This observation covers 7:41 am on 9 November to 7:41 am on 10 November 2026 in Australia/Sydney time and includes five technical changes verified against primary material.
Google announces TPU 8tx and TPU 8i Max systems
Google Cloud announced two eighth-generation TPU systems on 9 November. TPU 8tx uses the same silicon as TPU 8t with fifth-generation AMD EPYC hosts and adds hardware-level confidential computing for secure inference and small-scale training. Google reports latency up to 50% lower than TPU 8t; no public independent reproduction was identified. TPU 8i Max places two TPU 8i chips in one package, doubling compute, SRAM and HBM bandwidth relative to a single 8i. The announcement also labels TPU Agent Suite as a preview, opens XLA Serving Engine and TPU Sync, and confirms that Cloud Pathways is generally available on GKE. Google and Reuters confirm the systems announcement, but the material does not provide complete regional order dates, so this is recorded as an announcement rather than full delivery.
K-EXAONE 2.5 becomes available through a free API
LG AI Research opened free API access to K-EXAONE 2.5 on 9 November for individuals, developers, enterprises and research institutions using an API key. Official documentation describes a text model with 750 billion total parameters and 31 billion active parameters, Korean and English support, and a combined input and output context limit of 524,288 tokens. Its OpenAI-compatible chat completions endpoint supports tool calling, structured outputs and adjustable reasoning effort. AutomationBench 74.6, SWE-Bench Pro 59.21, BrowseComp 78.0 and Claw-Eval 77.29 are developer-reported results without a public independent reproduction. The new fact is operational free API access; model weights were not opened with the announcement.
NVIDIA introduces GR00T 2 and releases Human-1M-100K
NVIDIA introduced the GR00T 2 world action model at CoRL on 9 November and released Human-1M-100K, a dataset containing 100,000 hours of egocentric human video. The research page says the model jointly predicts future video and robot actions and can work across grippers, dexterous hands and humanoids. Its zero-shot, few-shot and long-horizon manipulation results come from NVIDIA testing and partner evaluations and have not been independently reproduced. The research page states that code will be available in December. The current delivery therefore covers research material and the dataset; it does not establish that the code, weights and full training stack are all publicly available.
OpenShell begins its move to vendor-neutral CNCF governance
CNCF records that OpenShell was accepted as a Sandbox project on 26 September. On 9 November, NVIDIA published the new implementation details: repositories are to move from the NVIDIA organisation to the neutral ai-openshell organisation, SDKs, drivers and installation tools may be separated into dedicated repositories, and the project is adopting public RFCs, vendor-neutral governance and the Cedar policy language. This series previously recorded the availability of OpenShell 0.1.0 as an agent sandbox runtime; the current change concerns project control and maintenance. Repository migration and working-group formation remain in progress, and Sandbox status is below CNCF Incubating and Graduated maturity.
NVIDIA Dynamo publishes deployment recipes for Korean open models
On 9 November, NVIDIA published Dynamo recipes for A.X K2, K-EXAONE 2.0, Solar Open 2 and Motif-3 and made them available on GitHub. The recipes include Kubernetes manifests, model-download jobs, smoke tests and benchmark procedures, with GPU counts, cache and storage requirements for aggregated or disaggregated prefill and decode deployments on B200 systems. NVIDIA reports up to twice the baseline throughput for some configurations; this is a developer measurement without independent reproduction. The recipes are available, while deployment still depends on access to B200 resources, each model’s licence, hundreds of gigabytes of weights and cache storage, and lengthy start-up times.
Sources
Google Cloud, Introducing TPU 8tx and TPU 8i Max: Next-generation systems for secure, dense AI, 9 November 2026, https://cloud.google.com/blog/topics/ai-infrastructure/tpu-8tx-and-tpu-8i-max
Reuters, Google to introduce two variants of its in-house chips, 9 November 2026, https://www.reuters.com/business/google-to-introduce-two-variants-its-in-house-chips-bloomberg-news-reports-2026-11-09/
LG AI Research, K-EXAONE 2.5 API Available for Free, 9 November 2026, https://www.lgresearch.ai/news/view?seq=755
EXAONE API Documentation, K-EXAONE 2.5, no publication date shown; accessed 10 November 2026, https://docs.exaone.ai/en/docs/k-exaone-2.5
EXAONE API Documentation, Models, no publication date shown; accessed 10 November 2026, https://docs.exaone.ai/en/docs/models
NVIDIA Blog, Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation, 9 November 2026, https://blogs.nvidia.com/blog/robotics-simulation-advancements-corl/
NVIDIA Research, GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation, November 2026, https://research.nvidia.com/labs/gear/gr00t-2/
NVIDIA OpenShell Research, OpenShell is now a CNCF Sandbox project, 9 November 2026, https://nvidia.github.io/OpenShell-Research/dev-notes/posts/2026-11-09-donated-openshell-to-cncf/
Cloud Native Computing Foundation, OpenShell, records acceptance on 26 September 2026; accessed 10 November 2026, https://www.cncf.io/projects/openshell/
NVIDIA Technical Blog, Optimizing Korea’s Leading Open Models with New NVIDIA Dynamo Recipes, 9 November 2026, https://developer.nvidia.com/ko-kr/blog/optimizing-koreas-leading-open-models-with-new-nvidia-dynamo-recipes/
NVIDIA Dynamo, Recipes README, no publication date shown; accessed 10 November 2026, https://github.com/ai-dynamo/dynamo/blob/main/recipes/README.md
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.