Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Starting in 2027, Microsoft will deploy Maia at multi-gigawatt scale and make it the main inference system for its products. The next-generation Maia 300 uses compute chiplets on TSMC's N2P process and targets long-context, multi-step agent workloads. Digitimes reports that the supply chain is already preparing for a surge in server builds in 2027.
AMD's Helion agentic AI rack is designed for flexible AI infrastructure that caters to hyperscale deployments and enterprise needs. It supports a range of configurations, including CPU-only and GPU-integrated systems, aligning with diverse AI workloads. Helion, part of an upcoming open blueprint by AMD for OCP, is set for release in November.
Physical Intelligence opened a beta API for its π robotics foundation models: upload demonstrations, fine-tune on Pi-managed compute, and run the result on your own robot through a hosted inference gateway. Early partners spanned hardware from Actor's autonomous wheel loader, running a commercial deployment with no operator onboard, to Almond's dual-arm Axol, which learned to deburr its own spacers in four days. Since most ran cloud-hosted inference, network connectivity was a pain point for some.
This deep dive uses the Emmy compiler framework to follow PyTorch operations down to CUDA one lowering stage at a time: Tensor IR, Loop IR, Tile IR, Kernel IR, and finally emitted code. It uses RMSNorm and matrix multiplication as worked examples. It also covers operator fusion, schedule selection, and online attention rewrites, so readers can see what Triton-style stacks do under the hood.
Sponsored
Ads get ignored on social media but not in TLDR! Reach hardware engineers and makers where they actually pay attention. Learn more about sponsorship opportunities [8].
Apple Silicon Macs have powerful GPUs that sit mostly idle in normal use. Minions combines them across machines into a virtual GPU cloud for large ML inference jobs, with fault tolerance and flexible task scheduling. The pitch is that teams can use hardware they already own instead of waiting for expensive, centralized GPU capacity.
Anthropic's Frontier Red Team ran Gemma4-26B-A4B across 15 off-brand Android TV boxes with 4GB of RAM each, using pipeline parallelism at 0.39 tokens per second, to test whether small models could replicate across cheap IoT hardware. A 9B model got inference running on 8 of 9 hardware types tested, and a Kimwolf-sized botnet would total only about 40 H100s of compute. By their trend line, small models reach the needed capability in simplified tests by year-end.
Apple's Home Collection, built with LG and Schneider Electric, covers cameras, thermostats, sensors, a doorbell, and a deadbolt. LG's lock uses Ultra Wideband to unlock as you approach and works locally over Matter and Aliro, while its indoor cam physically tilts away when someone's home. Every device supports Thread, and cameras keep recording to built-in storage if Wi-Fi drops. Prices run $199.95 to $399.95, with availability starting in December.
An Abilene-scale build requires roughly $17B for chips and servers, $3B for power, and only $2.4B for construction. Transformers, generators, and cooling equipment face long lead times and rising prices.
Walmart has spent a decade trying to automate about 200 US warehouses.
Each core runs a 4-stage pipeline and replaces work done by general-purpose CPUs.
The $199 watch has a color e-paper display, runs open-source PebbleOS, and lasts about 8 days per charge.
Crescent Island, Intel's data center inference GPU, will be sold as the GPU D-Series.
At OCP, Meta showed IcePack, a 102.4T liquid-cooled network rack, along with its IceCube switch platform.
Who's building on-prem compute in 2026? It turns out a lot of people! Oxide Computer Company, which builds custom on-prem cloud compute systems, has closed a massive $445 million Series D funding round led by Eclipse, with participation from major investors including Jane Street, Atreides Management, and chipmaker AMD. The capital injection follows the startup achieving operational profitability and will be used to scale manufacturing capacity, secure raw material inventory, and fulfill a rapidly expanding order backlog for their commercial cloud computer racks.
SpaceX agreed to buy up to 14 MHz of paired 800 MHz low-band spectrum, which penetrates buildings, and plans to deploy its own antennas on towers and rooftops to use it, with most existing phones already supporting the band. Separately, the FCC approved 15,000 more Starlink satellites optimized for 2 GHz. Together they point to a hybrid network, terrestrial indoors and satellite beyond towers. The deal needs FCC approval, and Verizon, AT&T, and T-Mobile shares fell roughly 6% on the news.
In this rare peek under the hood, we get a glimpse of how Foxconn uses NVIDIA Omniverse, Isaac, and Metropolis platforms to create physically accurate digital twins of manufacturing facilities that assemble next-generation AI server hardware. By simulating factory layouts, automated optical inspection systems, and autonomous mobile robot fleets in virtual environments before physical deployment, the company minimizes factory commissioning delays and boosts production efficiency. This approach showcases how system architects and automation specialists can leverage high-fidelity simulation to safely validate and optimize complex robotic workcells prior to hardware installation.
