Apple is reportedly developing an enterprise AI server using its M-series Ultra chips, expected to debut in 2029 in configurations with two or four future M8 Ultra chips. This would mark Apple's first server product in nearly two decades, driven by surging demand for Mac minis and Mac Studios among AI developers -- OpenAI has reportedly bought tens of thousands of Mac minis for reinforcement learning training, and Anthropic has rented them via AWS. Apple is also reportedly in talks with Nvidia about using NVLink Fusion networking technology to connect the chips.
The article profiles seven companies developing AI-powered humanoid robots for factories, warehouses, logistics, customer service, and homes: Figure AI, AgiBot, Unitree Robotics, Tesla, UBTech Robotics, Apptronik, and Boston Dynamics. It covers their robots, commercial progress, production figures, partnerships, and market shares, and notes that humanoid robots still struggle with unpredictable tasks, speed, reliability, and cost compared with people and specialized industrial robots.
Architect Labs is an AI research lab focused on co-designing, verifying, and building custom chips end-to-end with AI. Its posts cover Redwood, an AI-designed accelerator; the lab's AI system for silicon development; and functional RTL generation using curriculum learning and reward shaping.
Architect Labs announces Redwood, a frontier AI accelerator autonomously designed, verified, and deployed from a human-written spec in under two weeks, including RTL, verification, firmware, and compute kernels. A scaled Redwood Nano configuration runs Qwen3-0.6B on an AMD Versal FPGA at 12.1 tokens/sec, with projections showing significant throughput and power gains on a future Samsung 8nm target versus Jetson Orin Nano.
WiCi promotes a wireless GPU-compute layer for personal AI that lets devices use remote GPU capacity as if it were local. Its forthcoming WiCi One targets AI agents, local LLMs, video work, 3D rendering, and home robots.
The Register reports on Microduck, a $399, 25-centimeter robot offered through Hugging Face and built by Pollen Robotics. The open-source robot is designed to work out of the box and is pitched as a playful physical platform for AI coding and experimentation.
Perplexity and Nvidia launched Portable Computer, a locally run version of Perplexity’s agentic Computer platform for Nvidia DGX Spark and RTX-equipped Linux systems. The product keeps models, files, and work on-device by default, with optional permissioned escalation to cloud models.
In this Bloomberg Tech video, OpenAI says its new Jalapeno chips outperformed Nvidia’s current lineup in internal testing. OpenAI chip chief Richard Ho discusses gains in AI work per unit of power and response speed.
This YouTube video compares Mac Studio M4 Max, NVIDIA RTX 5090, AMD Strix Halo, and cloud GPUs for running local LLMs, covering Qwen 3 Coder Next, Qwen 3.6 27B, memory bandwidth, unified memory, VRAM limits, and why llama.cpp can outperform vLLM on consumer GPUs. It gives hardware buying recommendations by budget and workload, concluding that local AI currently works best alongside frontier cloud models rather than as a full replacement.
Nate B. Jones compares the NVIDIA DGX Spark, Mac Studio, and RTX 5090 for local AI use and lays out a full 'personal AI computer' stack: hardware choice by buyer profile, runtimes (Ollama, llama.cpp, vLLM, MLX), local models (Llama 4, GPT-OSS, Qwen, Gemma), an owned memory layer (his open-source 'Open Brain' project, Postgres/pgvector, SQLite), MCP-based tool access, and interface/voice layers. The core argument is that owning your compute and memory substrate compounds knowledge over time, even as cloud models remain useful for the hardest tasks.
This Tom's Hardware article reports that China's Moonshot AI allegedly used Nvidia Blackwell GB300-equipped servers and other advanced U.S. chips to train its Kimi K3 model despite U.S. export restrictions and Chinese import controls. It states Moonshot reportedly relied on two Chinese firms with Blackwell chips and joined multiple eight-chip servers across datacenters, with plans to seek more access for a future Kimi K4. The article notes Moonshot uses Nvidia HGX H20 GPUs for inference, while U.S. officials investigate possible unauthorized access to advanced chips in China.
