Strategic PartnersLatest editionIssue 66
NVIDIA
A KubeCon-driven edition on Kubernetes-based AI factory software and independent scrutiny, with no new SLED deployment evidence.
Full research summary
A KubeCon-driven edition on Kubernetes-based AI factory software and independent scrutiny, with no new SLED deployment evidence. NVIDIA contributed OpenShell to the CNCF Sandbox and announced NVKE and NVKR (public preview, not recommended for production) and Snapshot, which it reports restores GPU inference replicas 8x to 36x faster than cold starts on B200 with VAST storage. Independent reporting stresses that Sandbox is an entry tier and not a security audit, and an Oct 30 independent benchmarking paper finds B300 is not a uniform successor to B200, with a development vLLM build alone changing time-to-first-token by 2.1x. The common thread is that preview-stage tooling and vendor-measured results need local, workload-specific validation.
Read the edition Previous: Issue 65, November 9All NVIDIA editions
What this stream covers
NVIDIA AI platforms, software, infrastructure, reference architectures and ecosystem. Separate vendor claims from independent deployments, benchmarks and limitations. Explain Government/Education applicability where supported, without forcing irrelevant news into a SLED use case.
- Evidence records
- 5
- Cross-source patterns
- 2
- Also published November 10
- AnthropicAWSCampus OperationsCiscoCitizen Engagement & ServicesCrowdStrikeDellEmergency ServicesEverpureGoogleHealth & Human ServicesK–12Local GovernmentMicrosoftNetAppOpenAIPalo Alto NetworksPublic SafetyRed HatResearch UniversitiesServiceNowState GovernmentStudent SuccessTransportationUtility ManagementVAST Data
- Subscribe
- Atom feed for NVIDIA
What the evidence supports, and what it leaves open
2 patterns this edition will stand behind, each carried by at least two independent sources. Every one arrives with the question it has not answered.
- New open source AI-factory and agent-security tools arrive at preview or sandbox maturity
Operating questionWhat maturity, support and independent assurance evidence will we require before moving an open source preview or sandbox project into production?
- Measured GPU results depend heavily on software build, storage and workload
Operating questionWhich of our own workloads, software builds and storage configurations must we benchmark before relying on published GPU performance figures?
Research through your lens
Every resource includes source evidence and takeaways for all three roles.
Evidence in this micro-vertical
425 resources
Follow the outcomes
425 resources across outcomes in your selection. Counts include all outcomes.
Refine by evidence type and topic
Nvidia hands its AI agent sandbox OpenShell to the CNCF
Independent reporting on the same-day OpenShell donation says the move changes who steers the project, not whether developers can use it today, and that CNCF Sandbox is an entry tier for experimental projects, signaling neutral hosting rather than a security audit. It notes NVIDIA's adoption figures are its own and unverified, that NVIDIA does not say whether any independent review of OpenShell's isolation has occurred, that no date or shipping form is given for the OpenShell and Agent Substrate integration, and that NVKE/NVKR partner integrations are plans rather than shipping products. Partner Snapshot numbers are described as single-hardware-generation benchmarks without stated model sizes or test conditions.
Limitations & uncertainty
Secondary reporting derived mostly from vendor material; no hands-on testing; author-level independence from NVIDIA is not established beyond editorial framing.
Manage Kubernetes Clusters for GPU Infrastructure with NVIDIA Kubernetes Engine
NVIDIA announced two open source projects, NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR), for running managed Kubernetes services on GPU infrastructure. NVKR builds version-pinned node and container images and records build inputs; NVKE provisions and manages clusters through a REST API and controllers, with servers supplied by NVIDIA Infra Controller. Both are Apache 2.0, in public preview, and NVIDIA states they are not yet recommended for production workloads. NVKE is described as a Certified Kubernetes AI Platform and Installer, conformance tested on Kubernetes 1.36. Firmus and SK Telecom are integrating them; Mirantis, Rafay, Spectro Cloud, OpenNebula Systems, vCluster and Netris are incorporating components.
Limitations & uncertainty
Vendor-authored; public preview; not recommended for production; walkthrough prerequisites are substantial (multiple CPU servers, GPU pool, NICo installation, build host).
NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure
At KubeCon, NVIDIA said it is contributing OpenShell, its open source runtime for securing and governing autonomous agents, to the CNCF as a Sandbox project, and contributing to Agent Substrate, a Google-led CNCF Sandbox framework for long-running agent workloads on Kubernetes. It also announced three DSX AI Factory Platform components: NVKE and NVKR (public preview) and Snapshot. NVIDIA says hundreds of organizations have used OpenShell and that it has received tens of thousands of contributions in under a year. An OpenShell and Agent Substrate integration with Google is described as a collaboration, with no date given.
Limitations & uncertainty
Vendor announcement. Sandbox is CNCF's entry tier. No benchmarks, security audit or SLED deployment evidence in this post. Integration with Agent Substrate has no stated date.
Start and Restore GPU Workloads in Seconds with NVIDIA Snapshot
NVIDIA Snapshot is an open source checkpoint and restore system for GPU workloads on Kubernetes. On NVIDIA B200 GPUs with VAST NFS storage, NVIDIA reports restores of 3.6 to 42.7 seconds, 8x to 36x faster than cold starts of about 87 seconds to 18.5 minutes across vLLM, SGLang and TensorRT LLM, including models with checkpoints up to about 1 TB over eight GPUs. Checkpoint creation took 15 seconds to 5.5 minutes, and checkpoint plus restore was shorter than one cold start in every tested configuration. Design partners report Crusoe cutting GLM-5.3 readiness from about eight minutes to 31 seconds and Saturn Cloud cutting vLLM startup from 2 to 8.5 minutes to 6 to 34 seconds, both on HGX B300.
Limitations & uncertainty
Vendor-run benchmarks on one GPU model and one storage backend; same-GPU-and-driver restore requirement; workloads must meet compatibility requirements; image pull and provisioning time excluded; not independently reproduced.
AI demand rises: Deutsche Telekom expands AI factory in Munich
Deutsche Telekom says the first stage of its Munich Industrial AI Cloud, live since February 2026 and described as adding 50 percent to German AI capacity, is fully booked seven months later. It will add NVIDIA Blackwell Ultra B300 systems alongside B200 and RTX PRO 6000 Blackwell Server Edition GPUs, nearly another quarter of a billion in investment, and raise capacity to 20 megawatts. Users named include Siemens, Daimler Truck and research organizations; the company says the public sector can use it. A further Bavaria expansion is under discussion but no final investment decision has been made.
Limitations & uncertainty
Operator marketing with political endorsements. 'Fully booked' is not quantified; the minister's quote also says 'after one year', conflicting with the seven-month figure, which shows imprecision in the quoted claims. Not independently verified.
- Source
- MIT Technology Review (The Algorithm newsletter)
- Published
- November 9, 2026
The backlash about data centers might really just be about electricity
Reporting from Virginia's Data Center Alley argues that opposition to data centers is driven largely by electricity: grid-connected facilities (an average 54 MW in Loudoun County) have led Virginia to require data centers to pay the full cost of new transmission lines, and developers expect rules that could cut them off from grid power at peak. Larger AI sites (300 MW at minimum, ideally about 6,000 MW) tend to use on-site gas turbines, which bring air-pollution and neighbor concerns. A developer says industry ambitions are constrained less by public opposition than by a lack of electricity.
Limitations & uncertainty
Commentary and reporting by one author; not an audit or study. Contains no NVIDIA-specific information. Some claims (for example the Anthropic turbine purchase) are described as reported. Relevance to SLED is by inference.
Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation (GR00T 2, Isaac Lab 3.0, new benchmarks)
NVIDIA introduced GR00T 2, an open world action model built on Cosmos, plus Isaac Lab 3.0 (Newton physics integration) and Isaac Lab-Arena 0.4 developer preview with new DexBench and VABAR benchmarks. NVIDIA reports GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations and reached 71% average success across 50 tasks with one demonstration per task, 95% with five. It also reports up to 2.5x training throughput for Isaac Lab 3.0 with Newton versus Isaac Lab 2.3 with PhysX in a dexterous-manipulation benchmark. Named partner uses (Nexuni, Foxconn, Siemens, MicroAGI) are testing or evaluation work; no production rollout is stated.
