|
|
|
|
Bottom Line Up Front China's internet regulator, the Cyberspace Administration of China (CAC), will begin taking public complaints about AI across 14 categories in June 2026 under its Qinglang cleanup campaign, split between service failures like unregistered models and unlabeled synthetic content, and abuses like impersonation, harm to minors, and AI-run water armies, per ITHome. Huawei open-sourced Pangu 2.0 at its 2026 Huawei Developer Conference (HDC) in Dongguan, releasing a 505 billion parameter Pro mixture-of-experts model (18 billion active) and a 92 billion Flash model (6 billion active), both with 512K-token contexts and code shipping from June 30. Huawei also unveiled HarmonyOS 7, which embeds Pangu 6.0 in the operating system kernel and upgrades its agent framework; the Mate 90 phones and Kirin 2026 chip will debut the system this fall, on top of 66 million HarmonyOS 6 devices and 1.3 billion ecosystem devices. |
| The Cyberspace Administration of China (CAC) opens public AI complaints intake covering 14 categories from model filing to AI managed water armies The Cyberspace Administration of China (CAC), Beijing's lead internet regulator, said its Reporting Center will open a dedicated "AI Application Disorder Reporting Zone" from June 2026 under the "Qinglang" (Clean and Bright) special action against AI abuses, ITHome reported relaying the CAC's official "Wangxin China" WeChat account. The 14 categories split into "AI app service violations" covering failures around large model filing, platform safety auditing, training corpus security, AI data poisoning, synthetic content labeling, illegal uses of AI and open source model security. The second bucket, "AI content abuses," covers AI "remixing" of classics, fake information, impersonation, violent or vulgar content, harm to minors and AI managed online "water armies" (paid fake-engagement networks). Securities Times relayed the same notice via the People's Daily financial wire. |
| Huawei open-sources 505-billion-parameter Pangu 2.0 with rolling code release from June 30 Huawei executive director Yu Chengdong unveiled openPangu 2.0 at the company's HDC 2026 developer conference in Dongguan, releasing a 505-billion total and 18-billion active-parameter "Pro" variant and a 92-billion total and 6-billion active "Flash" variant, both with 512K-token context windows, ITHome reported. The mixture-of-experts design activates only the smaller "active" parameter count per token, cutting inference compute versus a dense model of similar total size. Huawei claims single-card throughput on its own Ascend AI accelerators is about double mainstream open source baselines and will roll out seven open-source components including pretraining code, post-training code and training operators starting June 30. Yu said the total parameter count was capped at 505 billion because most of Huawei's Ascend compute is being allocated to support other Chinese AI firms, with the remainder used to focus on latency and throughput improvements. |
| HarmonyOS 7 embeds Pangu 6.0 in the kernel, debuting on Mate 90 this fall as Huawei's first agent native mobile OS Yu Chengdong released HarmonyOS 7 at HDC 2026, calling it the first Chinese mobile operating system with the Pangu 6.0 large model embedded directly in the system kernel rather than called as a cloud service, ITHome reported. The system uses an "Agent-affinity" architecture and an upgraded Harmony Agent Framework 2.0. Huawei claims a roughly 15% performance improvement over HarmonyOS 6 and a built-in assistant called Xiaoyi that already has 180 million daily active users. The Mate 90 series will be the first device to ship with HarmonyOS 7 official build in fall 2026 alongside Huawei's new Kirin 2026 chip, with the HarmonyOS 6 install base at 66 million devices and the broader Harmony ecosystem covering more than 1.3 billion devices per Yu. Developer Beta API 26 opens for sign-ups today on Mate 80 Pro, Pura 90 Pro Max, nova 15 Pro, Mate X7, Mate XTs and Pura X devices. |
| Shanghai Stock Exchange sets June 15 STAR Market hearing for AI accelerator firm Enflame Technology The Shanghai Stock Exchange will hold a listing-review meeting on June 15 for Enflame Technology, a domestic cloud AI compute chip designer founded eight years ago that built its software stack independently of NVIDIA's CUDA ecosystem, Leiphone reported. Enflame revenue grew from 301 million yuan ($42 million) in 2023 to 990 million yuan ($140 million) in 2025, while annual net losses narrowed from 1.665 billion yuan ($230 million) to 1.164 billion yuan ($160 million) over the same period. The company said it expects to reach group profitability in 2026 or 2027, with cumulative R&D spending of 3.676 billion yuan ($510 million) across 2023 through 2025. Enflame's GCU-CARA hardware architecture and "TopsRider" software platform support roughly 1,000 models across more than 300 deployed scenarios, with Tencent named as the lead anchor customer. |
