Trajectory announced a $45 million Series A at a $300 million valuation, led by Sequoia with participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures, TechCrunch reported. The company’s platform is now publicly available and uses application feedback to retrain open-source models for specialized tasks, a capability Trajectory says can let nontechnical specialists improve models largely automatically.
Microsoft released Microsoft Corporation System Hardware Update 200.0.7.0 for the RTX Spark-powered Surface Laptop Ultra six days after the device went on sale. Microsoft says the update addresses extended reboot times, screen flickering and unexpected restarts, though reviewers differed on whether they encountered flickering. Windows Central's Zac Bowden reported that his reboot problem was fixed. The update is rolling out gradually, and installed firmware cannot be uninstalled or reverted.
NVIDIA released NVIDIA Recovery, an open-source workload-level service within NVIDIA DSX OS that automates recovery across distributed AI training jobs. It detects interruptions through application logs, GPU metrics, and scheduler signals, then attributes failures, isolates affected resources, and coordinates checkpoint restarts on healthy capacity. The service supports Slurm and can be deployed on an operator’s own cluster, while Kubernetes support remains on NVIDIA’s roadmap.
PC Games Hardware benchmarked a reader’s 22GB-modified GeForce RTX 2080 Ti against the stock 11GB card, focusing on nine memory-intensive games and demanding settings. The cards performed similarly at 1440p with DLSS Quality, while the stock card was about 20% slower and stuttered in Flight Simulator 2024 and Spider-Man 2, which exceeded 11GB of VRAM usage. In an AI image test, the 22GB card completed the toughest scene more than 200 seconds faster than a factory-overclocked 11GB card, but an RTX 5070 Ti finished the job in about 85 seconds.
NVIDIA says the AI Cluster Runtime (AICR), part of the NVIDIA DSX AI Factory Platform, now captures tuned Kubernetes configurations as reusable recipes, renders ordered Helm chart bundles, and validates running clusters so platform teams can be ready for NVIDIA Vera Rubin NVL72 compute trays on day one. The open-source NVIDIA/aicr repository publishes v1.0.0 preview recipes for the rke2 service and the vr200 accelerator, including training recipes validated across deployment, conformance, and an NCCL performance phase, plus an inference variant for Dynamo. The blog also notes Vera Rubin requires a new driver branch (R615.23 tested, R610 targeted for early support, R620 for official platform support) and swaps ConnectX-7/BlueField-3 for ConnectX-9 SuperNICs with BlueField-4 DPUs. NVIDIA also reports that validation surfaced an unmapped Rubin architecture value in go-nvlib, fixed in go-nvlib 0.12.0 and shipped in DRA driver v0.5.0.
German Chancellor Friedrich Merz said Germany and Europe must accelerate development of their own AI technology to avoid dependence on external providers, according to The Next Web citing Reuters. Speaking at an AI conference in Berlin, he called digital sovereignty an “existential question” and said Germany should control technology across the full supply chain, from hardware to software, while acknowledging the effort required. The Next Web also reported that Germany has created a digital affairs ministry, aims to quadruple its AI computing capacity by 2030 and has attracted pledges exceeding €225 billion from companies including Amazon and Nvidia.
NVIDIA introduced Cosmos as a platform of world foundation models and development tools for physical AI, with Cosmos Framework covering training, multimodal data processing, optimization, and inference. The accompanying workflow covers synchronized camera and LiDAR simulation, GPU-accelerated deployment through NIM microservices, and support for models including Alpamayo, GR00T, Qwen, and pi0.5; these capabilities are described in NVIDIA Technical Blog.
In a technical blog post, NVIDIA introduced Workload Map, an open-source component licensed under Apache 2.0 that provides a reusable declarative description of Kubernetes workload structures across frameworks. A JobSet definition can support workload inspection, gang scheduling, status interpretation, submission, mutation and platform visibility. NVIDIA says KAI Scheduler uses Workload Map as a fallback for workload types without native plugins, NVIDIA Run:ai uses it in production, and Elotl Nova uses it to interpret workload structures and resource requirements during multi-cluster scheduling.