This YouTube video documents a three-month effort to diagnose and fix slow performance on a 4-node Minisforum MS-01 Max Strix Halo cluster used for local LLM inference. The creator initially assumed bandwidth was the bottleneck but discovered the real issue was latency, leading to a switch from TCP-based multi-node setups to RDMA and tensor parallelism with vLLM. The video compares performance across models and configurations, highlighting the importance of matched RDMA hardware and proper quantization for multi-node AI workloads.
This Codemotion article explains DwarfStar, an MIT-licensed inference engine written in C by Salvatore Sanfilippo (antirez) for running large frontier language models locally, especially DeepSeek V4 Flash, on Mac computers. It describes how DwarfStar uses 2-bit quantization, SSD streaming, persistent KV-cache sessions, and distributed multi-machine inference to make very large MoE models practical on consumer hardware. It presents benchmark numbers and notes the project's current beta status.
This PrismML announcement covers 1-bit Bonsai, a family of 1-bit LLMs built to deliver high intelligence density on personal devices. The 8B model uses 1-bit weights across embeddings, attention, MLP, and LM head, runs with a 1.15 GB memory footprint, and is reported to be competitive on benchmarks with strong speed and energy efficiency on devices like an iPhone 17 Pro, M4 Pro Mac, and RTX 4090. It also introduces smaller 4B and 1.7B variants and discusses platform support and licensing.
This X/Twitter post from Cointelegraph reports that SpaceX unveiled its first AI satellite, featuring a 150 kW peak compute payload and a 70-meter wingspan. Elon Musk described the satellite as simpler but larger than a Starlink satellite, signaling SpaceX's move into orbital AI compute infrastructure.
This YouTube video/transcript from Dwarkesh Patel's podcast features Nvidia CEO Jensen Huang discussing the company's supply-chain moat, TPU competition, CUDA's software ecosystem, and whether Nvidia should become a hyperscaler. It also covers China export policy, AI chip security, and Nvidia's investments in AI labs like OpenAI and Anthropic.
This Forbes/CNBC video features Groq's cofounder discussing a landmark deal with Nvidia, described as Nvidia's biggest deal in its three-decade history, reached at a precarious moment for Groq. It explains that Nvidia is betting on Groq's inference-speed technology as part of a newly announced chip platform. The segment covers the strategic rationale behind the partnership for both companies.
This X post from Cointelegraph reports that China's Wuji Tech confirmed its 22-degree-of-freedom robotic hands were used in a viral demonstration by Genesis AI. The demo showcases advanced autonomous manipulation powered by a single neural network. The post highlights progress in AI-driven robotic hardware capabilities.
This video explains how a missile strike and shutdown at Qatar's Ras Laffan refinery complex threatens global helium supply, a critical input for advanced semiconductor manufacturing including EUV lithography and chip testing. It argues the disruption could raise energy costs and chip prices across South Korea, Taiwan, and Western fabs, while giving China a geopolitical advantage through domestic helium and energy sources like the Power of Siberia 2 pipeline.
Content behind a login wall / unavailable — inferred from link context: this Google share link could not be accessed directly, but the accompanying chat message indicates the linked article is titled 'Racing to Find the Best Robots at Nvidia GTC.' It likely covers robotics demonstrations and announcements from Nvidia's GTC 2026 conference, a recurring topic in this batch given other entries discussing Nvidia's robotics and orbital data center ambitions.
This product page introduces the OrangePi AI Station, an AI mini PC powered by an Ascend 310-series processor delivering 176 TOPS of AI compute, 16 CPU cores, up to 96GB LPDDR4X memory, and NVMe SSD support. It lists hardware specs, supported software, and target use cases such as smart security, transportation, healthcare, agriculture, and industrial automation. A chat message expresses surprise that such devices arrived faster than expected. Source: http://www.orangepi.org/html/hardWare/computerAndMicrocontrollers/details/Orange-Pi-AI-Station.html
This Reuters report says Beijing instructed several Chinese tech firms to stop ordering Nvidia's H200 chips while it considers mandating domestic AI chip purchases. China is seeking to reduce dependence on U.S.-designed semiconductors amid ongoing export-control tensions. The article notes U.S. export licenses for the H200 remain under review, while Nvidia's CEO said Chinese demand for the chip stays strong.
