Former Intel CEO Pat Gelsinger said at the COSM Technology Summit 2026 that AI computing efficiency needs roughly a 10,000-fold improvement, arguing that incremental chip improvements alone are insufficient. He said progress requires simultaneous advances in optics, power, cooling, materials, and packaging, while identifying power availability as a major constraint on hyperscale data centers. cnBeta also reported that he rejects the idea that semiconductor scaling has reached its physical limit.
US technology stocks suffered a broad sell-off on November 20, with the Nasdaq Composite down 5.28%, the S&P 500 down 3.70%, and Nvidia falling about 14%, according to 36氪 and other reports. Anthropic closed down 8.5% at $88.91 on its second trading day, giving back most of its first-day gain despite a reported $119.5 billion raised through the IPO and concurrent private placement. Reports tied the pressure to questions about AI investment returns, Nvidia’s accounting review and higher interest rates, while noting that strong demand and financial results prevent the move from establishing an industrywide collapse.
Microsoft announced a developer preview of USB4-based RDMA that lets two or more compatible Windows 11 PCs pool memory and GPU capacity to run a shared AI model through supported inference frameworks, without dedicated RDMA adapters or switches. Microsoft reported round-trip latency of about 28–50 microseconds for small transfers. In separate vendor tests, Microsoft used three AMD mini PCs to run a four-bit-quantized 235-billion-parameter Qwen3 model at about 18.5 tokens per second; AMD reported about 22 tokens per second for Qwen3-235B-A22B, while Intel reported 54.2 tokens per second for GPT-OSS-120B on four Core Ultra Series 3 PCs.
According to IT之家, Maxtang launched the IXH-WCL35 fanless embedded industrial computer, based on Intel's third-generation Core processor Wildcat Lake. The device measures 205×125×70.5 mm and includes one DDR5 SO-DIMM slot, one M.2 M-Key 2280 slot with PCIe Gen4 ×2 support, an FPC SATA III interface, four USB-A 3.2 ports, four USB-A 2.0 ports, six COM ports, two HDMI 2.0 ports, two 1GbE RJ45 ports and an 8-bit GPIO interface. It supports expansion with wireless cards and LTE cellular modules. Maxtang says the computer is intended for industrial automation and edge-computing deployments, including smart manufacturing, smart transportation, energy management and edge-AI applications.
Intel announced on November 18 that Microsoft will use its Foundry EMIB advanced packaging technology for Azure's in-house chips, making Microsoft the first publicly named external customer of the technology, though specific products and production scale were not disclosed. Supply-chain reports cited by TechNews say Google did evaluate EMIB-T for next-generation TPU advanced packaging as a second packaging source, but testing yield fell short of expectations and Google has reportedly resumed seeking TSMC support, with any order shift still uncertain. Analysts note the Google arrangement concerns back-end packaging only, with front-end fabrication remaining at TSMC, and that TSMC's CoWoS-L keeps an edge in production maturity and yield.
Intel says its forthcoming Diamond Rapids server processors will offer configurations with up to 256 performance cores, 1.6 TB/s memory bandwidth and 1.28 GB of last-level cache. The company also says GPU D-series accelerators will support up to 480GB of LPDDR5X memory and FP4, MXFP4 and FP64 formats, alongside continued work on silicon photonics, quantum computing and software compatibility. These specifications and benefits remain vendor claims without independent validation.
TechRadar found that the Geekom IT13 2026 Edition handled Microsoft Office, Adobe Creative Suite and 4K video editing well, though more demanding grading and graphics tasks exposed its limits. The compact mini PC offers a metal chassis, multiple ports, upgradeable memory and storage, and an integrated NPU for compatible local AI workloads. It was available from Geekom and Amazon, with a listed starting price of £1,099.