The New York Times, citing sources, reports that Anthropic’s compute agreements are expected to provide more than 3.5 gigawatts of capacity in 2026. The company is diversifying beyond Nvidia AI chips toward suppliers including Amazon, Google and AMD.
Researchers introduced Realtime-Bench, a 50-task benchmark for evaluating general-purpose AI systems whose decisions and actions must respond to changing conditions in real time. It includes 30 web tasks and 20 game tasks, measuring temporal state awareness, adaptive deliberation, and time-aware action selection under wall-clock, paused-clock, and token-clock conditions. Import AI reported that GPT-6-Astra achieved the highest wall-clock pass rate at 36%, while high-reasoning models approached 70% to 80% with the clock paused.
Nvidia told affected customers that unidentified activity on its Brev cloud development platform may have exposed stored information between October 4 and October 20. The potentially affected material included an SSH key used to access customer virtual machines, and Nvidia said it rotated the key and updated customer instances with a replacement. The company said it found no evidence of improper use but did not explain how access occurred or confirm whether customer data or workloads were accessed.
36氪’s analysis links three recent physical AI developments: Google DeepMind’s Genie 4, NVIDIA’s expanded Isaac robotics platform, and an Intel-SAP autonomous-warehouse proof of concept. It describes Genie 4 as an interactive world model with native audio and real-time natural-language scene changes, while NVIDIA added a humanoid reference platform, datasets, runtimes, and updates to its robot-learning framework. The Intel-SAP demonstration connected SAP EWM workflows with edge AI and autonomous robots that return task status and results.
ChillyWillMD demonstrated Nintendo Switch 2 game Mario Kart World running on a PC through an emulator with NVIDIA DLSS 5 Neural Rendering. IT之家 and Wccftech report that the demo reached 60 FPS on an NVIDIA GeForce RTX 5060 Ti with an Intel Core i9-13900K, while both outlets said the improved lighting gave the game an animated-film appearance. The demonstration still showed jagged character models and noisy headlight illumination.
IDC figures cited by TMTPost and republished by 36Kr show China shipped about 611,000 smart glasses in Q2 2026, down 8.0% year over year, while the global market grew 35.3%. 光子星球’s analysis says vendors are narrowing their focus to display-led products and vertical applications such as driving, translation, meetings, accessibility and first-person data collection, while also exploring personal agents. The report cautions that one quarter may be distorted by new-product launch timing.
This tutorial explains how to evaluate LLM inference services by balancing throughput, request latency, and memory utilization, with latency further divided into TTFT and TPOT. It recommends profiling bottlenecks before tuning vLLM parameters, and maps options such as batched-token limits, concurrency, prefix caching, chunked prefill, and speculative decoding to different workload goals. The author also warns that some listed vLLM commands, environment variables, and metric names are invalid, and that NVIDIA-specific tooling must be replaced when running on Ascend NPUs.
Anthropic’s amended S-1 reported third-quarter revenue of about $17 billion, up from $11.6 billion in the second quarter, but the company returned to a $567 million operating loss from $974 million in operating profit. Compute and infrastructure costs rose to $9.9 billion, or 58% of revenue. The filing disclosed $895.4 billion in purchase commitments and lease agreements, substantially all related to compute and infrastructure, although the supplied reports do not provide a complete breakdown of their terms. Anthropic said the quarterly result was affected by $5 billion in short-term compute purchases, which it said cost more than comparable capacity under multiyear commitments, and an approximately $2 billion noncash employee charitable-matching expense.
Anthropic’s 335-page S-1 prospectus, submitted on October 25, 2026, reportedly disclosed about $637 billion in compute commitments to seven named suppliers over the next decade, according to TMTPost. TechNews reported a broader total of about $895.4 billion for computing and infrastructure commitments, including about $447.5 billion due by the end of 2030. The filings include take-or-pay obligations and a reported $161.2 billion Broadcom lease covering primarily Google-designed TPU chips.
