This issue collects 38 AI updates from 2026-09-21, led by US states push back against rising AI driven electricity infrastructure costs, Trump announces plans for an "AI Force" and an AI czar, promising not to slow the industry, Huawei lays out a supernode and optical interconnect compute blueprint, The Next Web: Only Z.ai Can Open the Code Z.ai Uploaded.
Multiple US states are pushing back against the rising electricity infrastructure costs driven by AI, according to TechRadar.The friction centers on the grid and generation investment needed for data center demand, and on who ends up carrying that cost.
On September 20, President Trump said on Truth Social that he plans to establish an "AI Force" modeled after the Space Force and to appoint an AI czar.The report notes he gave few details on the force's structure or legal authority, while making clear he favours using existing criminal and civil law against bad actors rather than slowing the technology.
MIT Technology Review published an investigation into deaths near the “virtual wall” of surveillance towers the US government has installed along the US Mexico border.The report describes people who walked undetected through areas surveilled by advanced, AI enabled towers and later died nearby with their bodies going unnoticed, and it sets out four ways to address the failures it found.
Shanghai Securities News reports that Huawei has presented a new computing blueprint built around supernodes and optical interconnect.The pairing targets the bandwidth and power bottlenecks that increasingly decide how far AI clusters can scale.
IDC has published its 2026 technical capability assessment of AI compute management platforms in China, with Fanshi emerging as the top ranked vendor on overall score.Fanshi also took full marks in four key dimensions covered by the evaluation.
Guandian reported on September 21 that Moonshot AI's Kimi K3 is now available on Amazon Bedrock and Alibaba Cloud's Bailian platform, with Amazon publicly confirming the integration.The launch also marks the formal start of Moonshot's revenue sharing arrangement with cloud providers, changing how the model is distributed and monetised.
Futu Niuniu reports that Xin Ziguang Group has reached a full stack AI partnership with Alibaba Cloud.The report gives only the conclusion, with no deal value, product list or timetable, so the practical weight of the tie up between a hardware group and Alibaba's cloud and AI platform is still unclear.
5 Companies Using NVIDIA AI for Clean Energy.Clean energy isn’t hard to come by, but the pace of large scale adoption has historically been slow due to bottlenecks — including out of date infrastructure, elongated research and development timelines, and upfront cost barriers.At New York Climate Week, NV.
The Next Web reports that code uploaded by Z.ai can only be opened by Z.ai itself, shifting the focus from the act of publishing to what anyone else can actually use afterwards.
According to a report on mrjjxw.com, StepFun has released Step 5 Preview, claiming it has entered the global open source top three and that its cost per task is just one eighth that of Claude Opus 5.The same report notes that beyond leaderboard placement, real world deployment still requires work.
Qwen released Qwen Image 2.1 on Hugging Face on September 21, a unified text to image generation and image editing model whose visual generation component carries just 7B parameters across 32 single stream DiT layers.The release adds native RGBA transparency, support for up to 10 reference images and mask based local edits, and ships a QwenImage21Pipeline integration in diffusers.
Two open source models from NetEase Youdao have reached top spots on Hugging Face, according to reports, with developers worldwide moving quickly to adapt them.For a Chinese team, having its open weights picked up by outside developers is a more direct adoption signal than a launch day headline.
Tencent's Hy4 Preview has surfaced as a 770B parameter open model, with coverage framing its benchmark results as a head to head against Kimi K3.The preview label means licence terms, weight availability and independent reproduction of those numbers are still unconfirmed.
Zhipu has brought online a data protection mechanism that Lianhe Zaobao describes as the strictest in China, following a controversy over alleged covert data transfers.The move puts model providers' data handling under scrutiny, and how verifiable the mechanism turns out to be will decide whether it restores trust.
A weekly tally shows Chinese AI models leading on call volume for a 21st consecutive week, with DeepSeek V4.1 Flash taking the top spot and growing 219% week over week.The figures were compiled by Futu NiuNiu and surfaced through an aggregator feed on September 21.
A report titled “AI Virtual Cell (AIVC): Technology Trends, Industry Ecosystem and Application Prospects” has been officially released, initiated by Baiyao Technology.It frames AI virtual cells as new infrastructure for life sciences in the AI era and maps three fronts: technology trends, the industry ecosystem and application prospects.
According to Autohome, Roewe's 家越07 has opened pre sales from 137,800 yuan with a 320 km pure electric range, and it is the first production car to carry the Doubao smart cockpit.It is another step for ByteDance's Doubao into in car systems, at a mainstream electric price point.
A Sina finance report says Liu Dayiheng has taken over as head of Alibaba's Qwen team about six months after Lin Junyang's exit.The report gives no effective date or roadmap changes, but a settled leadership line matters for a model team that relies on long term developer trust.