Castrol launched Castrol CORE in China at the Data Center Standards Conference, combining design support, cooling technologies, monitoring hardware and lifecycle services for liquid-cooled AI data centers. The company also issued a white paper on operating and maintaining cold-plate liquid-cooling systems, while reporting initial orders and shipments for a monitoring sensor and simulated heating device. Castrol says its cooling technologies have been tested with Intel and Nvidia, but the supplied report provides no independent test or shipment data.
llm-d 1.0 adds multi-tenant resource management, request-priority scheduling, high availability, rolling upgrades and context-aware routing for agent workflows. An IBM Research benchmark deployed an approximately 753-billion-parameter model through llm-d, supporting up to 3,000 concurrent coding agents at peak output throughput above 6.6 million tokens per minute. Red Hat AI 3.5 also added governance capabilities, while IBM and Red Hat jointly disclosed that Lightwell Clearinghouse had reached general availability.
Morgan Stanley reiterated an Overweight rating on TSMC and set a target price of NT$3,088, representing 42% upside from the reported NT$2,170 closing price. TechNews also reported that TSMC remains confident in AI demand, targets more than doubling 2nm capacity in 2027, and is assessing another US fab site while discussing a potential TeraFab collaboration with Tesla.
The Verge AI reports that Nvidia generated about $8,300 in profit per second, with quarterly profit reaching $65.3 billion as data-center and AI demand drove a sharp rise beyond earlier crypto and GPU booms. Following Jensen Huang’s resignation, interim CEO and CFO Colette Kress described operations as “business as usual,” said demand was accelerating and supply was growing, and projected $130 billion in quarterly revenue. Nvidia also plans to return all excess free cash flow net of strategic uses to shareholders, while its gross margin is expected to ease to 71–72 percent.
Microsoft and Intel announced a long-term agreement under which Azure will secure multi-year Xeon supply capacity, including the next-generation Diamond Rapids processor. Microsoft also became the first publicly announced adopter of Intel’s EMIB advanced-packaging technology for its own Azure silicon, with the broader collaboration covering CXL memory, QAT acceleration and TDX security.
cnBeta, citing newly exposed roadmap information, reports that Intel plans to divide Razor Lake into S, U and L series spanning desktops, mainstream mobile devices and ultra-low-power systems. Razor Lake-S is reportedly aimed at traditional desktops, Razor Lake-U at thin and light laptops, and Razor Lake-L at ultra-low-power devices. Intel has not confirmed the launch schedule, final specifications or market positioning, while claims about advanced CPU cores, integrated graphics and stronger AI acceleration remain unverified.
theCUBE’s coverage of Dell Technologies’ AI Leadership Symposium focused on the cost and control challenges enterprises face when moving AI workloads into production. Interviews with representatives from Dell, Intel, Deepgram, Singulr AI and MisaLabs addressed cloud token budgets, existing data-center capacity, runtime governance for AI agents, controlled data environments and private-cloud deployment. Deepgram described its work with Dell and Fortanix to run voice AI models natively on customer devices, including air-gapped systems; this was presented as an objective rather than a completed rollout. Dell sponsored the symposium, and theCUBE was a paid media partner, although sponsors did not have editorial control.
SAP executives told InfoQ that AI-agent proofs of concept may perform differently in production because enterprise data is inaccurate, business rules vary across systems, and integrations and exceptions are harder to handle. The company has released more than 300 agents and is organizing them around its existing ERP processes through Joule Work and Joule Studio, while SAP Business AI Platform provides governance and tracking.
MiTAC Computing Technology showcased a full-stack AI infrastructure ecosystem at SC26, combining high-density liquid cooling, disaggregated inference and end-to-end resource scheduling. The company and Skymizer proposed assigning Prefill to AMD Instinct GPUs and Decode to Skymizer HTX301 accelerators, while also showing rack systems, server platforms and the MiCoreView pod-management software. These capabilities and performance figures were presented by the vendors and were not independently tested in the supplied report.
