---
date: 2026-06-10
subject: "MIIT AI+telecom plan to 2028 | MusaCoder trains on MTT S5000 | He Xiaopeng heads XPeng robotics"
---

**China's industry ministry (MIIT) issued a 2026-2028 plan** to fuse AI with the country's telecom networks, requiring by 2028 that 75% of metro areas sit close enough to computing power for near-instant AI responses, that 30+ proven use cases emerge and that networks largely configure and repair themselves. **GPU maker Moore Threads open-sourced MusaCoder**, a coding model trained entirely on its own Chinese-made chips, and says the 27 billion parameter version beat Anthropic's Claude Opus 4.7, DeepSeek-V4 Pro, GLM-5.1 and Kimi K2.6 on a key benchmark for GPU code. **XPeng founder He Xiaopeng took direct command** of the company's humanoid robotics business June 10, comparing the moment to the launch of its first car eight years ago as it races toward year-end mass production of the IRON robot. **ByteDance is spinning off its AI drug discovery unit** to raise outside money while keeping control, moving the roughly 50-person team, its algorithms and drug pipeline into a new company in what insiders call the firm's first attempt to commercialize AI for science.

# 1. Top Stories

- **MIIT binds AI telecom integration to 2028 milestones, with 75% metro 1-millisecond latency coverage target** — The Ministry of Industry and Information Technology (MIIT), China's lead regulator for telecom and industrial information policy, issued the "AI + Information and Communications Innovation and Development Implementation Opinions (2026-2028)" this week, CLS reported June 10. By 2028 the plan requires metro compute 1-millisecond latency circle coverage of at least 75%, more than 30 high-value typical scenarios and information communications networks reaching initial high-grade autonomy at internationally advanced operations and service levels. By 2030 the plan targets significant breakthroughs in core technologies for AI telecom network integration and improved integrated sensing-computing-intelligence service capability. The document organizes work into AI driven network architecture research, 5G-A and 6G integration, network-native intelligence, space-based compute networks and a network-agent internet, alongside high-end optoelectronic chip R&D and intelligent compute super-node interconnect. [CLS](https://www.cls.cn/detail/2395872)

- **Moore Threads open-sources MusaCoder code model post-trained end-to-end on domestic MTT S5000 GPUs** — Moore Threads released and open-sourced MusaCoder, a code generation model for GPU low-level kernel code, in 9 billion and 27 billion parameter sizes, ITHome reported June 10. The complete post-training pipeline ran on the Kuae intelligent compute cluster built on Moore Threads' MTT S5000 GPUs, which the company describes as the industry's first open-source code language model completed end-to-end on a domestic full-function GPU stack. Moore Threads reports MusaCoder-27B-RL hit KernelBench Overall Pass@8 of 93.2% and Avg.@8 of 88.60%, ahead of Claude Opus 4.7, DeepSeek-V4 Pro, GLM-5.1 and Kimi K2.6. Model weights are posted on Hugging Face and the technical paper is on arXiv as paper 2606.04847. [ITHome](https://www.ithome.com/0/962/509.htm)

- **XPeng founder He Xiaopeng takes direct CEO role at robotics business ahead of year-end IRON mass production** — XPeng Chairman and CEO He Xiaopeng announced via a company-wide letter on June 10 that he will personally serve as CEO of the humanoid robotics business effective immediately, Wallstreetcn reported. He compared the current IRON humanoid commercialization juncture to the eve of XPeng's first mass-produced vehicle G3 launch eight years ago, framing the move as a step in repositioning XPeng from a smart vehicle maker to a physical AI company. The letter said the robotics unit has integrated XPeng's hardware, AI model, supply chain and precision manufacturing groups, with He's direct oversight intended to mobilize cross-department resources through the mass-production push. He named four product targets: a humanlike form factor; on-device natural language and reasoning rather than cloud reliance; full-body coverage with light motion and SEP data privacy; and first-to-scale high-end humanoid production through full-stack in-house R&D and cross-domain integration. [Wallstreetcn](https://wallstreetcn.com/articles/3774318)

- **ByteDance carves out AI drug discovery unit for independent financing, with parent retaining controlling stake** — ByteDance has launched the spin-off and independent financing process for its AI drug discovery line, with the parent retaining a controlling stake post-spin, ITHome reported June 10 relaying Intelligent Emergence. The roughly 50-person core team, algorithms, technology platform and existing pipeline assets transfer to the new entity, which will continue to receive compute support from Volcano Engine. The new company is led by Liu Kai, who has run ByteDance's AI pharma effort since 2021, and the internal protein-structure-prediction team has been merged into Liu's group. A person close to the spin-off described the move as ByteDance's first attempt at AI4S industrialization and said the company hopes to "run through a Chinese path for AI4S industry." [ITHome](https://www.ithome.com/0/962/437.htm) [36Kr](https://eu.36kr.com/en/p/3846956646124036)

