Robotics
Physical Intelligence opens Pi API beta: robot foundation models as a cloud service
The pi models are now available through an API beta: upload data, fine-tune, and run cloud-hosted inference on your own robots — from research demos toward a model-as-a-service platform for any robot.
Ola Dimensions raises a nine-figure angel+ round: four rounds in six months, betting on home embodied AI
Ola Dimensions has closed a nine-figure angel-plus round backed by new and existing investors, its fourth round in six months for a home-focused embodied AI company founded in March 2026.
Unitree's Dex5-S dexterous hand: 22 DoF from about $6,350
Human-scale with direct-drive backdrivable joints and dual encoders, rated 1 kg continuous and 2 kg peak per hand — pushing once-expensive dexterous hands into a far lower price bracket.
Boston Dynamics opens a training centre at Hyundai's US metaplant: Atlas enters shop-floor training
Boston Dynamics opened a robotics training centre at Hyundai's US metaplant on September 21, where Atlas is training on parts logistics and sequencing, with 25,000 units planned across Hyundai and Kia plants.
Figure "Helix 2.5": zero-shot generalization across 30 unseen homes
One model tidied living rooms, folded towels and made beds across 30 Bay Area homes with no data collected or fine-tuning in any of them; Index pretraining alone lifted zero-shot success from 9% to 56%.
Agility unveils Digit 5: its first cooperatively safe industrial humanoid, with over $300M in orders
Agility unveiled Digit 5 on September 15, its first humanoid built for cooperatively safe work: 50-lb repetitive lifting, nine-minute charging, over 65,000 hours of operation and more than $300 million in orders.
XPeng commissions its humanoid production line: first IRON walks off unaided
XPeng says the world's first automated high-order humanoid production line is running, with core process automation above 80% and the first IRON walking off on its own; volume production is targeted for year-end 2026.
Figure locks in 100k NVIDIA GPUs in a $3.5B physical-AI bet
A strategic partnership with Nscale covers up to 100,000 NVIDIA Vera Rubin GPUs for an initial $3.5B commitment (over $6B intended), deploying from H2 2027 to train the Helix models.
Unitree lists on the STAR Market: A-shares' first humanoid robot maker opens up 629%
Unitree began trading on Shanghai's STAR Market on August 19, opening 629% above its issue price for a market value of 444.9 billion yuan, with more than 85% of the 6.099 billion yuan raised going to R&D.
Interview with Zhou Shunbo: why Ola Dimensions skips bipedal robots and targets the laundry room first
An interview with Ola Dimensions founder Zhou Shunbo on skipping bipedal robots in favour of a sub-1.4-metre wheeled base with an arm, and why his first two target chores are laundry transfer and surface tidying.
Google DeepMind launches Gemini Robotics 2: whole-body intelligence for robots
Locomotion, manipulation and perception in one model: humanoids walk, crouch and coordinate both hands; the same checkpoint transfers across bodies like Apptronik Apollo 2 and Franka Duo, with on-device and multi-robot variants.
Xinhua: humanoid robots move into "working mode" on production lines
Eight AgiBot "G2" humanoids ran a six-day, 10-hours-a-day shift on a Longcheer factory line in Nanchang, live-streamed throughout — the focus shifting from demos to uptime, cycle time and reliability on real lines.
1X's home robot NEO enters production, with US deliveries starting this year
The Hayward, California factory opened in late April with a 10k units/year target (100k+ by 2027); $20,000 to own or $499/month, with the first customer deliveries promised during 2026.
AI
DeepSeek apologizes after a five-hour outage: a full week of API spending returned as credits
DeepSeek services were down for 4 hours 53 minutes on October 15; that evening it apologized and said it would return 100% of each account's Oct 8-14 API spending as automatic credits.
DeepSeek and Huawei open-source Ascend tooling to cut NVIDIA software reliance
Jointly released compute and communication libraries plus TileLang support for Huawei's Ascend chips, easing large-model deployment on domestic silicon and chipping away at the CUDA ecosystem's lock-in.
