Tech weekly, Oct 12-18, 2026: Microsoft's Maia 300, an unpatched Siemens advisory, and humanoids on a parts line

Microsoft puts Maia 300 in front of the Open Compute crowd

Microsoft presented Azure Maia 300, its third-generation custom AI accelerator, at the OCP Global Summit in San Jose on October 12, and told Reuters it is planning multi-gigawatt deployments from next year that could involve more than a million chips. Each chip carries eight HBM4 stacks for 288 GB of memory and 18 TB/s of bandwidth inside a 950-watt budget, with compute chiplets on TSMC's N2P process. Racks hold 96 chips drawing about 167 kilowatts, against 72 GPUs and roughly 142 kilowatts in Nvidia's GB300 racks. Microsoft claims about 40% more useful work per megawatt than unnamed "leading GPU-based systems," and up to seven times the token throughput of a Maia 200 rack. Maia 200 already runs Copilot on Microsoft infrastructure, and GitHub Copilot moves onto Maia before year end.

Sources: Reuters via The Hindu BusinessLine, October 16, 2026; Dataconomy, October 13.

Meta charges for the model it calls frontier

Meta released Muse Wonder and Muse Wonder Full on October 12, the second generation from the lab run by Alexandr Wang. Mark Zuckerberg told the Financial Times that Meta has caught up with rivals at the frontier. Neither model ships open weights, ending the release habit Meta built around Llama. Wonder costs $4 per million input tokens and $20 per million output through the Meta Model API, with a one-million-token context window, while Full stays inside Muse Code subscriptions at $50 or $200 a month. On Artificial Analysis's Intelligence Index v4.3 at top reasoning settings, Wonder scored 52 and Full 56, against 48 for Muse Spark 1.3, which shipped six weeks earlier. Meta's long-duration software engineering test put Full at 57.5, ahead of Claude Opus 5.5 at 46.3 and GPT-6 Astra at 42.5, while Terminal-Bench 4.0 moved from 33% to roughly 57%.

Sources: Financial Times, reported by JW View (中新经纬), October 13, 2026; index figures via China AI News; TMTPost, October 13.

Gemini 4 Argon reaches general availability

Google opened Gemini 4 Argon to paid tiers on October 13 through the Gemini API, AI Studio and Gemini Enterprise, with Ultra subscribers getting it in Antigravity and Ultra and Pro subscribers getting it in the Gemini app and Search's AI Mode. The context window holds 1,048,576 tokens of combined input and output, and default synchronous output caps at 65,536 tokens unless long decoding is enabled. Introductory API pricing is $2 per million input and $10 per million output tokens, with cached input discounted 95%; both double to $4 and $20 after November 30. Google says the model ranks first on the Vals Index for finance, coding, legal and tax work. Internally, Antigravity-assisted code submissions rose over 35% with rollbacks down nearly a third. Argon also lifted Gemma 4 E2B throughput on vLLM from roughly 37,900 to more than 108,000 tokens per second.

Sources: Google AI for Developers model documentation; Google Blog, October 13, 2026.

Nvidia and Cisco draft a denser pluggable module

At the same summit on October 12, Nvidia and Cisco announced the Ultra-Dense Small Form Factor Pluggable specification. USFP runs 400 Gb/s electrical lanes, twice what today's pluggables carry, and the first planned transceivers fit 16 full-duplex lanes into an OSFP-sized package for up to 12.8 Tb/s, with room to go past 32 lanes later. Against current 200 Gb/s pluggables the companies claim 4x front-panel bandwidth density and about 1.3x better power efficiency, letting operators attach more compute without another switching tier. The connector seats the module before its high-speed contacts engage, a two-step insertion intended to reduce contact wear.

Sources: Nvidia announcement, reported by The Paper (澎湃新闻), October 12, 2026; C114, October 13.

Siemens leaves eight S7-1500 flaws without a patch

Siemens published 17 industrial control advisories on October 14, spanning SIMATIC S7-1500 CPUs, Industrial Edge devices, RUGGEDCOM controllers, SIMATIC IPCs, WinCC and others. The highest score, CVSS v3.1 9.8, belongs to SQLite flaws in the RUGGEDCOM CROSSBOW Station Access Controller below V6.0. More pressing for plant operators is SSA-591731 from October 13: eight cross-site scripting flaws in the web server of S7-1500 CPUs and related families including ET 200SP and ET 200pro, three of them rated critical at CVSS v4.0 9.4, with no fix available at publication. The bugs sit in the Trace page, which fails to sanitize comment fields in imported trace files. An attacker would have to get a user holding the Read diagnostics right to open a rigged file, after which injected JavaScript can act through the web server within that user's authorization. Siemens advises disabling unused web servers, restricting TCP ports 80 and 443 to trusted addresses, and importing only trusted trace files.

Rockwell Automation patched two flaws in the 1732E ArmorBlock, one of them critical, where an unauthenticated attacker could change the web server password and take the administrator account. Schneider Electric issued advisories for Pro-face STW6000 and Harmony STW6 operator panels and Sage RTUs. On October 16, VulnCheck separately disclosed an out-of-bounds write in Snap7 through version 1.4.3: opReadArea() trusts a DataLength field from the PLC and copies up to 65,535 bytes into a caller's buffer, which can crash SCADA, HMI or gateway processes.

Sources: securityonline.info, October 13, 2026; VulDB entries CVE-2026-63353 and CVE-2026-111280; SecNews, October 14.

KUKA brings a plant platform and a wheeled humanoid to Shanghai

At the China International Industry Fair on October 12, KUKA introduced KaiTian KAI, an industrial intelligence platform that gathers equipment data across a plant, tests decisions in a digital twin before passing them to machines, and uses multi-agent units to execute them. Economic Observer reported that KUKA describes the move as going from automation supplier to plant-wide intelligence services, and that the developer community opened alongside the platform reports 85,000 registered users across more than 20,000 companies. MIRO, the new industrial wheeled humanoid, stands 1.39 to 1.85 meters tall with 21 degrees of freedom, carries 20 kg, and runs at least six hours per charge, with KUKA citing an equivalent 15,000-hour life test. The same stand brought the LBR NEXUS cobot at 7 to 20 kg payloads, a welding cell with repeatability of 0.04 mm, and from Midea's Welling unit a force-controlled joint module that resolves torque changes of 0.02 N·m.

Sources: Economic Observer (经济观察网), October 15, 2026; Guangzhou Daily (广州日报), October 13.

Fifteen humanoids take loading stations at a parts maker

Leju Robotics and Zhaofeng Electromechanical announced an embodied-intelligence loading deployment on October 15 at Zhaofeng's Hangzhou plant. Fifteen humanoid robots cover 15 stations and more than ten part types, completing each loading cycle in under 40 seconds at roughly 98% success. The line has run for over two months and close to 1,000 hours, handling more than 60,000 parts. Jiemian News reported that a skilled worker finishes the same transfer in about 15 seconds against the robots' roughly 35, and quoted Zhaofeng chairman Kong Aixiang estimating another six to twelve months before cycle times support wider rollout. IDC data released September 29 put global humanoid shipments for the first half of 2026 near 25,000 units, up 432.1% year on year with China above 19,000, while research, education, exhibition and government data-collection buyers accounted for 69% of them.

Sources: Jiemian News (界面新闻), October 16, 2026; Securities Times (证券时报), October 15.