| Sugon, a U.S. Entity Listed company, will unveil a "Hundred-Core Fusion" HPC platform on June 15 Sugon, a U.S.-sanctioned Chinese high-performance computing supplier, said it will launch a new general purpose computing platform on June 15 under the tagline "Hundred-Core Fusion, A New Realm of Supercomputing," signaling a high core count CPU design, ITHome reported. Sugon's largest shareholder is Beijing Zhongke Suanyuan Asset Management, an entity wholly owned by the Chinese Academy of Sciences Institute of Computing Technology, and Sugon itself holds 27.96% of Hygon Information as of March 2026. Hygon, which listed on the STAR Market in 2022, makes server CPUs under an x86 architecture license and accelerator chips compatible with a CUDA-like programming environment. The launch falls on the same day as Enflame Technology's Shanghai Stock Exchange listing review. |
|
| Huawei is to fully open-source release its NearLink protocol stack on July 15th, pitching it as a Bluetooth alternative Gong Ti, president of Huawei's Terminal BG Software Department, said the NearLink short-range wireless protocol stack will be fully open-sourced on July 15, 2026 with 150,000 lines of code and mature standard libraries, ITHome reported from the HDC 2026 keynote. NearLink is a Chinese-developed short-range wireless protocol pitched as an alternative to Bluetooth and Wi-Fi, claiming higher throughput, lower latency, longer range and stronger security and reliability, with some 5G techniques included in its design. The protocol is deployed today in Huawei phones, tablets, PCs, routers, headphones, styluses and keyboards as well as in third-party tablets, smart lights, shavers and air conditioners. |
| Huawei launches the Hongtu Plan to scale OpenHarmony across 200 chip types and 1,200 device categories Yu Chengdong launched the Hongtu Plan (literally "Grand Vision") at HDC 2026 to push Huawei's OpenHarmony open source operating system across more than 20 industries, 200 chip types and 1,200 device categories, ifeng reported relaying the keynote. OpenHarmony is the open source variant of Huawei's HarmonyOS that non-Huawei vendors can build on without the closed source layers tied to Huawei devices. Yu said OpenHarmony has accumulated more than 13,000 code contributors, more than 140 million lines of contributed code, more than 3,200 ecosystem partners and more than 1.3 billion ecosystem devices. |
| Mature process chip work gets equal billing with advanced process under 15th Five-Year Plan, Caixin opinion piece argues Beijing's 15th Five-Year Plan outline explicitly placed "refining and perfecting mature process nodes" alongside "raising advanced-process manufacturing capability" as parallel core strategies for the Chinese semiconductor industry, a Caixin opinion piece argues. Mature process nodes are chips at older geometries, roughly 28 nanometers and above, used in cars, appliances and industrial controls, distinct from leading edge 3-nanometer or 2-nanometer logic targeted by TSMC and Samsung. The argument pushes back on three views commonly held inside the industry: that mature processes are backward, that they are already in oversupply and that they no longer need fresh R&D investment. |
|
| III | AI & FOUNDATION MODELS | |
| DeepSeek and Moonshot are converging operationally despite divergent founder postures DeepSeek founder Liang Wenfeng and Moonshot AI founder Yang Zhilin are leading companies on increasingly similar paths after years as polar opposites in development strategy, per a Huxiu analysis. Both labs are raising outside capital, have started capital market tracks and have shifted technical focus toward long-context, agent and coding work. They are also scaling up their "harness" teams that handle inference efficiency and large cluster operations. DeepSeek's first external round will fund GPU and data center buildout, including a senior data center delivery hire, while Moonshot's capital is flowing into product, user acquisition and commercialization after OpenClaw, a third-party open-source agent framework that integrates Moonshot's Kimi models, drove sharp revenue growth early this year. Moonshot has closed three funding rounds over the past six months at rising valuations, while DeepSeek's previously self-funded position has shifted as it launched external fundraising this year. |