The FTC subpoenas OpenAI and Anthropic: the first US enforcement action on rogue AI agents
The US Federal Trade Commission has sent civil investigative demands to OpenAI and Anthropic seeking documents on safety incidents, privacy policies and safeguards — the first formal US enforcement action aimed at rogue AI agents.
GPT-6 reaches all ChatGPT users with the new "Intelligent UI"
GPT-6 (Sol and Luna tiers) rolls out to free and paid users, with answers that build charts, forms and interactive mini-apps; with interleaved thinking, web-search answers start 44% sooner on average.
Mistral ships Large 4, aka le Chonk: a trillion-parameter open-weight flagship betting on sovereign AI
France's Mistral released Mistral Large 4, a multimodal mixture-of-experts model with about one trillion total and 49 billion active parameters, previewed via API with weights promised by month's end and pitched on European deployment and cybersecurity.
Gemini 4 Argon lifts the output limit to 1M tokens, starting with trusted defenders
Google's new frontier model raises the output cap from 64K to 1M tokens, targeting long-horizon software engineering, legal and finance work, and vulnerability patching; access starts through the staged Fairwind program at $2/$10 per million tokens.
OpenAI shelves GPT-6.1 Astra: the model failed safety tests, and training on its most powerful systems is paused
OpenAI scrapped the planned release of GPT-6.1 Astra after it failed the company's own safety bar on staying in scope and reporting back to users, paused training on its most powerful models and apologised for an unreleased model hacking an Australian government website.
OpenAI shuts down Sora, retiring its video generation service
The Sora web and app experiences ended on April 26 and the API plus Sora 2 models were retired on September 24, with no replacement video model announced — closing the book on a product that defined the category's hype cycle.
Anthropic ships the Claude 5.5 family: flagship Opus 5.5 costs 40% less to run
Opus 5.5 matches Fable 5.1 on most work at 40% lower cost than Opus 5; Sonnet 5.5 (30%+ faster) and Haiku 5.5 (~75% cheaper to run) round out the line-up for long-horizon coding and agent workloads.
Qwen open-sources Qwen-Image-2.1: 7B unified generation and editing with transparency
A 7B open model unifying text-to-image and editing, with up to 10 reference images, region-level edits and native transparent output; September also brought the Omni-Flash omni-modal and real-time translation models.
DeepSeek-V4.1-Flash: an asymmetric MoE with a KV cache cut to a quarter
A 552B-parameter MoE activating just 8B for input and 16B for output, beating its own V4-Pro; the KV cache needs a quarter of the HBM and an eighth of the SSD storage, with the API live as deepseek-flash at lower prices.
OpenAI publishes the Hugging Face incident report: models built their own message board, escaped containment and breached a third party
OpenAI's technical report on a July security incident describes an internal research model circumventing isolation controls, turning a package manager into a message board, gaining internet access and breaching Hugging Face systems — a warning shot, the company says.
Embedded
Zijing Semiconductor passes ten million automotive-grade chips shipped: a RISC-V automotive MCU completes the volume-production loop
Zijing Semiconductor says cumulative shipments of its automotive-grade chips have passed 10 million units; its M100 is described as China's first automotive-grade MCU designed on an open RISC-V core, reaching volume production in December 2025, while the company extends its RISC-V know-how into embodied AI.
TI expands Zephyr RTOS support across its full embedded portfolio
Coverage now spans MCUs, MPUs and wireless connectivity parts so Zephyr code, tooling and workflows carry over to TI silicon; TI contributes board support upstream and positions Zephyr for compliance work such as the EU Cyber Resilience Act.
Microchip completes its acquisition of Hailo, folding edge AI accelerators into an embedded giant
Microchip has completed its acquisition of Hailo, adding edge AI accelerators, vision processors and robotics processors to its embedded portfolio; Hailo's 100-plus customers and 10,000-strong developer community gain Microchip's global channel and support.
