPU 8i.

Google listed 384MB of SRAM for TPU 8i at launch, which means the doubled figure for an 8i Max package is 768MB. Google has not disclosed the package’s HBM capacity. The company says fitting twice the silicon into the existing rack footprint increases compute per square foot and per watt. As with TPU 8tx, Google has not announced customer availability or pricing.

Google maps the four systems to distinct stages of agent workloads: TPU 8t for scale-up training, TPU 8i for low-latency reasoning, TPU 8tx for confidential computing and TPU 8i Max for large context windows and high compute density per rack. The descriptions indicate how Google is segmenting its TPU offerings by workload, while leaving deployment timing for the two new systems unspecified.

Software additions address memory and cluster operations

Alongside the hardware, Google open-sourced TPU Sync, a C++ library that treats on-chip HBM, host DRAM, shared storage and the data center network as a unified memory hierarchy. It can tier key-value (KV) caches off-chip, stream context from prefill nodes to decode nodes and synchronize reinforcement-learning weights from training systems to samplers.

Google reports that TPU Sync reduces KV transfer latency by more than 90 percent and cuts time to first token by an average of 90 percent on open-weight mixture-of-experts models with 235 billion to 480 billion parameters. Those are company-reported results for the specified models; the announcement does not provide additional benchmark methodology.

Cloud Pathways, the Google DeepMind runtime behind Gemini, is now generally available on Google Kubernetes Engine. It uses one controller to coordinate TPU systems across multiple slices for pretraining, reinforcement learning and multi-host inference. Google says it can keep long-running jobs operating through hardware disruptions using sub-minute slice hot-swapping and elastic resharding.

GKE dynamic slicing is also generally available. It uses optical circuit switching to assemble TPU cubes into larger topologies or divide them into smaller, isolated workloads without recreating node pools. Google says the feature can reduce job startup times by up to five times and improve fault recovery speed by up to 4.5 times.

A TPU agent suite is in preview and integrates with coding tools including Claude Code, Cursor and Google Antigravity to help onboard workloads, diagnose bottlenecks and autotune kernels. Google has also introduced an open-source XLA Serving Engine, which serves Python-trained models from a C++ runtime across TPUs, GPUs and CPUs. Together, these software releases extend the announcement beyond chip configurations to the management, performance and serving tools used to operate AI workloads.