South Korea’s Ministry of Science and ICT announced that the sixth-generation national supercomputer “Hanjin” is now operational. The hybrid CPU-GPU system uses NVIDIA GH200 chips, with 9,936 CPUs and 8,496 GPUs, and the ministry reported 614 PF of theoretical performance. It also claimed Hanjin offers 24 times the computing performance, 10 times the storage capacity and four times the network performance of the 2018 Nurion system, but these comparisons and a projected TOP500 top-10 ranking have not been independently verified.
Nvidia showcased two research-stage optical-interconnect technologies at SEMICON West 2026, according to a report republished by 创业邦. Optics-on-Interposer places the optical engine on the silicon interposer to shorten electrical paths, while an eight-wavelength DWDM demonstration transmitted 256 Gb/s over one fiber using eight 32 Gb/s data channels. Nvidia has not disclosed product adoption, commercialization schedules, measured power improvements, or bandwidth-density data.
A 36Kr report says Alphabet shares fell 12.64%, their largest one-day decline since listing, as investors focused on heavy AI capital expenditure, negative free cash flow and external financing. The report links Alphabet's share decline to these financial pressures while noting that causes differed across companies, including Intel's product availability and guidance. It also compares Meta, CoreWeave, a copper-clad laminate manufacturer, Nvidia and selected suppliers, arguing that AI demand remains strong while working-capital needs, debt costs and financing structures increasingly constrain expansion. The report gives conflicting figures for Alphabet's quarterly free cash flow, so its exact amount is unclear.
科技新报 cited Yahoo Finance statistics showing that 47 AI and chip stocks lost about $5 trillion in market value from October 6, while the semiconductor ETF fell roughly 17% during the month. Nvidia and Alphabet each shed about $835 billion in value, as selling pressure remained concentrated on AI-related stocks while 625 other large U.S. companies gained about $1.3 trillion in combined value.
US media reported that Megaspeed executive Huang Le, also known as Alice Huang, is suspected of diverting NVIDIA AI chips to China and passing information to Beijing, with surveillance showing close ties to NVIDIA CEO Jensen Huang. President Donald Trump said he would investigate the allegations carefully and indicated Jensen Huang could remain a technology adviser unless problems emerge. NVIDIA said its relationship with Huang was proper and that it had checked that Megaspeed did not illegally ship NVIDIA AI processors to China.
The supplied article surveys large-model pretraining, covering data, tokenization, optimizers, memory, distributed parallelism, checkpoints, and model evaluation. It compares Megatron-LM, DeepSpeed, and torchtitan; explains the different configuration guidance in Kaplan’s and Chinchilla’s scaling laws; and discusses cost estimation, loss spikes, checkpoint recovery, and resuming distributed training. Framework and performance characterizations are the article’s assessments, while its cost calculation is presented without accompanying source calculations.
Auras Precision Technology reportedly entered Nvidia’s supply chain for metal heat spreaders after passing certification in 2024. The Taiwanese manufacturer developed the co-packaged thermal components from its metal-stamping expertise in LED lead frames, and TechNews, citing Commercial Weekly, described it as one of only two firms able to make and supply the components to Nvidia. The report also says Google and Amazon use Auras thermal products in internally developed AI chips.
At SEMICON West 2026 in San Francisco, NVIDIA researcher Nikola Nedovic introduced Optics-on-Interposer (OOI), a research approach that moves optical engines from the package substrate onto the silicon interposer shared with compute chips. The design shortens high-speed electrical paths with denser interconnects; NVIDIA says it can reduce power use and increase interconnect bandwidth density, but has not disclosed measured results or a commercialization timeline. NVIDIA also presented dense wavelength division multiplexing research using eight wavelengths over one fiber to achieve 256Gb/s of total bandwidth.