This Medium article argues China advanced its EUV lithography capabilities by recruiting experienced engineers, including a former ASML scientist, rather than by reverse-engineering hardware. It claims a working prototype was completed at a high-security Shenzhen facility, with EUV light generation reaching 3.42% conversion efficiency and chip production targeted for 2028-2030. The piece frames this within broader discussion of semiconductor export controls and the strategic value of human expertise.
This How-To Geek article covers Lisuan, a Chinese startup that launched the G100, a domestic gaming GPU on a 6nm process positioned to compete with Nvidia's RTX 4060. It reviews earlier Chinese GPU efforts from Jingjia Micro and Moore Threads that struggled with driver maturity, and notes early G100 benchmark leaks appear weaker than claimed, though the chip's ability to boot Windows and run DirectX 12 is still an engineering milestone.
This Sharp Tech podcast episode with Ben Thompson discusses the tradeoffs between Nvidia GPUs and Google TPUs in AI hardware. It explains that GPUs are more programmable and flexible while TPUs are simpler and more efficient but harder to program, affecting adoption and software support. The conversation also covers why Nvidia's CUDA software layer sustains its market strength and why Google's TPU strategy focuses on internal use rather than external chip sales.
This CNBC video explains the major categories of AI chips, comparing Nvidia GPUs with custom ASICs like Google's TPUs and Amazon's Trainium/Inferentia. It covers why GPUs became central to AI training, why growing inference demand favors smaller, cheaper custom chips, and how hyperscalers including Google, Amazon, Meta, Microsoft, and OpenAI are designing their own silicon. It also touches on edge AI chips, TSMC manufacturing, and geopolitical and power constraints on the AI chip supply chain.
This video explains the reported Nvidia-Groq deal as a non-exclusive licensing agreement combined with an acquihire-style transfer of Groq's founder and key leaders to Nvidia, while Groq remains independent under a new CEO. It argues the deal reflects a broader AI industry shift from traditional acquisitions to 'capability transfers' focused on buying inference technology and talent. The discussion centers on memory bandwidth, high-bandwidth memory, CoWoS packaging, and SRAM-heavy inference chip designs shaping competition in low-latency AI serving.
This Reuters-produced YouTube news roundup covers a report that China secretly built a prototype chipmaking machine in a six-year effort compared internally to the Manhattan Project, aimed at rivaling Western AI chip technology. It also briefly covers unrelated news including a Trump holiday address, Warner Bros Discovery rejecting Paramount Skydance's takeover bid, and a murder case involving Rob Reiner's son.
This Reuters exclusive (via Yahoo Finance) reveals that China has built a prototype extreme ultraviolet (EUV) lithography machine in Shenzhen, developed by former ASML engineers and coordinated with Huawei, as part of a six-year push for semiconductor self-sufficiency. The article notes major remaining technical hurdles, with insiders targeting functional advanced chips by 2028, though some consider 2030 more realistic.
This VentureBeat article (accessed via a Google share link) reports that startup Quilter used physics-driven AI to design a fully functional two-board Linux computer with 843 components in about 38.5 hours of human labor, versus an estimated 428 hours for human engineers. The resulting board, running NXP's i.MX 8M Mini platform, booted Debian Linux successfully on the first attempt with zero design rule violations, highlighting AI's growing role in automating printed circuit board design.
This product page describes Pocket, an AI-powered wearable device paired with software that transcribes conversations into notes, action items, mind maps, and searchable summaries. It highlights features like speaker labeling, cloud history, privacy controls, and tiered pricing (Free, Pro, Enterprise), alongside customer testimonials.