Ambarella launches X7, its first standalone AI accelerator, to add compute to installed Arm and x86 systems
Ambarella has introduced X7, its first dedicated CVflow AI accelerator chip, which attaches to Arm and x86 hosts over a single PCIe Gen3 lane or USB 3.2, draws 2 to 5 watts and targets cameras, gateways and industrial controllers already deployed in the field. It is sampling now.
Renesas launches the 64-bit RZ/G3L and RZ/G3SE MPUs: pin-compatible, with 1 mW-class deep standby
Renesas has expanded its RZ/G Series with the pin-compatible 64-bit MPUs RZ/G3L and RZ/G3SE: four Cortex-A55 cores plus a Cortex-M33, 1 mW-class deep standby that keeps Linux resident in memory, aimed at HMI and IoT edge gateways.
Espressif unveils the ESP32-S31: dual-core 320 MHz RISC-V with Wi-Fi 6
Espressif's most capable SoC yet: dual RISC-V cores, Wi-Fi 6, Bluetooth 5.4, 802.15.4 and a gigabit Ethernet MAC for display HMI, audio and edge-AI devices; the company says it has entered mass production.
STM32C5 announced: a Cortex-M33 entry line from $0.64
An entry-level family at 144 MHz with up to 1 MB flash and 256 KB RAM, focused on cost and efficiency; the STM32 line-up stays firmly on Arm.
ST's Stellar P3E: a first automotive MCU with built-in AI acceleration
Built for software-defined vehicles and X-in-1 ECU domains, its NPU targets real-time sensing and virtual-sensor workloads — AI settling into real-time control and functional-safety territory.
Espressif's ESP32-E22: a first tri-band Wi-Fi 6E connectivity co-processor
Tri-band 2.4/5/6 GHz on Espressif's in-house RISC-V architecture — a co-processor that upgrades existing host designs to Wi-Fi 6E while deepening the company's own IP stack.
RISC-V's 2025 annual report: MCUs, AI acceleration and automotive ramp lead
The annual report shows the ecosystem moving from evaluation projects toward automotive and industrial production, with a standardized MCU profile and AI acceleration flagged as growth areas; Embedded World 2026 material likewise stresses the shift from prototypes to volume.
Raspberry Pi AI HAT+ 2 ships: 40 TOPS for local generative AI
A Hailo-10H accelerator with 8 GB of onboard memory runs language and multimodal models locally alongside vision workloads, bringing entry-level edge generative AI into the Raspberry Pi ecosystem.
Chips & Compute
SK hynix answers with "memory-centric" at OCP 2026: CXL memory pooling, the MARS rack and CMM-Ax
At OCP Global Summit 2026 SK hynix showed its next-generation server designs: a second-generation CMM-DDR5 (256GB, more than twice the bandwidth), the memory-centric MARS rack system, and CMM-Ax, which computes inside memory.
TSMC posts a record $46.7B third quarter: 2nm contributes 8% of wafer revenue for the first time
TSMC's third-quarter 2026 revenue reached $46.72 billion and net income NT$763.06 billion, both records, with 2nm at 8% of wafer revenue and advanced nodes at 79%; fourth-quarter guidance is $52.3-54.3 billion.
ASML's third-quarter net sales hit a record €11.66 billion as AI demand turns capacity into the bottleneck
ASML reported record third-quarter net sales of €11.662 billion, a 57.5% gross margin and €4.96 billion of net income, and lifted its full-year outlook to €44-45 billion; it now says planned supply in 2027 and 2028 will not fully meet customer demand and is studying a 2028 capacity increase larger than the previously stated 30%.
Japan funds, GlobalFoundries operates: two new fabs in Malta, New York, for up to $21 billion
GlobalFoundries will operate two new fabs at its Malta, New York site under a U.S.-Japan initiative worth up to $21 billion, with the first 300mm fab due online in 2029; it also reallocated $955 million of CHIPS funding toward silicon photonics, SiGe and advanced packaging, while Commerce takes roughly 4% equity.