Wccftech, citing Commercial Times and unnamed industry sources, reports that NVIDIA is considering PTFE to reduce signal loss in next-generation NVLink switch boards for AI servers. Possible applications include a Rubin Ultra switch board or a next-generation orthogonal midplane. The report says PTFE remains at the sample stage, with industry estimates putting overall yields below 5%; some suppliers are skeptical. Wccftech assigns the rumor a 45% plausibility rating.
An analysis syndicated by 36Kr and 创业邦 argues that AI expansion depends on more than electricity, linking it to power generation, grids, transformers, cooling, memory, packaging, capital and local permitting. It cites the International Energy Agency’s April 2026 report for rising infrastructure spending and data-center electricity demand, while noting that fast-evolving chips and data centers face slower grid and equipment timelines. The analysis also describes constraints shifting from GPUs to HBM and advanced packaging, then to transformers, gas turbines and grid connections.
《科创板日报》 reports that Chinese private-capital firms remain active in embodied intelligence while becoming less willing to discuss physical AI publicly and more cautious about current investment conditions and deal pace. Its incomplete tally found at least 185 financing deals worth about RMB33.335 billion in China in Q3 2026, below the more than RMB46 billion average quarterly amount recorded in the first half. Investors are reportedly asking more about real-world deployment and business progress as financing slows, while some are concerned about potential entry by large AI companies.
A Reddit user said they bought an EVGA RTX 3090 FTW3 with 24GB of memory for $1,100 to expand local AI workloads beyond the 12GB capacity of an RTX 4070. They plan to install it in a headless Ubuntu Server 26.04 LTS system that serves Ling-3.0-tiny, Omnivoice, Parakeet, coding inference, and other home-lab services over a LAN. The post asks for advice on long-context coding models, on-demand model swapping, and managing the FTW3's Linux power and thermal settings without EVGA Precision X1. The seller's claims about low use and no mining or AI workloads were not independently verified.
Agility Robotics plans to list on Nasdaq on November 5, 2026 under ticker AGLT, following a shareholder vote and transaction closing. TMTPost’s analysis says the company reported $1.782 million in 2025 revenue against a $2.5 billion pre-money valuation, while raising only about $465 million after substantial SPAC redemptions. It also examines Digit 5’s RaaS economics, deployment utilization, funding needs, and management’s going-concern warning.
What problem it solves
AI agents need to choose among models while balancing task success, cost, and latency for their workloads. Applications also need flexibility in where routing runs and who controls model clients, credentials, retries, and service operation.
Key features
Multiple routing algorithms — Routing can be configured with Auto, Task, Execution, or Composite strategies, including mode…
What problem it solves
Serving large language models requires efficient memory management, flexible model execution, and practical integration with existing models and APIs. vLLM addresses these needs with inference and serving capabilities spanning batching, caching, quantization, distributed execution, and multiple hardware platforms.
Key features
Efficient memory and serving — vLLM manages atte…
What problem it solves
AI agents need product-specific instructions to use NVIDIA software reliably across physical AI, robotics, simulation, CUDA-X libraries, RAG, and platform workflows. The catalog centralizes these skills while keeping them maintained in their respective product repositories and synchronized through an automated pipeline.
Key features
Agent-discoverable skill catalog — The NVI…
What problem it solves
AI agent skills need to be checked for quality, security, duplication, and their actual effect on agent behavior. Static checks alone cannot establish whether guidance improves task results, so evaluation must also support live comparison with and without a skill.
Key features
Three-tier evaluation pipeline — Independent tiers validate skill safety and structure, detect over…
What problem it solves
Running large language and vision-language models can require substantial setup and may depend on the available hardware. The project aims to make inference accessible across a wide range of local and cloud hardware with minimal setup.
Key features
Minimal-setup inference — The project provides installation options for common platforms, including scripts, Docker, pre-built b…
What problem it solves
Serving language models requires reading checkpoints, tokenizing requests, and running accelerated kernels while managing a local model server. TensorFold addresses these requirements with a native binary that can run on Apple Silicon and NVIDIA GPUs without requiring Python or MLX for serving.