| Tencent's WeChat AI agent rollout marks its late-cycle AI bet after lagging behind Baidu, Alibaba and ByteDance competitors Tencent's WeChat AI agent transformation is the company's main strategic AI bet after lagging Baidu's Ernie, Alibaba's Tongyi Qwen and ByteDance's Doubao for most of the post-ChatGPT cycle, per TMTPost. The June 2 news of an upcoming WeChat AI agent sent Tencent shares up more than 10% intraday and added roughly HK$415 billion (about $53 billion) to market capitalization before most of the gain was reversed over the following sessions. As of March 2026, Tencent's standalone Yuanbao assistant had a monthly active user base of about 57.3 million, compared with about 345 million for ByteDance's Doubao and about 166 million for Alibaba's Qwen, per QuestMobile data. Tencent released its Hunyuan large language model in September 2023, almost 10 months after ChatGPT and roughly six months after Baidu's Ernie. |
| ACE Robotics' Kairos world model claims top rank on four embodied intelligence benchmarks ACE Robotics, also known as Daxiao Robotics, said its Kairos world model topped the RoboTwin 2.0, LIBERO-Plus, WorldModelBench Robot and DreamGen leaderboards for generative video and state prediction in embodied AI, beating NVIDIA's Cosmos 3, Physical Intelligence's PI, MotuBrain, being-H0.7, Abot, Fast-WAM and Wan2.2, Leiphone reported. A world model is an AI system that learns to simulate the physical environment and predict how it will respond to actions, used in robotics and autonomous driving to plan motions before executing them in the real world. Kairos uses a unified architecture that fuses understanding, generation and prediction at the base layer rather than bolting prediction on top of a video generation model, ACE says. The company puts Kairos at a 96.1% average success rate on RoboTwin 2.0, a bimanual robot manipulation benchmark; the leaderboard claims are self-reported and not yet independently corroborated. |
|
| TSMC plans 15% price hike on 3-nanometer process as AI demand outpaces capacity TSMC plans to raise prices for its 3-nanometer process by about 15% in the second half of 2026 in response to AI demand growth "far exceeding market expectations," Jiemian News reported citing industry sources. TSMC's 3-nanometer node is its most advanced process shipping in end products at scale, used in current generation NVIDIA AI accelerators, Apple chips and high-end smartphone system-on-chips, while its newer 2-nanometer node entered volume production in late 2025 and is still ramping. The Q2 monthly capacity for the line is expected to reach 160,000 to 175,000 wafers, but customer queues have not eased meaningfully. |
| Memory chip and lithium price surges add 15,000 to 20,000 yuan in cost per AITO vehicle, Seres chairman says Seres Group Chairman Zhang Xinghai said at the 2026 China Auto Chongqing Forum that memory chip prices have risen roughly fivefold, from about 20 yuan ($2.80) per unit to nearly 100 yuan ($14) per unit, ITHome reported. Zhang said rising input costs, led by memory chips and lithium carbonate, add an average of 15,000 to 20,000 yuan ($2,100 to $2,800) in cost per AITO brand vehicle, the EV line Seres makes in partnership with Huawei, even as average selling prices in the Chinese auto market continue to drop. Zhang cited lithium carbonate moving from about 80,000 yuan ($11,200) per ton a year ago to about 180,000 yuan ($25,200) per ton today. AITO models delivered 34,320 vehicles in May 2026, up 48.2% month over month. |
| Rare Metals ETF rises 4.15% as SK Hynix swaps molybdenum for tungsten in 375-layer NAND China Molybdenum, Zhuye Group and Jinduicheng Molybdenum hit their 10% daily limit, the CSI Rare Metals Theme Index closed up 3.26% and the Rare Metals ETF Huaxia (159053) closed up 4.15%, Jiemian News reported. The rally follows SK Hynix's completion of production verification for a 375-layer NAND flash memory that substitutes molybdenum for tungsten in word lines, with the South Korean memory maker planning to begin mass production at its Cheongju M15 fab by year-end. At equivalent feature sizes, molybdenum offers lower electrical resistance than tungsten, accelerating read-write speed in dense NAND stacks. Dongguan Securities flagged tungsten, molybdenum and germanium as the minor metals benefiting from combined new energy and AI demand, with the underlying CSI Rare Metals Theme Index's top 10 weights including China Molybdenum, Xiamen Tungsten, Yunnan Germanium and China Tungsten and Hightech. |
|
|
|