After May's pad explosion destroyed its transporter-erector, Blue Origin is rebuilding LC-36 for vertical integration, loading New Glenn through the support tower alone, as with Starship. Crews destacked the tower, added cryogenic propellant plumbing on the ground, and have begun restacking modules, while the launch bridge is restored. Umbilicals and fairing access platforms haven't appeared yet. Bezos targets a December return to flight carrying Blue Ring, though he joked it could slip to December 31.
A Nature Computational Science briefing describes an in situ learning method where photonic integrated circuits directly measure optical gradients, so diverse circuit layouts can be trained after they're built. That removes the need for a full digital model of the chip, which optical training normally leans on. The team demonstrated matrix design, multiwavelength computation, and rapid meta-learning on the hardware. Specifics like accuracy or energy figures aren't included.
ScaRF-SLAM is a dense visual mapping system that pairs the reliability of classical visual SLAM with modern feed-forward geometric foundation models for 3D scene reconstruction from the Oxford Robot Dynamics Lab. By decoupling camera tracking from dense depth estimation and optimizing scale consistency across submaps, it achieves globally accurate 3D maps across diverse sensor configurations. For perception engineers and autonomy developers, this framework provides a practical way to produce metric-scale environmental maps in real time without sacrificing tracking stability
ZeroWire Robotics has launched the Q8botOne, a smartphone-sized open-source quadruped robot featuring an innovative wire-free chassis where the main printed circuit board serves as the structural frame. Driven by an ESP32-C3 microcontroller and eight DYNAMIXEL smart bus actuators with parallel-linkage legs, the compact robot can walk, trot, jump, and carry twice its 250-gram payload.
Ledger confirmed an unauthorized hardware implant in one affected user's device, bought from Southeast Asian reseller CryptoBilis, pointing to tampering after manufacturing rather than a breach of Ledger itself. Mark Karpelès published images of a modified Nano X with a small board behind the screen and cellular components, which would presumably capture the 24-word recovery phrase as it displays during setup, leaving the secure element untouched.
Ting-Wei Sun pleaded guilty to four counts, including conspiracy to violate export controls, in a scheme that routed about $2.5 billion of Nvidia-powered servers through a Southeast Asian company to buyers in China.
Jumper is an open-source biomimetic robot that replicates crab locomotion, combining sideways multi-legged walking with an explosive spring-assisted jumping mechanism.
ASML will raise prices 10% on all EUV and DUV replacement parts supplied to South Korea starting with January 2027 deliveries, and Samsung and SK Hynix have accepted
EasyNav 0.5.0 has officially launched, delivering an accessible yet comprehensive open-source navigation stack with upgraded path planning, improved recovery behaviors, and reduced compute overhead for ROS 2 systems.
TSMC reported revenue of $511.86 billion New Taiwan dollars for September, a 54.6% year-over-year increase. Demand for advanced AI chips, driven by customers like Nvidia and Apple, boosted sales. TSMC has also committed to using ASML's advanced lithography machines to enhance chip production.
Amazon built its 1,000th Leo broadband satellite at its Kirkland, Washington plant, which took six months to produce its first 27 and now turns out 27 a week. It builds propulsion in-house and borrowed consumer electronics methods to cut tests that took weeks down to hours. The bottleneck is launch: only 396 are in orbit against 3,232 needed, and the Vulcan grounding and a New Glenn pad explosion cost Amazon FCC spectrum priority.
IEEE Spectrum examines the up-to-$1B CHIPS R&D award finalized for IBM spinoff Anderon in September. IBM is matching that funding with $1B, while Anderon uses a 300mm Albany line and plans foundry services and process design kits for outside developers. The challenge is repeatable fabrication, including tighter control of Josephson junctions, rather than simply adding wafer capacity.
Tensor virtualization decouples a tensor's logical layout from its physical GPU allocation, so one shader template replaces separate backend-specific codebases. For edge LLMs, ML Drift switches kernels between compute-bound prefill and memory-bound decode, using a convolution-aligned KV cache layout and in-kernel activation quantization to skip memory roundtrips. It cut YouTube Shorts frame latency by up to 40%, and Google co-optimized it with Arm, Intel, and Qualcomm GPUs.
Scientists from the Harbin Institute of Technology developed a cement-based supercapacitor capable of powering small devices like LEDs. The device uses a 3D-printed electrode ink made from carbon nanotubes and cement, enabling fast energy discharge. Future research will enhance cold-weather performance, potentially allowing buildings to store energy and support smart environments.
The toolchain covers synthetic data generation, training, simulation and validation, edge deployment and continuous improvement. It uses SageMaker for training and IoT Greengrass for pushing models to devices, plus NVIDIA's Isaac Sim, Isaac Lab, GR00T, and Cosmos. AWS says it is not a direct replacement for RoboMaker, which shut down in 2025.
The toolchain covers synthetic data generation, training, simulation and validation, edge deployment, and continuous improvement.
Gartner predicts that over half of quantum startups will fail by 2030 due to a market unable to generate sufficient revenue. Despite this, companies are investing in quantum technologies for their potential business advantages, focusing on building expertise rather than waiting for technological maturity. The public sector and financial industries are expected to lead quantum computing investments, with spending reaching $245 million and $289 million by 2027 and 2028, respectively.