Key features
Native Metal and CUDA inference — The Zig-based engine executes mode…
What problem it solves
Interactive multimodal and speech systems often require generation to span heterogeneous stages with different compute patterns, dependency structures, and resource needs. Serving these pipelines requires coordinated stage lifecycles, inter-stage transport, model integrations, and a unified API surface.
Key features
Coordinated multi-stage runtime — Models generation as coor…
What problem it solves
Data analysis workflows often require expressive query composition, flexible execution models, and interoperability with columnar data. Polars addresses these needs by allowing the DataFrame API and SQL to share the same optimizer and engine, with support for large workloads and custom extensions.
Key features
Unified DataFrame and SQL Queries — Queries can be written with e…
What problem it solves
Running frontier-scale open-weight models locally requires an inference platform that can make efficient use of the heterogeneous hardware available on a personal device. The project addresses this by treating GPUs, CPUs, host memory, and interconnects as a unified, elastic inference platform.
Key features
Edge-native MoE runtime — The runtime provides MoE serving with bandw…
What problem it solves
Autonomous AI agents often need to read files, install packages, call APIs, and use credentials, but unrestricted access can expose host data, secrets, and networks. OpenShell provides these capabilities within an isolated runtime where declared policies constrain agent access.
Key features
Kernel-level runtime enforcement — Agents run in isolated sandboxes where kernel cont…
What problem it solves
AI agent skills execute with implicit trust and often receive minimal vetting before installation. SkillSpector helps users assess whether a skill contains vulnerabilities, malicious patterns, or other security risks before installing it.
Key features
Multi-format scanning — Scan agent skills supplied as Git repositories, URLs, zip files, directories, or individual files.
So…
What problem it solves
Mosaics obscure portions of videos and images, making it difficult to view the affected content without changing the rest of the picture. Users need a restoration workflow that can identify selected mosaic regions and process them with GPU-accelerated video handling.
Key features
Targeted mosaic restoration — Detects mosaics in videos and restores them while leaving the rest…
What problem it solves
Deep learning models often need to be adapted for efficient deployment across different inference frameworks. Model Optimizer addresses this by providing a unified set of optimization techniques and export workflows for quantized checkpoints.
Key features
Post-training quantization — Compress models with quantization techniques available for Hugging Face, Megatron-Bridge, Di…
What problem it solves
Platform teams need to create and operate isolated Kubernetes clusters for customers on infrastructure they supply and control. They need a management plane that can provision GPU machines, manage cluster lifecycle operations, and establish node identity before issuing join credentials.
Key features
Independent workload clusters — Create workload clusters through the NVKE AP…
What problem it solves
Running AI models locally can require integrating multiple inference engines, model formats, hardware accelerators, and API surfaces. LocalAI addresses this with a small core, on-demand backends, and compatible APIs that work across supported hardware without requiring a GPU.
Key features
Composable inference backends — Backends are packaged separately and pulled only when a…
What problem it solves
LLM and Visual Gen inference requires specialized kernels, efficient runtime orchestration, and flexible deployment across NVIDIA GPU configurations. Developers also need a modular framework that supports customization while integrating with established inference tooling.
Key features
Optimized inference kernels — Custom kernels accelerate common inference operations, includ…
What problem it solves
Rust developers need host-side access to the CUDA runtime and tools for expressing GPU kernels while preserving Rust's ownership and borrowing rules. The platform provides separate programming models for workloads that require direct thread-level control and workloads naturally expressed through tensor operations.
Key features
Shared CUDA runtime — Host-side crates provide d…
What problem it solves
Agents can retrieve documents and query databases, but unreliable results undermine agentic analytics. Teams need a context layer that records where data lives, what its definitions mean, whether it is usable, how it should be queried, its provenance, and related concepts.
Key features
Enterprise ontology generation — Generate semantic ontologies from database metadata, infe…