This Medium article argues that AI has begun to dominate the quantum computing stack, using machine learning to help build, debug, and stabilize quantum systems. It frames this shift as accelerating a broader U.S.-China race for computational power, funding, and scientific dominance in supercomputing.
This Wikipedia article covers MetaX Integrated Circuits, a Shanghai-based company founded in 2020 that designs general-purpose GPUs and AI accelerators, often compared to Nvidia. It details the company's founders, funding history, patents, 2025 IPO filing, and its chip product line including the C500 and N100.
Futurism reports on a video, shared by Salesforce's Marc Benioff, showing Tesla's Optimus humanoid robot struggling to answer a simple question and walking awkwardly through an office. The piece contrasts Musk's ambitious robotics claims with reported production delays and technical setbacks, and notes competition from China's Unitree, whose cheaper robot appeared more capable.
Content behind a login wall / unavailable — inferred from link context: based on the URL, this Register article likely covers the release of DeepSeek's V3.1 model and reported hints about the Chinese-made chips used to train or run it, relevant to the broader US-China AI hardware competition.
The Trump administration is reportedly in talks with Intel about the US government taking an equity stake in the chipmaker, according to Bloomberg sources. The potential deal is described as part of the White House's growing willingness to blur lines between government and industry. It could support Intel's delayed Ohio factory hub, once billed as poised to become the world's largest chipmaking facility.
This article covers a reported Chinese breakthrough in mass-producing high-quality indium selenide (InSe) semiconductor wafers, framed as a step beyond silicon-based chips. It explains potential advantages for military, AI, cloud computing, consumer electronics, and medical applications, while noting that silicon remains dominant due to established lithography and fabrication ecosystems. The piece frames this as a strategic materials-science development in the broader US-China semiconductor competition rather than an imminent 'silicon killer.'
This article reports on Chinese researchers using a 2-watt laser from a satellite at 36,000 km altitude to reportedly transmit data at 1 Gbps, claimed to be faster than Starlink's capability. It describes the technical approach, called AO-MDR synergy, combining Adaptive Optics and Mode Diversity Reception to counter atmospheric turbulence, tested at the Lijiang Observatory with a path-picking algorithm to improve signal stability. The chat context notes skepticism about the sensational headline, suggesting the underlying technical claim is more nuanced than presented.
This Bloomberg article reports that Neuralink expects to implant its brain chips in 20,000 people per year by 2031, generating at least $1 billion in annual revenue, according to documents reviewed by Bloomberg. The company plans to operate about five large clinics within six years and offer at least three device versions: Telepathy for brain-to-machine communication, Blindsight for vision restoration, and Deep for treating tremors and Parkinson's disease. This reflects Elon Musk's brain-computer interface company's expansion plans.
This video explains IMEC's semiconductor technology roadmap progressing from 2nm chips in 2025 toward 0.2nm by 2037, covering transistor evolution from FinFET to gate-all-around and CFET designs. It discusses backside power delivery, next-generation lithography like High NA EUV, and emerging materials such as MoS2 and carbon nanotubes. The video connects these hardware advances to AI compute growth and implications for companies like TSMC, ASML, NVIDIA, and Intel.
This article reports that Johns Hopkins University researchers trained an autonomous surgical robot called SRT-H (Surgical Robot Transformer-Hierarchy) to perform gallbladder removal surgery. The robot analyzed surgical videos, responded to voice commands, and adapted in real time during trials on a realistic model, completing the procedure multiple times with reported 100% accuracy and no human intervention. The system is not yet ready for use on actual patients, and the piece discusses ethical considerations and future directions.
This Hackster.io article describes a project by Binh Pham that packages a local large language model and its inference hardware onto a USB stick using a Raspberry Pi Zero and llama.cpp. The build required removing ARMv8-specific instructions because the Pi Zero's ARMv6 processor couldn't compile the software as-is. The system works best only with very small models, with a 15M parameter model processing tokens in about 200 milliseconds versus 2.5 seconds for a 77M parameter model.