Google AI Updates: November 10, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Google Adds TPU 8tx for Confidential Inference and TPU 8i Max With Double the Compute and HBM Bandwidth of TPU 8i
Google Cloud / TPU. Google announced two additions to its eighth-generation TPU portfolio on November 9, bringing the year’s total to four systems alongside TPU 8t (training) and TPU 8i (low-latency reasoning). TPU 8tx is built on the TPU 8t silicon with 5th Generation AMD EPYC CPU hosts and is aimed at confidential inference: Google says it keeps data and model weights protected while in use, offers up to 2x lower latency than TPU 8t through the Boardfly network, and carries a near-zero performance penalty. TPU 8i Max packages two TPU 8i chips together, giving double the compute, SRAM and HBM bandwidth of TPU 8i within the existing rack footprint. The post also lists software changes for developers, including a TPU agent suite in preview that plugs into Claude Code, Cursor and Google Antigravity, an open-source XLA Serving Engine, and Agent Compute Engine, which suspends and resumes idle agent sessions. Google gave no pricing or general availability dates in the post. Source
2. TPU Sync Open-Sources a Tiered KV Cache and Weight-Transfer Layer for TPU Inference and RL, Cutting KV Transfer Latency Over 90%
Google Cloud / TPU. TPU Sync is an open-source C++ library that treats TPU HBM, host DRAM, shared storage and the data center network as one memory hierarchy, moving TPU-tiled tensors with asynchronous DMA that overlaps with compute. Google reports that, in internal July 2026 tests on TPU v7x, it cut KV transfer latency by more than 90% and time to first token by an average of 90% for open-weight MoE models of 235B to 480B parameters in prefix-caching and disaggregated-serving workloads; for Qwen3 235B, median KV transfer fell from several seconds to about 230 ms. It integrates as a KV connector for vLLM-TorchTPU, SGLang-TorchTPU and llm-d on GKE, with Python wheels tpu-sync-torch and tpu-sync-jax. The RL weight-synchronization module is marked experimental, as is the POSIX storage tier that spills KV blocks to local SSD or Managed Lustre. Source
3. Cloud Pathways, the Single-Controller Runtime Behind Gemini, Reaches GA on GKE
Google Cloud / GKE. Google made Cloud Pathways generally available on GKE on November 9, with Preview support on Slurm TPU clusters through Cluster Director. Pathways uses a single JAX client (JAX_PLATFORMS=proxy with pathwaysutils) to orchestrate multiple TPU slices as one machine, with colocated Python sidecars and elastic training that reshards down to N-1 slices after hardware failures. It is available in all regions that support TPU v5e, v5p, Trillium (v6e) and Ironwood (TPU 7x), with container images supporting JAX 0.10.1 and later. Google says customers processed more than 130 million TPU chip-hours with Pathways on GKE over the previous 90 days. Source
4. GKE Dynamic Slicing for TPUs Goes GA, Forming Super-Slices and Sub-Slices Without Recreating Nodepools
Google Cloud / GKE. GKE dynamic slicing is now generally available for TPUs. It separates logical slice formation from physical nodepool provisioning using optical circuit switching: super-slicing stitches 16-VM 4x4x4 TPU cubes into larger topologies, and sub-slicing partitions a single cube into smaller isolated slices for concurrent jobs. Google reports up to 5x faster job startup and up to 4.5x faster recovery, with unhealthy cubes swapped in one to two minutes. For platform teams, the practical change is that a topology change no longer requires deleting and recreating a nodepool. Source
5. Google Cloud Describes an Open Kubernetes Stack for Post-Training and Serving Gemma With OpenRL, llm-d 1.0 and KubeRay
Google Cloud / GKE. A November 10 post, published around KubeCon North America, assembles Google’s CNCF Sandbox projects into one open-weight model workflow. OpenRL exposes four Python primitives (forward_backward, optim_step, sample, save_state), time-slices training and rollouts on the same accelerators, and ships only parameter deltas to rollout workers. llm-d 1.0, co-founded with Red Hat, IBM Research, CoreWeave and NVIDIA, adds prefix-cache-aware routing; Google says production deployments doubled KV cache hit rates from 35% to 70% and cut median time to first token by 70 to 85% on multi-turn agent workloads. On Harvey.ai’s open Legal Agent Benchmark, Google reports Gemma 4 accuracy rose from about 50% with standard SFT to 70% with OpenRL reinforcement learning loops. The CNCF donations of Agent Substrate, gVisor and OpenRL were covered on 2026-11-07. Source
6. Google Research Launches Language Explorer, an Interactive Map of Linguistic Diversity Built on LinguaMeta
Google Research. Google Research and Google DeepMind published Language Explorer on November 9, an interactive web platform for exploring language data worldwide, developed under the 1,000 Languages Initiative and built on the open-source LinguaMeta dataset. The post cites Google Translate support for nearly 250 languages and Gboard support for more than 900 language varieties, and points to open speech data collected through Project Vaani in India and Project Waxal in Africa. The tool is available online and also runs as an installation in the exhibition “Real Talk! Language. Power. World.” at the Museum of Communication Berlin, through February 2028. The post does not name any AI model behind the platform. Source