Limitations & uncertainty
Vendor announcement; numbers not independently verified in this edition (see the independent RoboColosseum evaluation of the prior-generation GR00T N1.7). The 2.5x throughput is for one benchmark against one earlier release. Conference-week marketing context.
- Source
- NVIDIA Investor Relations (GlobeNewswire press release)
- Published
- November 9, 2026
NVIDIA, Rebellions and SK Telecom Announce Collaboration to Advance Sovereign AI Infrastructure With NVIDIA NVLink Fusion
NVIDIA, Rebellions and SK Telecom announced a collaboration, supported by Korea's Ministry of Science and ICT, in which Rebellions will connect its R200 NPUs to NVIDIA's platform through NVLink Fusion and the MGX rack architecture, and SK Telecom plans to deploy the systems in its AI data center for sovereign AI services including the national AI for All project. The companies plan a disaggregated-inference approach using NVIDIA GPUs for prefill and Rebellions NPUs for decode. R200 systems are expected to be ready for deployment in 2028.
Limitations & uncertainty
Vendor press release with executive quotes; no independent verification; deployment is three years away and all performance statements are aspirational.
- Source
- NVIDIA Kubernetes Engine documentation
- Published
- Undated documentation page; NVKE preview surfaced publicly Nov 6–7, 2026 per independent reporting
NVIDIA Kubernetes Engine (NVKE) preview documentation: supported features and known limitations, including plain-HTTP API, no authorization, and no management-plane backup or upgrade
NVIDIA's NVKE documentation describes an open-source preview management system that lets an AI-factory operator create isolated Kubernetes workload clusters (own control plane, etcd, node pools) on its own GPU bare metal through NVIDIA Infra Controller (NICo). The known-limitations page states NVKE is not recommended for production; the API serves plain HTTP on port 30080 with authentication off after install; NVKE does no authorization, so any admitted caller can create or delete every workload cluster; customers receive only a non-revocable admin kubeconfig; and the management cluster has no self-management, upgrade, or backup/restore procedure.
Limitations & uncertainty
Vendor documentation for a preview; behavior and APIs may change. Page is undated, so the publication date is left null. No measured outcomes, customer deployments or U.S. public-sector use are described.
CNBC: Nvidia investigates CEO's relationship with a woman affiliated with Megaspeed, a customer accused of helping restricted Chinese companies obtain chips
CNBC reports, citing New York Times and Financial Times reporting, that U.S. intelligence agencies have expressed concern about a businesswoman affiliated with Megaspeed, an NVIDIA customer accused of helping restricted Chinese companies acquire NVIDIA AI chips; NVIDIA says Megaspeed was previously investigated for evading export controls. CNBC notes the U.S. restricts exports of NVIDIA's most advanced chips to China, that NVIDIA reports Q3 earnings Nov 17 and holds a Washington conference at month end.
Limitations & uncertainty
Allegations are attributed and unproven; CNBC relays NYT/FT reporting rather than primary evidence; investigation outcome unknown. Personal details are deliberately omitted here as not relevant to SLED applicability.
NVIDIA board opens independent investigation after reports its CEO was the target of a foreign intelligence effort
On Nov 8 NVIDIA's board said a committee of independent directors, with outside counsel, will investigate recent reporting that CEO Jensen Huang was the target of a foreign intelligence effort. The board says it has no reason to believe national security or NVIDIA secrets were compromised and expresses confidence in Huang. This is NVIDIA's own statement; it contains no findings and no timeline.
Limitations & uncertainty
Vendor statement only; outcome of the inquiry is unknown, and independent confirmation of facts comes from separate reporting (see CNBC item). No claim here about product or supply-chain impact.