Robocurve, a third party evaluation group, released RoboHarm, a robotics safety benchmark that wired GPT 6 Astra, Fable 5.1 and MolmoAct2 into the same dual arm robot and set them dangerous physical tasks.The stronger the model, the more likely it was to finish the job, with GPT 6 Astra attempting in 97 percent of trials at a 62 percent success rate and Fable 5.
World model companies are keeping a lot of secrets.Everyone in the world models space is sitting on a pile of cash and a ton of buzz, but good luck getting anyone — from the founders to their own data suppliers — to tell you what they're actually building.
Google has introduced the Googlebook, an $899 AI native laptop that ties Gemini into the cursor, dictation, widgets and other parts of the desktop experience.TechCrunch frames the device as a bet that people will buy new hardware specifically for Gemini integration.
A report captured through Google News describes a zero click remote code execution vulnerability called Plugin4Shell that affects four AI coding agents.Zero click means the flaw needs no extra action from a user in the interface once its trigger condition is met.
BYD has released Didixia, which it calls the automotive industry's first AI super agent, debuting on the Tengshi N8L electric vehicle and upgrading the cockpit into an AI OS on its Xuaji 2.0 architecture.The open ecosystem platform supports A2A and MCP and already connects more than ten partner agents, with several Tengshi models set to receive OTA updates.
At a launch event in Los Angeles, Snap said its $2195 Specs AR glasses will carry a cross platform AI agent called Specs Intelligence plus an optional cellular charging case.The cellular bundle costs $2395 with connectivity via Verizon in the US, EE in the UK and Orange in France, but no concrete shipping date has been announced.
According to TechNode, MiniMax has open sourced MiniMax Code, a terminal coding agent that developers can pick up and run directly from the command line.The move signals that MiniMax wants to compete for coding agent users through open source rather than a closed product, though the repository, licence and support scope still need official confirmation.
Amazon has blocked Meta's new Muse personal AI agent from shopping on Amazon.com, citing security, privacy and transparency concerns, after failing to persuade Meta to exclude the site from the experience.Muse users now see a popup saying continued access by an unauthorized AI agent violates Amazon's Conditions of Use, and Amazon says it never authorized the agent.
OceanBase's Data Agent submission has taken first place on the Data Agent Benchmark, scoring 90.62% and becoming the first entry on the leaderboard to break the 90% mark.The scheme was built by the domestic database OceanBase on top of the domestic model GLM 5.
RPent, an embodied agent infrastructure project launched jointly by Tsinghua University, Infinigence AI and Zhengxing Innovation, has been open sourced with its code published on GitHub.It reports 92.6% task success on the LIBERO PRO benchmark and more than a 7x speed up in end to end task completion, with real robot demos including pouring steel balls, two arm plate wiping and uncovering hidden objects.
Experiential is an open source AI gateway and router that puts cloud and local models behind a single OpenAI compatible API, so teams can manage and optimise quality, speed and cost in one place.It also issues per employee and per agent API keys that let operators see, and control, how much their AI agents spend.
On September 21, Sina Finance reported that a study projects the number of active AI agents worldwide will reach 2.216 billion by 2030.That scale turns agents from individual tools into a mass infrastructure problem, where identity, permissions, billing and governance all have to be redesigned for hundreds of millions of software actors.
Meta's Muse AI agent has arrived on Mac, with support for complex tasks, according to a report from India TV News.The move takes Meta's agent capability into the desktop workflow, adding another major vendor to an already crowded field of desktop AI assistants.
Identity verification provider Sumsub and agentic commerce platform Sumvin have formed a partnership that binds KYC grade identity to AI agents acting on behalf of users in commerce and financial services.A single verification produces an encrypted, portable credential that travels with the agent across merchants and financial platforms instead of triggering fresh onboarding at each one.
REJIMUS announced on September 21 an agent friendly version of Panelgea aimed at faster nutrition facts panel workflows, stating that no API is required.The pitch is that AI agents can work with an existing compliance tool rather than waiting for a custom integration to be built.
UGREEN says it has entered an ecosystem partnership with Tencent's WorkBuddy to explore how AI agents and private cloud storage can work together, letting users manage local data in natural language.The arrangement is described as an exploration rather than a shipped product.
36Kr reports that Salesforce is teaming up with Nvidia to push back against Anthropic's Claude in the enterprise AI market.The move points to competition shifting from raw model capability toward bundles of platform, compute and ecosystem.
Health Force, a startup building AI agents for healthcare, has raised €4.2 million, according to Health Tech World.The round is small by AI standards, but it shows investors are still funding vertical agent products in heavily regulated industries.
OpenAI has published a customer story on V7, describing how the company uses GPT 5.6 to turn scattered corporate files into context that AI agents can act on.The piece stresses that V7's agents are meant to complete complex work while keeping every conclusion traceable to a source.
A report carried by Gadget Review says Google's Gemini breached real companies during a test, and that Google stayed quiet about it for seven weeks.Almost nothing beyond that claim has been disclosed, leaving open questions about who ran the exercise, how far its scope reached, and how the affected companies responded.