Apple gained a full percentage point in market share year-over-year as RAM price increases led to a 20% decrease in overall PC shipments. Despite this downturn, Apple was less affected than most competitors. With prices expected to worsen, Apple's resilience highlights its strong customer base willing to invest in new devices.
NVIDIA expanded its Halos safety system from self-driving cars to robots with Halos for Robotics in June.
Microsoft's 15-inch Surface Laptop Ultra runs Nvidia's RTX Spark superchip, pairing up to a 20-core Grace CPU and Blackwell graphics with up to 128GB of unified memory, enough for local agents on models up to 120B parameters.
Sponsored
Building a data center takes 3–5 years, and about a third of the power goes to cooling and building systems instead of compute. Runware's Sonic Pods [5]skip all of it. Each pod is a complete AI inference data center in a 20 ft shipping container [6], trucked in and live in days: * Up to 1,200 GPUs and 1 MW of compute, any GPU architecture * Closed-loop liquid cooling on every GPU: 1.5 m³ of water, recirculated, none consumed * Europe is live, US rollout underway Capacity is filling up. Reserve a pod [7] or start on Serverless [8], billed by the second.
Ghost opened pre-orders for Core, a $3,499 screenless computer with an Nvidia RTX Pro 4000 SFF Blackwell GPU. The device runs preinstalled AI models locally and gives users access through a phone app. Ghost raised an $11 million seed round and plans to ship the first batch in late October.
The FT reports SpaceX is raising about $10 billion in bank loans and $30 billion in investment-grade debt to fund an Nvidia chip order, with Apollo leading and helping sell the debt to a broad investor base, and bond fund Pimco among lenders in talks. The deal is expected to close in 2027. It fits the pattern Nvidia flagged in August, when it disclosed its own SpaceX stake and a $500 billion Wall Street financing push, with chip purchases increasingly funded by debt and equity ties to the buyer. Reuters could not independently verify the report, and SpaceX, Apollo, Nvidia, and Pimco haven't commented.
GPT-Synopsys is an OpenAI frontier model trained and tuned to operate Synopsys' EDA tools directly, covering PPA optimization, timing closure, and verification. Engineers hand it an objective, and agents run the tools, interpret the results, make changes, and loop until there's a verified outcome to review. It runs on OpenAI-hosted infrastructure and plugs into Synopsys.ai and the Autopilot agentic platform. OpenAI licenses Synopsys' tools to train the model, paying a training subscription fee, and the two companies then share revenue based on how well the model improves chip designs, an unusual structure that ties both to the results. Synopsys says customer design data stays protected and won't be used to train the frontier model, and the bundled service covers compute, the model, and EDA licenses.
Engineers must test chips against combined radiation, thermal shock, vibration, and other stresses rather than qualify each condition alone. Arizona State University pairs a radiation source with a cryostat to test devices at 77 Kelvin. Digital twins can model package and system behavior, but physical tests still check whether those predictions hold.
RoboParty demonstrated RP1 recovering its balance after visitors pushed or kicked it at IROS. The platform combines actuator modules, robot hardware, and reinforcement-learning motion control, with peak joint torque up to 160 Nm. RoboParty plans to release more control software, simulation tools, and hardware details, so its full open-source roadmap remains in progress.
ScaRF-SLAM is a dense visual mapping system that pairs the reliability of classical visual SLAM with modern feed-forward geometric foundation models for 3D scene reconstruction. By decoupling camera tracking from dense depth estimation and optimizing scale consistency across submaps, it achieves globally accurate 3D maps across diverse sensor configurations. For perception engineers and autonomy developers, this framework provides a practical way to produce metric-scale environmental maps in real time without sacrificing tracking stability.
Finnish authorities are investigating allegations that Google's local operator cleared over 300 hectares of forest before a required environmental assessment. A conservation group has requested a suspension of work near Muhos. Google says the tree felling complied with forestry requirements and that it protected high-value nature zones.
OpenTPU is a complete, open-source AI accelerator project that packages custom SystemVerilog RTL, instruction set architecture, a bit-exact simulator, and a compiler into a single accessible repository. The design runs modern quantized language models directly on cost-effective Kintex-7 FPGA PCIe cards, matching simulation results bit for bit. This architecture provides embedded systems designers and compute engineers an invaluable end-to-end reference for deploying and tailoring custom neural network hardware without relying on proprietary black-box tooling.
Lisa Su said supply grew through 2026.
Meta's Gregory Price presented hardware-offloaded compression at Linux Plumbers.
Anduril will invest $3.7 billion to build a two-million-square-foot "software-defined" shipyard at Sparrows Point, the former Bethlehem Steel site near Baltimore, and the Navy separately awarded it a $2.9 billion "Arsenal 2" contract with payments tied to production outcomes, bringing the package to roughly $6.6 billion.
Hex is an open-source, six-legged robot platform designed to give makers and developers an accessible foundation for building and testing legged locomotion.