- Source
- University of Utah CHPC News
- Published
- November 2026 (CHPC newsletter)
Utah CHPC: Redtail AI Factory formally launched; Winter 2027 allocations open Nov 9 (15k GPU-hours/quarter)
CHPC's November 2026 newsletter states Redtail AI Factory formally launched Oct 6, 2026, will move to standard quarterly allocations beginning Jan 1, 2027, opens request window Nov 9–Dec 1, 2026 for Winter 2027, and caps new projects at 15,000 GPU hours and/or 100,000 CPU core hours per quarter, allocated by project rather than research group.
Limitations & uncertainty
Operational announcement, not an outcome evaluation. Hardware counts and NVIDIA role are not restated in this newsletter (covered in earlier archived Redtail materials). 'Last Updated' footer noise on page; substance is the allocation calendar.
heise online: Nvidia previews its own Kubernetes engine (NVKE) — tight NICo/BlueField coupling, two stacked management layers, missing access control
heise reports that NVKE, published as an Apache 2.0 preview on GitHub, lets GPU data-center operators offer EKS-like managed Kubernetes, but is tightly coupled to NVIDIA's stack: provisioning runs through NICo, which itself needs a three-node Kubernetes cluster, so real deployments run two stacked management layers; NICo's default tenant isolation uses BlueField DPUs; the API accepts unencrypted HTTP on port 30080 and lacks access control; GPU drivers come separately via GPU Operator; and NVIDIA has given no production date. heise notes Gardener and Kamaji already handle similar control-plane hosting.
Limitations & uncertainty
Reporting based on project documentation rather than measured deployment; article is a translation of a German original; does not evaluate GPU workload performance.
NVCRE GitHub: workload certification reports failing nodes but never cordons; not continuous monitoring or a benchmark leaderboard
The NVIDIA Cluster Readiness Engine repository states NVCRE runs real training and NCCL workloads across topology-aware node groups, reports every bad node with a reason, and leaves quarantine to the platform: it never cordons, taints or otherwise modifies a node. Scope and non-goals say it is not continuous monitoring, not a general training platform, and not a benchmark leaderboard—measurements exist to find faults, not to rank hardware. Newly relevant as the companion limit statement to NVIDIA's Nov 5 DSX OS configure-certify-operate post.
Limitations & uncertainty
Vendor documentation; no independent measurement of detection quality; living README may change.
NVSentinel docs: default install is monitoring-only; quarantine, drain and remediation are opt-in as confidence grows
Current NVSentinel documentation describes a Kubernetes GPU fault detection and remediation system that can cordon, drain and trigger break-fix. It claims deployments on AWS, GCP, Azure and OCI up to 1,100+ nodes and about 40,000 GPUs. Critically for operators, by default only health monitoring is enabled and is described as safe because it only observes and reports; fault quarantine, node drainer and fault remediation are enabled via Helm values as confidence grows. Newly relevant because the Nov 5 DSX OS post promotes enabling remediation after monitoring trust is established.
Limitations & uncertainty
Vendor docs; scale claims unverified independently; version Latest v1.30.0 at retrieval.
NVIDIA: open-source Aerial and Sionna 5G/6G tools pass 1 million downloads and 8,000+ research institutions; university and partner results on semantic comms, neural receivers and RAN power savings
NVIDIA says developer adoption of Aerial and Sionna grew 10x in a year, its open 5G/6G tools passed 1 million downloads, and more than 8,000 research and academic institutions use them. It summarizes partner results: King's College London reported up to 52.6% faster robot task completion with semantic communications on Jetson AGX Orin; SoftBank's neural receiver on Aerial ACAR beat a conventional receiver at all 15 tested interference settings (max ~60% uplink gain); ETH Zurich cut block error rates by more than one-third with locally tuned receivers; University of Leeds reported 45.6% RAN power reduction on a DGX Spark network-in-a-box; CTTC cut simulation time from ~3 hours to 10–20 minutes.
Limitations & uncertainty
Vendor blog aggregating others' results; methods, baselines and sample sizes are not shown; several results are 'controlled tests' or 'early evaluation'. Download and institution counts are NVIDIA-reported. Independent work (see neural-receiver calibration-drift preprint) shows in-distribution gains may not hold outside the training envelope.
