At an ITRI seminar on Taiwan’s communications industry, speakers outlined AI-RAN integration, silicon photonics, space computing and 6G as major technology directions for 2027. ITRI estimated Taiwan’s communications-industry output could reach NT$1.4471 trillion in 2027, up 3.0% annually, while Yole Intelligence projected the global silicon-photonics market could grow from US$1.82 billion in 2023 to US$25.2 billion in 2030. The report also relayed ITRI’s account of a SoftBank and Ericsson commercial-network trial in Japan, claiming AI-native scheduling raised spectral efficiency by up to 25% and downlink user throughput by up to 50%.
According to IT之家 AI, citing Business Insider, Anthropic's S-1 filing describes a unified compute pool spanning Amazon Trainium, Google TPU, Nvidia GPUs and AMD GPUs, with software assigning workloads to the most suitable chips. Anthropic has spent years and substantial capital developing this cross-platform scheduling capability and is also investing billions in Nvidia compute, while analysts cited by the report say migration difficulties limit the near-term threat to Nvidia.
An analysis from the WeChat account Zhengjing She reports that Nvidia generated $96.2 billion in revenue in its fiscal 2027 second quarter, including $89 billion from data centers, at a 75% gross margin. It contrasts these figures with negative quarterly free cash flow at Meta, Alphabet and Alibaba, while noting that Microsoft reported positive free cash flow in its fiscal 2026 fourth quarter as Azure revenue grew 43%. The analysis says GPUs, data centers and power contracts require substantial spending before revenue and profits are realized, so rising valuations do not necessarily mean companies are generating cash. The figures cited in the analysis were not independently verified for this review.
AI data-center demand is reportedly diverting PCB and chip-substrate production capacity from PC and consumer-electronics makers. Nikkei Asia reports that some PCB lead times have stretched to 30 months, copper-clad laminate costs have risen at least 12-fold since July 2026, and one major supplier's order book extends into 2030.
SK hynix reported record third-quarter operating profit of 83.3 trillion won and revenue of 104.6 trillion won, attributing stronger memory-chip prices to AI infrastructure demand. Its shareholder-return plan includes 30 trillion won in cash dividends and 60 trillion won in share buybacks and cancellations. The company also said DRAM and high-bandwidth memory capacity is booked through 2029 and forecast chip shortages could continue until 2030.
Anthropic’s public S-1 says the company generated $16.286 billion in first-half 2026 revenue and reported $974 million in operating profit during the second quarter, its first disclosed quarterly operating profit. The filing shows compute and infrastructure costs falling from 201% of revenue in the first half of 2025 to 55% in 2026, while also revealing hundreds of billions of dollars in capacity commitments and major risks involving customer concentration, cloud partners, autonomous systems and AI-related harm. TechCrunch, Understanding AI, Tom’s Hardware and 36Kr also examine strategic-investor stakes, research-heavy compute spending and safety and legal disclosures in the filing.
Zankore, an Indonesian AI infrastructure platform backed by Nvidia, is syndicating a $3.1 billion GPU loan to the broader market, according to people familiar with the matter. The transaction adds to a flurry of financing for Asia’s AI buildout.
TechNews reports that unconfirmed market rumors could push NVIDIA’s RTX 60 series to 2028 and likewise delay AMD’s RDNA 5, potentially extending consumer GPU generations from roughly two years to about three. The analysis links this possibility to strong AI accelerator demand, NVIDIA’s data-center revenue and rising memory costs, including a rumored pause for RTX 50 SUPER. It also points to narrower performance gains from TechSpot testing and improving integrated GPUs as factors weakening the case for new consumer graphics cards.
NVIDIA Technical Blog says CUDA 13.4 adds transparent, application-aware memory enforcement for GPU time-sharing while retaining the SM-sharing compute mode available through Multi-Process Service. Workloads can request an exact memory allocation or a guaranteed minimum with optional burst capacity, and the new open-source GPU Fractioning Operator applies those limits and compute modes in Kubernetes through pod annotations. NVIDIA also says KAI Scheduler v0.18.0 can select GPU fractions and write the required annotations, though no independent test results were supplied.
Dell Technologies says enterprises securing agentic AI should combine supply-chain vetting, policy enforcement, runtime sandboxing, telemetry, narrowly scoped permissions and observability across hardware, runtime, data and applications. Dell executives also said some customers’ planned chatbots could have been exploited during penetration testing, while the company is working with Nvidia to move agentic AI from deskside systems into data centers.
Oracle confirmed that it will offer Google TPUs through Oracle Cloud Infrastructure, making it the first cloud vendor outside Google to sell Google’s AI accelerators, according to TechRadar. TPU 8t for training and TPU 8i for inference are planned for OCI availability in 2027. Separately, The Information reported that OpenAI plans to purchase Google TPUs, but final terms are unknown and both OpenAI and Google declined to comment.