科技新报, synthesizing analysis from Engadget and other publications, argues that laptop NPUs mainly trade some generality for low power consumption rather than improving AI performance across the board. It notes that TOPS figures are generally theoretical INT8 peaks, while practical NPU use is concentrated in background features such as meeting blur, captions, translation, and local search; in a cited OpenVINO test, Mistral 7B reached 98% NPU utilization. The source identifies limited use cases and software support as constraints, recommending Copilot+ PCs mainly for battery life and frequent video calls, and sufficient memory and bandwidth for local AI models.
Intel disclosed the first benchmark results for its graphics-free D270 data-center GPU, claiming up to 5.9 times the application throughput of NVIDIA's RTX PRO 6000 Blackwell Server Edition across four FP64-heavy financial workloads. The 350W accelerator also delivered 1.12 times the throughput on Qwen 3 30B-A3B and up to 29% lower latency on Drex 2.0 than the comparison card. TechSpot and other outlets noted that these were Intel's own results from pre-production chips and non-matching systems, so independent verification remains outstanding.
Hardware 360º used ChatGPT Astra to decompile and compare two ASRock BIOS versions, one containing SkyOC and one with the feature removed, according to Tom’s Hardware and Wccftech. The enthusiast says the AI identified the relevant workaround code and produced a modified BIOS targeting an ASUS ROG Maximus XI Extreme Z390 motherboard. The modified firmware has not been fully tested, so its ability to restore SkyOC on modern Intel platforms remains uncertain.
Lattice reported fiscal 2026 third-quarter revenue of $291 million, up 118% year over year, and approximately 90% growth in FPGA revenue, which CEO Ford Tamer attributed partly to accelerating AI demand from data centers. Citing Semiconductor Industry Observation, 36Kr says AMD, Altera and Lattice are deploying FPGAs in server management, security, rack control and data-path functions as AI infrastructure grows more complex. The report frames this as a shift from competing with GPUs for core computation toward programmable infrastructure roles that require management and control.
Cornelis has launched CN6000, a family of 800G network cards and switches designed for AI clusters and supercomputers, positioning the products as an alternative to Nvidia’s InfiniBand. The main card uses Intel’s Omni-Path technology and supports chips from AMD, Intel, Nvidia and Qualcomm, while Dell, Lenovo and Supermicro plan supporting systems. Cornelis says its card handles 1.6 billion messages per second and is 21% faster for small-message transfers, based on its own tests; partner systems are expected in early 2027.
Israeli investors say Nvidia's leadership transition may create an opening for AI-chip and inference-software competitors, though TechCrunch AI reports that stronger economics—not the change alone—will determine whether they win business. Nvidia employs about 6,000 people in Israel, where it operates a major R&D facility following its $6.9 billion Mellanox acquisition and has planned a multibillion-dollar campus. Colette Kress is serving as interim CEO while the permanent succession remains tentative.
LMSYS Org launched the SGLang Consortium during the opening keynote of the inaugural SGLang Summit in San Francisco, with NVIDIA, AMD, Intel, Google, Meta, Cloudflare, Modal and RadixArk as founding members. Members will assign technical representatives and attend two roadmap meetings annually, contributing engineering expertise, deployment feedback, testing resources or joint development, while SGLang maintainers retain authority over technical decisions and code acceptance. The stated priorities include frontier-model support, reasoning and multimodal workloads, distributed inference and hardware optimization.
36Kr reported that all three major U.S. stock indexes closed lower on November 12: the Nasdaq fell 0.39%, the Dow fell 0.36%, and the S&P 500 fell 0.35%. Nvidia fell more than 4%, with its market value reportedly shrinking by $182 billion. SK Hynix fell more than 3%, while Netflix and Intel fell more than 2%. Nvidia GPU competitors Broadcom and AMD gained more than 5%; Google and Tesla rose more than 1%, while Microsoft and Amazon posted modest gains.