NVIDIA Reshard collective: NeMo RL weight refit 2.7x–5.7x faster than gather-and-broadcast on DGX H100/GB200
NVIDIA's Nov 6 post describes NCCL M2N Reshard as a DTensor-to-DTensor weight-refit path for disaggregated RL post-training, reporting 2.7x–5.7x lower refit latency versus gather-and-broadcast in NeMo RL on DGX H100 and GB200 across models including Qwen 3 235B and DeepSeek v3 671B, with hybrid Reshard for dominant FFN weights.
Limitations & uncertainty
Vendor blog; results are NeMo RL reference measurements on NVIDIA DGX systems, not third-party replication. Hybrid path still uses gather-and-broadcast for some weights. NeMo RL refit guide URL cited in post returned 404 when checked during this run.
NVIDIA tutorial: FLUX 3 Image Klein on Blackwell with TensorRT LLM VisualGen reports up to 3.46x faster text-to-image on one B200, with quality trade-offs
NVIDIA reports combined optimizations (NVFP4, fused kernels, torch.compile, FP8 attention, sparse attention, step caching) deliver up to 3.46x faster text-to-image and 4.45x faster image-to-image for FLUX 3 Image Klein Base versus uncompiled BF16 at 2048x2048 on one B200. Individual options trade fidelity: LPIPS scores are reported for FP8 attention and caching. Eight B200s cut median T2I latency from 14.90 to 3.00 seconds for Base.
Limitations & uncertainty
Vendor-run measurements on B200 only; quality measured with LPIPS on selected settings; no independent replication; speedups vary by option and not all combinations preserve fidelity.
- Source
- GitHub (NVIDIA/AI-Factory-Operations-Agent)
- Published
- Undated repository README (linked from the Nov 4, 2026 NVIDIA blog)
NVIDIA's AI Factory Operations Agent repository is Apache-2.0 but experimental, and its Helm chart and images need an early-access invitation to a private NGC registry
The README labels the project an experimental release under Apache-2.0. Supported workflows are read-only Kubernetes inspection, Prometheus/VictoriaMetrics queries and Grafana dashboards, Slurm job root-cause analysis, optional cluster-management integrations and sandboxed command execution. Browser and MCP clients do not receive cluster credentials. In the early-access stage the published Helm chart and container images require an invitation to a private NGC registry; an nvcr.io login alone does not grant pull access. It offers browser, HTTP API, CLI and MCP interfaces, including a skill file for coding agents.
Limitations & uncertainty
README has no release date; experimental status; access gating may change. Not an independent evaluation.
NVIDIA on Hugging Face: four open-weight Cosmos 3 Nano checkpoints for interactive world simulation (driving multi-sensor data, streaming camera/LiDAR, simulator-to-RGB, bimanual robot action-to-video)
NVIDIA released four post-trained Cosmos 3 Nano checkpoints (16B trainable parameters each, OpenMDW1.1 license): Transfer-Auto for offline multi-view driving data with joint RGB–LiDAR, Sim-Auto for streaming closed-loop driving observations, Sim-Transfer for turning physics-simulator controls into RGB, and Sim-Bimanual for predicting video from two-arm robot actions. The post itself says the open problem is measuring where these models faithfully support policy learning and evaluation and understanding their limits; inference requires Linux, an NVIDIA GPU, Docker and Container Toolkit, with the published vLLM-Omni setup tested on one GB200.
Limitations & uncertainty
Vendor publication; no independent benchmark of these exact checkpoints was found. The model cards date V1.0 to Oct 20, 2026 per secondary reporting; the blog date is Nov 5. Tested configuration (one GB200) is not a minimum requirement.
NVIDIA co-design blog: Dynamo KV routing, disaggregation, and up to 30x throughput/MW claims for Vera Rubin NVL72
NVIDIA's Nov 5 co-design post positions five levers—NVFP4, wide expert parallelism, Dynamo KV-aware routing/disaggregated serving/KV Cache Runner via NIXL, Vera CPU staging, and CMX/BlueField-4 storage—as a stacked memory system, claiming Vera Rubin NVL72 can deliver up to 30x throughput per megawatt versus GB300 NVL72 from the same 288 GB HBM per GPU.