Oracle will offer cloud services based on Cerebras Systems computers in its data centers, according to a report. SiliconANGLE, citing CNBC, also reported that Oracle intends to deploy Cerebras' wafer-scale AI accelerators in its cloud.
NVIDIA released the open Nemotron-Personas-France and Nemotron-Personas-Belgium locale-grounded persona datasets, created with Pleias and KU Leuven. The datasets belong to an open Nemotron data program covering 12 countries and can support post-training of open models for application-specific local knowledge, cultural nuance and regional languages. NVIDIA's accompanying tutorial covers generating synthetic conversations, fine-tuning with LoRA, merging the adapter and evaluating the resulting model.
NVIDIA and Oracle announced that Vera Rubin NVL72 systems and standalone Vera CPUs are generally available on Oracle Cloud Infrastructure Superclusters. They also plan to expand the Rubin deployment to multi-gigawatt scale and offer an HGX Rubin NVL8 configuration with eight GPUs, a dual Intel Xeon 6 host and 4 TB of DDR5 memory. The vendors say CPU and GPU resources can scale independently for model training, inference and agent workloads; Tom’s Hardware separately benchmarked Vera on a non-OCI server, so those results do not independently verify OCI production performance.
The 18-institution C-BRAIN consortium is working with NVIDIA on AI systems for Alzheimer’s and other dementia research, using a hybrid centralized and federated approach to let models learn from institutional data without moving it beyond institutional walls. NVIDIA Blog reports that C-BRAIN runs 64 NVIDIA data center GPUs on premises alongside Microsoft Azure cloud nodes, incorporates BioNeMo and MONAI, and has built the open-source agentic research tools OpenScientist and InsightEngine. The consortium also aims to develop a multimodal biological foundation model using neuroimaging, multi-omics, fluid biomarkers and longitudinal clinical data.
An NVIDIA Technical Blog post details three optimizations for Ling-3.0-flash on NVIDIA DGX Spark: the Humming MXFP4 MoE backend, online FP8 quantization of the LM head, and NEXTN multi-token-prediction speculative decoding. Profiling showed batch-size-1 decoding was memory-bound, with weight reads dominating, and the Mixture-of-Experts computation plus the output-layer projection standing out as the costliest operations. On a synthetic workload the baseline ran at 27.71 ms TPOT and 36.09 tok/s, while NEXTN MTP reached 9.72 ms and 102.88 tok/s; across the full 1,319-question GSM8K evaluation the three optimizations together raised decode throughput 42.5% and cut mean end-to-end latency 24.5%, with no statistically detectable accuracy change (96.89% to 96.82%). The post also warns that combining online FP8 LM-head quantization with MTP requires the draft model to share the entire LM-head module, otherwise the framework silently falls back to a configuration slower than baseline.
Nvidia denied reports that it has discontinued the RTX 5090, amid renewed speculation about constrained supply for both desktop and laptop variants. XMG said deteriorating procurement conditions, sharply higher GPU prices and reduced replenishment orders have constrained availability, prompting the company to raise the price of the Schenker Key 18 Pro by €250 and warning that other RTX 5090 laptops could see similar or larger increases. TechRadar also cited an unnamed board maker that reportedly could not obtain the GB203 chip and 32GB GDDR7 memory package, although that account remains unverified.
AMD announced a $1 billion investment in Meta’s equity offering alongside an expanded infrastructure partnership covering EPYC CPUs, Instinct GPUs, Helios rack-scale systems, Pensando networking and ROCm software. AMD and Meta said the first Helios systems have been installed at Meta and are undergoing testing, with broader deployment planned in the near future. The companies also set a goal of building gigawatt-scale AI infrastructure, while AMD said Meta has deployed millions of EPYC processors and uses Instinct GPUs for training, ranking and recommendation workloads.
Nvidia joined Accel, Shastra VC and Micromax in Simplismart’s all-primary Series B, led by Dallas Venture Capital, which values the Bengaluru startup at about $100 million including the round. Simplismart says its software automates model deployment and optimizes latency, request handling and computing consumption, and that it helped InVideo cut GPU serving costs by 56%. Nvidia and Simplismart are also collaborating on integrations involving Nvidia Dynamo, while funding will support US expansion, hiring and inference software development.
A 36Kr report, citing Menlo Ventures investor Deedy Das, says a group of Silicon Valley AI startups reached more than $1 million in monthly revenue within months. Their model is to rent NVIDIA GPUs, host models including DeepSeek, Kimi and GLM, and sell tokens to customers. Das estimated that an NVIDIA B300 rented for $8 per hour could generate $10–$20 worth of tokens per hour at sufficient efficiency and demand, though the report says rising revenue was accompanied by thin margins amid an intensifying price war.