WIRED AI reported that Nvidia announced Jensen Huang had resigned from the company. Less than an hour after markets opened, Nvidia had lost roughly $546 billion in market value. Although the stock had rebounded somewhat, the preceding weeks of scandal had reduced the chipmaker's valuation by more than $1 trillion; the report did not attribute the entire decline to Huang's resignation.
In a Bloomberg Power On interview, NUVACORE co-founder Gerard Williams III described the company's effort to develop powerful, energy-efficient CPUs for AI data centers. Williams said NUVACORE's “Core First” approach does not commit to x86, ARM or RISC-V at the core-design stage. According to IT之家, Williams also said AI tools reduced early performance modeling and design evaluation from seven to 12 months to two to three months; the report did not provide independent verification of that figure. Williams, who worked on Apple processors from the A7 through M1, argued that CPUs remain necessary alongside GPUs for memory scheduling, security, operating-system execution and serial tasks.
36Kr reported on November 12, 2026, that most large US technology stocks were lower in premarket trading. As of publication, NVIDIA was down more than 7%, SanDisk nearly 6%, SK hynix and Micron Technology more than 4%, and Intel nearly 3%. The report did not state the cause of the declines, and the figures reflected a premarket snapshot.
ASUS is expanding its Agent Computer initiative across motherboards, mini PCs, workstations and software, letting users choose local, cloud or hybrid AI for agentic workloads. The company says the ProArt GR1X mini PC, built on NVIDIA RTX Spark with up to 128GB of unified memory, can run locally hosted models with up to 120 billion parameters. ASUS is also developing the Zenni Claw agent platform and Multi-LM Tuner, while ASUS EZ Installer can deploy OpenClaw on mini PCs and NUCs.
The Next Web reports that GITEX GLOBAL 2026 will open with a one-day summit on 7 December at Dubai World Trade Centre, followed by a four-day expo from 8 to 11 December at Dubai Exhibition Centre in Expo City Dubai. The summit programme covers the compute economy, capital and future frontiers, while the expo includes AI infrastructure, physical AI, robotics and cybersecurity, with Huawei, Microsoft, AWS, AMD and other companies named as headline participants.
During China’s 2026 Double 11 period, AI token packages appeared on major e-commerce platforms, with JD.com reporting a 20-fold increase in package sales during a two-hour livestream and Tmall reporting nearly 100,000 Qwen Office memberships sold in 30 minutes. CCID Research Institute figures cited by 36Kr said China’s inference data volume had exceeded training data volume for the first time, while daily token calls topped 175 trillion by June. 36Kr also reported discounted packages from official sellers alongside unofficial resellers whose reliability, pricing transparency, data privacy and operational accountability remain unclear.
What problem it solves
Documents and office files often need to be transformed into a structured, text-friendly format before they can be processed by language models or text analysis tools. MarkItDown addresses this by preserving important document structures such as headings, lists, tables, and links during conversion to Markdown.
Key features
Broad format conversion — MarkItDown converts PDF, P…
What problem it solves
Businesses often need to connect multiple payment providers and add capabilities such as routing, retries, vaulting, observability, and reconciliation without rebuilding their payment stack. Hyperswitch addresses this by allowing teams to select and integrate the modules they need without unnecessary complexity or vendor lock-in.
Key features
Intelligent payment routing — Ro…
What problem it solves
Teams deploying AI inference need to optimize models, reduce resource demands, and run them across different frameworks and hardware platforms. The toolkit addresses model conversion, compilation, and inference workflows for applications ranging from computer vision and speech recognition to generative AI and language models.
Key features
Inference optimization — Optimize de…
What problem it solves
Modern documents mix text, images, tables, equations, charts, and multimedia, which traditional text-focused RAG systems cannot effectively process. This project provides an integrated workflow for parsing, analyzing, retrieving, and querying these content types through a cohesive interface.