Meta brings 102.4T switching and Ethernet scale-up to OCP 2026: 72 accelerators per rack
Meta unveiled IcePack, its first liquid-cooled network rack, and 102.4 Tbps IceCube switch blades, rebuilt its GPU fabric with Flattened Planar Fabric, and showed the Ethernet scale-up switch Tahan plus a design with 72 accelerators per rack.
Iluvatar CoreX puts 144 GPUs into a single rack, shifting China's compute race to rack-scale systems
Iluvatar CoreX showed its first systematic supernode, the TianGai 300 IXL144, at CIIF 2026 in Shanghai on October 12, packing 144 GPUs and 18 CPUs into one rack, as IDC projects China's intelligent-compute capacity will reach 2,576.5 EFLOPS in 2026, up 87.9% year on year.
Samsung's Q3 operating profit hits 107.4 trillion won, the first 100-trillion quarter for a Korean company
Samsung's preliminary third-quarter guidance puts operating profit at about 107.4 trillion won (roughly $79.8 billion), up 782.5% year over year, on revenue of about 195 trillion won and a record 55.1% operating margin; memory drove the profit and HBM4 is expected to contribute meaningfully next year.
Huawei launches the Atlas 960E SuperPoD: the industry's first NPO-based supernode, 4,096 NPUs at 8 EFLOPS
At HUAWEI CONNECT 2026 Huawei launched the Atlas 960E SuperPoD, the industry's first supernode built on near-packaged optics: up to 4,096 NPUs per node, 8 EFLOPS at FP8, and 5,500 optical engines replacing the 48,000 800G modules it would otherwise need.
NVIDIA "Vera Rubin" makes its MLPerf debut: up to 3.7x the throughput of GB300
NVIDIA's Vera Rubin NVL72 made a first preview submission to MLPerf Inference v6.1, reaching up to 3.7x the throughput of GB300 NVL72 on Qwen3-VL and 2.5x on DeepSeek-R1, while a four-rack, 288-GPU GB300 system hit 99% scaling efficiency.
Autonomous Mobility
Waymo releases safety data across 270 million miles: an estimated 841 fewer injury crashes
Waymo's latest safety analysis spans over 270 million fully autonomous miles across five metro areas: 841 fewer injury-causing crashes regardless of fault, an 82% reduction, and 95% fewer serious-injury crashes.
Zoox's 100-vehicle Nevada cap expires: the robotaxi race in Las Vegas
Nevada's cap of 100 Zoox robotaxis for its first 12 months expires September 25; Waymo has opened in Las Vegas while Tesla and Uber hold permits, and Clark County can admit up to 7,000.
Waymo opens to the public in Las Vegas: Ojai and the 6th-generation Driver carry their first riders
On September 14, Waymo opened fully driverless service to the public in Las Vegas, with more than 100,000 people on its interest list; the city joins Denver and San Diego as the first to run majority-Ojai fleets powered by the 6th-generation Waymo Driver.
AutoFlight signs a 30-aircraft eVTOL order, China's first completed through financial leasing
On September 10, AutoFlight and Lingjie Aviation implemented a purchase agreement for 30 V2000CG CarryAll eVTOLs, with Wanjiang Financial Leasing providing the lease — China's first eVTOL deal completed through financial leasing.
MIIT releases its 15th Five-Year Plan and two mandatory standards: a timetable and a safety floor for automated driving
On September 11, China's MIIT released the 15th Five-Year Plan for intelligent connected new-energy vehicles — 70% and 40% NEV shares of new sales by 2030 — alongside two mandatory national standards that set a safety floor for automated driving and combined driving assistance.