Limitations & uncertainty
Pure vendor architecture claim. Throughput/MW figure is NVIDIA comparison under its agentic assumptions. Storage vendor validation numbers are not campus production measurements. Independent Dynamo tests in this edition show routing and disaggregation benefits are workload- and link-dependent.
NVIDIA DSX OS: configure with AICR recipes, certify with NVCRE workload tests, operate with NVSentinel (monitoring-first)
On Nov 5 NVIDIA published a DSX OS walkthrough that ties three open-source Apache-2.0 projects into a Kubernetes GPU lifecycle: AICR for version-locked configuration recipes rendered to Helm/Argo CD/Flux/Helmfile; NVCRE for NCCL and training burn-in with optional performance thresholds (none configured by default); and NVSentinel for continuous health monitoring that can cordon, drain and remediate when enabled. The post states NVSentinel installs in monitoring mode by default and will not cordon, evict or remediate until operators enable those features. Lifecycle validation can gate new or repaired nodes until NVCRE tests pass; in-place driver/operator/fabric changes do not auto-trigger revalidation.
Limitations & uncertainty
Vendor walkthrough; no published accuracy or goodput results for NVCRE thresholds; remediation behavior depends on local Helm values.
Report: Nvidia explores retrofitting Apollo-backed Brightspeed telecom buildings into small AI data centers to avoid new-build opposition
Relaying The Information, Yahoo Finance reports NVIDIA approached Apollo about converting existing telecom buildings, centered on Brightspeed's 130+ sites with about 1 GW of available power, into a network of small 'inference' sites across as many as 20 states, with possible retrofits using Vera Rubin racks from Q1 2027. Retrofits would still need local permits for specialized cooling, and buyers of the capacity are unclear.
Limitations & uncertainty
Unconfirmed, relayed reporting; NVIDIA has not announced the plan; Brightspeed is described as nearing bankruptcy; purchasers and counts of sites unknown; stock-market commentary in the article is not evidence of technical outcomes.
Preprint FastAFD: attention-FFN disaggregation gains up to 1.17x over the best wide-EP baseline on GB200 NVL72, and not on every model
The preprint builds FastAFD for GB200 NVL72 and compares attention-FFN disaggregation with the best vLLM wide expert-parallel setting. AFD gains up to 1.17x on Qwen3 and 1.16x on Qwen3.5 but only matches wide EP on MiniMax-M2.5. An analytical model predicts AFD throughput within 3% from colocated profiles, so operators can estimate payoff before deploying.
Limitations & uncertainty
Preprint, not peer reviewed. Authors' systems and models; only abstract inspected here, full methods not independently reproduced; the 2.48x Groq LPX/Vera Rubin figure is a projection, not a measurement.
Karnataka minister welcomes Dayananda Sagar University AI First Factory with 20 NVIDIA DGX B200 nodes
On Nov 4 The Bengaluru Live reported that Karnataka Home/IT minister Priyank Kharge welcomed Dayananda Sagar University's academic AI First Factory built with NVIDIA. Pro-Chancellor Dr Premchandra Sagar said the facility has a high-performance computing system with 20 NVIDIA DGX B200 supercomputing nodes so students can develop, train and deploy large-scale AI on real infrastructure rather than textbooks alone. The piece is a ministerial/university announcement without measured research or teaching outcomes.
Limitations & uncertainty
Announcement journalism; university self-description; non-U.S.; no utilization or learning-outcome metrics.
NVIDIA describes the AI Factory Operations Agent it runs on its own DGX SuperPOD clusters: Nemotron-based sub-agents that investigate read-only and recommend, with SREs deciding
NVIDIA says it built the AI Factory Operations Agent to reduce manual work for its own SREs on DGX H100, B200 and GB300 production clusters. A NemoClaw-based orchestrator routes chat questions to Job Recovery, Hardware, Observability, Research, Kubernetes and Cluster Management agents over MCP, using locally served Nemotron models. Kubernetes access goes through a server-side tool with read-only credentials; the post shows the agent allowed to get pods but denied secrets and pod create/update. NVIDIA lists five lessons: clear agent boundaries, read-only first, cross-layer correlation, precise MCP tools and transparent reasoning. It says it measures RCA accuracy and SRE agreement but publishes no figures.