Key features
End-to-end multimodal pipeline — The workflow covers document ingestion…
What problem it solves
Stock analysis often requires users to gather market quotes, news, technical indicators, financial information, and AI-assisted interpretation across multiple sources. The project addresses this by bringing supported-market data, analysis tools, AI agents, alerts, and review workflows into a locally focused desktop application.
Key features
Multi-market market data — The app…
What problem it solves
Coding and research workflows often require persistent context, reusable operating patterns, and coordinated agents across long-running sessions. Prime Agent addresses this through a programmatic environment, durable harness state, background continuity, and mechanisms for refining supplemental state from evidence.
Key features
Persistent programmatic environment — A persist…
What problem it solves
Monitoring global news, geopolitical activity, markets, conflicts, infrastructure, and other live signals requires bringing information from many domains into one view. The project addresses this with a unified situational awareness interface that combines curated feeds, maps, correlation, analytics, and programmatic access.
Key features
Curated global intelligence — Curated…
What problem it solves
Developers need an editor that combines code-aware assistance, flexible editing, remote workspaces, extensibility, and command execution in one environment. Lapce addresses these needs with built-in language server support, modal editing, remote development, WASI-compatible plugins, and a terminal.
Key features
Language Server Protocol support — Built-in language server inte…
What problem it solves
AI agents often produce plausible SQL without the business definitions, approved joins, units, and examples needed to make results trustworthy. Teams also need a reviewable way to manage that context and turn governed answers into shareable dashboards across existing data sources.
Key features
Governed semantic and context layers — Business meaning is represented through the…
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
Developers and coding agents need to understand how code is connected before making changes. GitNexus provides a graph-based view of symbols, callers, dependencies and execution flows to support code discovery, impact analysis and refactoring.
Key features
Graph-based code intelligence — Index a codebase as a knowledge graph covering symbols, calls, dependencies and executio…
What problem it solves
Server operators need to detect malicious behavior across logs and HTTP requests, then apply remediation at the appropriate layers of their infrastructure. They also need a way to benefit from threat intelligence shared by the community.
Key features
Integrated Security Engine — The Security Engine combines IDS/IPS and WAF capabilities in one solution. It analyzes log source…
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
Organizations need to structure, store, organize, and visualize technical and non-technical cyber threat information while linking it to primary sources. They also need to relate that information, infer useful connections, and exchange data with other platforms and tools.
Key features
Structured threat intelligence — Organize technical and non-technical information through a…
What problem it solves
Developers need a development environment that combines an IDE with terminal multiplexing while remaining accessible across machines. Rune also addresses language tooling, workspace organization, extensibility, and remote development through a connected application and headless mode.
Key features
Connected development workspaces — Rune instances can form an end-to-end encryp…
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
Legacy SASE platforms combine connectivity and security in heavyweight, closed, cloud-locked stacks assembled from years of patchwork. Administrators need an open, self-hostable platform that unifies networking and security while remaining straightforward to deploy and operate.
Key features
Connect remote networks — Site connectors use outbound tunnels and intelligent NAT tr…
What problem it solves
AI agents, assistants, and chatbots need access to repository context and GitHub capabilities to complete development and collaboration workflows. Without a direct connection, these tools cannot conveniently read repositories and code files, manage issues and pull requests, analyze code, or automate workflows through natural language.
Key features
Repository and code access …
What problem it solves
Complex financial questions require researchers to break them into manageable tasks, gather current market information, and evaluate the resulting analysis. The project addresses this by automating research planning, tool-driven data collection, and iterative self-validation.
Key features
Intelligent task planning — Complex queries are automatically decomposed into structure…
What problem it solves
OpenPencil lets users describe UI in natural language and see it appear on an infinite canvas, then modify existing designs by selecting elements and chatting. It also lets multiple AI agents work on different sections of complex pages in parallel while keeping `.op` files human-readable, Git-friendly and diffable.
Key features
Prompt-to-canvas generation — Describe a UI in …