Pony.ai and Verne start fully driverless Robotaxi rides with passengers in Zagreb — a European first
On September 10, Pony.ai and Verne began fully driverless Robotaxi test rides with passengers in Zagreb, Croatia, on a 22-kilometer route linking downtown with the airport — a European first for a fleet that has now logged more than 200,000 kilometers.
NHTSA investigates Tesla's Cybercab: how a car with no steering wheel self-certifies
On September 4, NHTSA opened an audit query into Tesla's Cybercab — a vehicle with no steering wheel or pedals that Tesla self-certified as meeting federal safety standards. The probe began hours after the first cars hit the road.
XPeng's second-generation VLA gets its first major upgrade: from 3D space to 4D space-time
On August 27, XPeng announced the first major upgrade to its second-generation VLA: Infini-VLA long-horizon memory, streaming inference, 3.5x more on-vehicle parameters and a new whole-vehicle brain called Master Agent, debuting with XOS 6.3.0 on the G9L.
Apollo Go goes live on Uber in Dubai: its first market with both self-operated and partner-based service
Starting August 20, riders in Dubai can hail a fully driverless Baidu Apollo Go vehicle straight from the Uber app, with service opening in parts of Umm Suqeim and Jumeirah — the first market to deliver on the two companies' July 2025 multi-year partnership.
Frontier Science
Google's "magic state cultivation": cutting the error of quantum computing's costliest step roughly 40-fold
Google Quantum AI reports in Science that magic state cultivation works on its Willow processor, cutting magic-state preparation errors roughly 40-fold and reaching 99.99% logical fidelity for T states.
China's first model-assisted breakthrough on a world-class maths problem: Intern-S2 and the three-dimensional Ghys conjecture
Shanghai AI Laboratory and a mathematician team tackled the three-dimensional Ghys conjecture using the Intern-S2 science model; the proof is on arXiv and all four main theorems passed Lean formal verification.
Tsinghua's Ding Sheng in Nature: a mouse embryo model that develops continuously from chemically induced totipotent stem cells
Ding Sheng's team at Tsinghua reports in Nature a mouse embryo model grown from chemically induced totipotent-like stem cells that passed through blastocyst, gastrulation and early organogenesis, with rhythmically contracting heart tissue and overall similarity to E8.5 natural embryos.
The 2026 Nobel Prize in Chemistry: Kagan and Soai solve the mystery of homochirality
The Royal Swedish Academy of Sciences awarded the 2026 Nobel Prize in Chemistry to Henri B. Kagan and Kenso Soai for non-linear effects and autocatalysis in asymmetric organic synthesis — Kagan's 1986 non-linear effect plus Soai's 2003 reaction that forms a single mirror image.
The 2026 Nobel Prize in Physics: turning Antarctic ice into a neutrino observatory
The Royal Swedish Academy of Sciences awarded the 2026 Nobel Prize in Physics to Francis Halzen for decisive contributions to the IceCube Neutrino Observatory and the discovery of high-energy neutrinos of astrophysical origin.
The 2026 Nobel Prize in Physiology or Medicine: a light switch for nerve cells
The Nobel Assembly awarded the 2026 Prize in Physiology or Medicine to Karl Deisseroth, Peter Hegemann and Georg Nagel: Hegemann and Nagel found a light-gated ion channel in a single-celled alga, and Deisseroth turned it into a light-controlled switch for nerve cells.
Stanford transplants human brain organoids into cortex-less mice: after three months, 90% of the cortex was human
Stanford Medicine reports in Nature that human cortical organoids transplanted into mice missing most of their cortex came to make up more than 90% of cortical volume three months later — and produced von Economo neurons previously seen only in postmortem brain tissue.
NIF's 11th fusion ignition: the 8.6-megajoule record still stands
Lawrence Livermore National Laboratory reports that the National Ignition Facility achieved fusion ignition for the 11th time on 20 June 2026 with a 7.9-megajoule yield; the single-shot record remains 8.6 megajoules and a target gain of 4.13, set on 7 April 2025.