Limitations & uncertainty
Vendor blog about NVIDIA's own clusters; no quantitative results; orchestration framework is experimental and APIs may change between releases.
NVIDIA NeMo Compass and Eval Author: turn agent-trace failures into tests and harness fixes; Dassault example claims 28% score gain
On Nov 4 NVIDIA published a NemoClaw agent-optimization tutorial using NeMo Compass (trace analysis), NeMo Eval Author (Harbor/NeMo Gym evaluations) and harness edits. The workplace-assistant example withholds two held-out tasks from coding agents and reports development passes rising from 3/6 to 6/6 and held-out from 4/6 to 6/6 after a SOUL.md change. A Dassault Systèmes LEO car-door example claims malformed tool calls fell from 142 to zero, a 28% relative evaluation-score improvement and 35% fewer tool calls across six scenarios after an overnight unattended loop. The post itself warns to keep held-out cases hidden to avoid overfitting.
Limitations & uncertainty
Vendor and partner claims; small example task counts; overnight loop results not independently audited.
CUDA 13.5 makes Vera Rubin support generally available and adds MPS v3 GPU-sharing controls: weighted timeslicing, SM borrowing/lending and cgroup-based routing
CUDA 13.5 moves Vera Rubin support from early developer access to general availability and extends Windows on Arm support to DGX Station for Windows. Multi-Process Service v3 adds weighted timeslicing, SM partition load balancing with borrowing and lending, cgroup v2 subtree-to-namespace routing, per-namespace SM partition assignments, and a polling-based fallback for memory accounting when cgroup notifications are missing or unreliable. Other changes cover CUDA Tile multi-GPU communication, cuda.core 1.2 and CCCL 3.5, including reproducible floating-point scans.
Limitations & uncertainty
Vendor announcement; feature claims are untested here; borrowing behaviour under contention is not quantified.
Fermaca becomes Mexico's first NVIDIA Cloud Partner; phased AI factory targets 60 MW by end-2027
Fermaca announced it is the first NVIDIA Cloud Partner based in Mexico and unveiled an AI factory platform, starting with Fermaca Digital Origins in Mexico City with an initial NVIDIA Blackwell Ultra (B300 NVL72) cluster and a Vera Rubin roadmap. Inventec supplies compute and DDN storage. It targets 60 MW of total facility power for operational AI facilities by Dec 31, 2027, and a longer-term ambition of up to 1 GW of IT capacity at Durango.
Limitations & uncertainty
Company press release; plans, not delivered capacity; no named customers, timeline for first service or pricing; Mexico-specific and not transferable to U.S. SLED without caveats.
ITMAX: SamurAI V2 on NVIDIA VSS and Cosmos runs on 1,000+ of 16,000+ Kuala Lumpur cameras; claims 70%+ better alert verification and 26% shorter peak corridor times
ITMAX, a Smart City-as-a-Service provider, described its NVIDIA stack: RTX PRO 6000 Blackwell for inference, H100 for fine-tuning, Jetson Orin at the edge. SamurAI V2 is built on the Metropolis VSS blueprint with Cosmos 3 fine-tuned on Malaysian roads; it runs on 1,000+ of 16,000+ managed cameras, with 10,000 planned in 2027. ITMAX claims alert verification accuracy improved by more than 70% and incident search time fell more than 90% in the first KL deployment. On 10 corridors with adaptive signals, peak travel time is 26% shorter on average; TeraSim simulation of a three-junction cluster cut delay 29.8 s per vehicle at about 700 vehicles/hour. City Maestro (Nemotron) agent runs on-demand simulations.
Limitations & uncertainty
Self-reported metrics; '26% shorter' lacks a stated baseline; simulation results are not field results; non-U.S.
Stream editions
Each edition carries its own synthesis and evidence ledger.
November 10, 20261 edition
November 9, 20261 edition
November 8, 20261 edition
November 7, 20261 edition
November 6, 20261 edition
November 5, 20261 edition
November 4, 20261 edition
November 3, 20261 edition