A TMTPost analysis cites a Chinese internet report that Apple sought a lower LPDDR5X price from ChangXin Storage but was refused. The article presents the claim as the backdrop for a broader shift in Apple’s supply-chain position as AI infrastructure absorbs memory capacity. Citing TrendForce, it says LPDDR5X contract prices rose 78% to 83% in the second quarter as manufacturers shifted production toward HBM and other server products. It also cites Ming-Chi Kuo’s forecast that tight LPDDR supply could reduce Apple’s A20 chip supply by 10% to 20%, and TrendForce’s estimate that storage could account for about 34% of iPhone 18 Pro material costs.
Goldman Sachs research, as reported by IT之家 citing Business Insider, found that semiconductor moves explained about 54% of cross-asset performance dispersion over three weeks, compared with 2% for oil prices. Goldman strategists also said investor sentiment has become closely linked to technology stocks, although it has not yet reached levels that historically indicate buying opportunities.
36Kr reports that Nomura’s global robotics study finds the industry is in an accelerated phase after the AlexNet moment but before the GPT-2 stage. Although new robot models are approaching GPT-2-like capability, none has yet demonstrated the reliability needed for practical work, mass production or factory deployment. Nomura says a true ChatGPT moment will be confirmed through operating metrics, a closed-loop data flywheel, consistent hardware and real-world validation rather than a single model launch.
Advantech opened its new 10-acre North America headquarters in Tustin, California, on October 20 and followed the ceremony with its Advantech Technology Summit. The site combines regional operations, engineering, systems integration, customer service and application displays, including a new center for custom assembly, hardware-software integration, testing and delivery. The summit with AMD, Intel, NVIDIA and Qualcomm focused on Edge AI, edge computing and applications in healthcare, robotics and industrial automation.
GF Securities analyst Jeff Pu expects SpaceX to receive 20,000 NVIDIA Vera Rubin NVL72 racks in 2027, making it the largest single buyer in a projected 96,000-rack market. Pu says the six major neoclouds are expected to buy 34,000 racks, or 35% of projected demand, and estimates Foxconn could handle up to 45% of SpaceX's data-center orders.
NVIDIA released SkillSpector under the Apache License 2.0 as a security scanner for AI agent skills. It uses static analysis and optional LLM semantic evaluation to identify risks such as prompt injection, data exfiltration, malicious patterns, dependency vulnerabilities and supply-chain issues before installation in tools including Claude Code, Codex CLI and Gemini CLI.
The documentation says SkillSpector does not sandbox the host or isolate an installed skill. It may miss attacks in non-English content, images, encrypted or binary code and runtime behavior. It also notes that LLM analysis sends file contents to the configured provider, while dependency checks send dependency names and versions to OSV.dev.
NVIDIA cites a 2026 study in which 26.1% of an analyzed subset of 31,132 skills contained at least one vulnerability and 5.2% showed likely malicious intent; these figures were not independently verified in the supplied report.
NVIDIA and Meta announced an expanded partnership centered on Muse and the Muse model family, with workflow optimizations intended to improve training and inference performance and total cost of ownership. NVIDIA will invest $3 billion in Meta’s equity offering, while Meta plans to expand deployments of Vera Rubin, Vera CPU, Spectrum-X networking and BlueField-4, and the companies will consider OpenShell and Sentry collaboration for agent safety and monitoring. Meta claims Muse has delivered frontier-level benchmark performance and rising consumer downloads, but the reports provide no independent figures.
Kazakhstan joined a U.S.-led AI supply-chain strengthening mechanism in Washington in June and, three weeks later, became a founding member of China’s global AI initiative in Shanghai, TechNews reported. The report says Kazakhstan is pursuing advanced technology, high-tech jobs and supply-chain opportunities through the U.S. system while seeking influence over regional and global AI standards through China’s platform. It also cites critical minerals, inexpensive electricity, a potential Nvidia data-center agreement worth up to $10 billion, and record Chinese direct investment in Kazakhstan.
XTX, the quantitative trading company founded by Alex Gerko, uses machine learning and neural networks to predict prices across more than 53,000 financial instruments, according to 36Kr’s republication of 投中网. The report says its research relies on a supercomputer with 25,000 AI chips, while the firm announced plans in 2025 to invest more than €1 billion in five Finnish data centers. It also cites 2025 UK entity filings showing combined revenue of about £3.93 billion for XTX’s three main UK operating entities.
TechNews reports that NVIDIA appointed former Google Cloud executive Alison Wagonfeld as its first chief marketing officer and Microsoft veteran Nick Parker as sales chief. Several longtime executives also retired or handed over responsibilities, with Shanker Trivedi’s portfolio divided among younger managers as NVIDIA adopts a flatter structure. Finance, legal, operations and engineering remained comparatively stable.
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…