- **China's Foreign Ministry warns EU on Russia sanctions package potentially targeting Chinese firms** — Foreign Ministry spokesperson Lin Jian said at the June 10 regular press conference that Beijing firmly opposes illegal unilateral sanctions lacking UN Security Council authorization, Wallstreetcn reported relaying Beijing Daily. Lin was asked about EU High Representative Kaja Kallas's comments that the next Russia sanctions package may include companies based in China, Turkey and India. Lin said China has made repeated solemn representations to the EU urging it to correct its erroneous approach and withdraw illegal unilateral sanctions. Lin said China "will closely monitor relevant developments and will take necessary measures to firmly safeguard its legitimate rights and interests." [Wallstreetcn](https://wallstreetcn.com/livenews/3117298)

# 2. Policy & Regulation

- **MIIT plan accelerates 400G/800G backbone build and east-central-west compute hub channel optimization** — The MIIT 2026-2028 AI + Information and Communications plan calls for accelerating 400Gbps and 800Gbps backbone transmission network construction and optimizing transmission channels between eastern, central and western national hub nodes, National Business Daily reported June 10. The plan orders a faster buildout of China's highest-capacity internet backbone lines and better data routes between the country's eastern, central and western computing hubs, National Business Daily reported June 10. It calls for upgrading city networks to all-optical high-speed systems, improving the exchange points where networks hand off traffic, and flattening network layers so city users can reach computing power within milliseconds. [National Business Daily](https://www.nbd.com.cn/articles/2026-06-10/4422213.html)

- **MIIT directs R&D on high-end optoelectronic chips, co-packaged optics and AI compute super-node interconnect** — China's industry ministry is ordering stepped-up research on high-end chips that move data using light, including optical switching components and co-packaged optics, along with trials of hybrid optical-electronic networking, ITHome reported June 10. The 2026-2028 plan also calls for optical interconnect research for large AI computing clusters, plus advanced networking work (lossless wide-area networks, task-based scheduling and AI agents for network operations) aimed at shuttling data between computing centers faster and more cheaply. [ITHome](https://www.ithome.com/0/962/498.htm)

- **MIIT plan slates AI smartphones, PCs, smart home and wearables for "vigorous development" with embodied intelligence integration** — China's industry ministry wants "vigorous development" of AI smartphones, PCs, smart home devices and wearables under its 2026-2028 plan, ITHome reported June 10. The plan also calls for ensuring robots and other physical AI systems work smoothly with networked communication hardware, part of building an integrated ecosystem of AI devices. It encourages telecom carriers to develop AI agent services for consumers, including personal assistants, home management, caregiving, interactive fitness and 3D viewing. [ITHome](https://www.ithome.com/0/962/511.htm)

# 3. Chips & Semiconductors

- **TMTPost documents domestic compute usability and procurement gaps, with next-generation chips framed as inflection point** — Domestic compute remains both hard to use and hard to procure, with industry participants pointing to the next generation of domestic chips entering mass production as the key transition window, TMTPost reported June 10. Inference vendor Shishi Technology relies on NVIDIA for more than 70% of compute, and Chinese banks reportedly continue routing AI workloads through third-party compute services that themselves use nondomestic GPUs, satisfying "first-layer" localization rules without fully shifting silicon. Kylin Software vice president Yao Ling said no unified compatibility standard exists across domestic chip vendors, leaving single-chip adaptation cycles at three to six months. Huawei plans roughly 600,000 Ascend 910C and 750,000 Ascend 950PR units in 2026 for total production of about 1.6 million chips, still below market demand against Huawei's own estimate of $3 trillion to $4 trillion in AI infrastructure spending by 2030. [TMTPost](https://www.tmtpost.com/8022414.html)

- **Shin-Etsu, SUMCO and GlobalWafers raise 12-inch wafer prices on AI demand, Chinese makers gain mid-node share** — Global silicon wafer suppliers Shin-Etsu Chemical, SUMCO and GlobalWafers simultaneously raised 12-inch wafer prices in 2026, with AI and high-performance computing specialty wafers seeing the largest increases, Wallstreetcn reported June 10 relaying Guosheng Securities analyst Hua Xiaowei. SUMCO projects AI demand for advanced-process wafers will reach 1 million wafers per month in 2026, more than 10% of global 12-inch demand, and SEMI projects the global wafer market will exceed $20 billion by 2030. Chinese domestic fabs will account for 47 of an estimated 108 new global fabs by 2028 per SEMI, with the Chinese share of 22 to 40 nanometer capacity projected to climb from 25% in 2024 to 42% in 2028. Domestic 12-inch wafer maker Xi'an ESWIN hit 850,000 wafers per month in monthly capacity at end-2025 for 6.8% global share, with Q1 2026 12-inch revenue at silicon wafer specialist Lion Microelectronics up 88.12% year over year. [Wallstreetcn](https://wallstreetcn.com/articles/3774317)

- **TSMC CFO confirms price increase posture but rejects "four or five times" surge framing** — TSMC Chief Financial Officer Wendell Huang told media at the Hsinchu Science Park on June 9 that inflation has pushed up the foundry's operating costs and price adjustments are not ruled out, but explicitly denied any "four or five times" surge, Wallstreetcn reported. Huang said pricing will reflect TSMC's "technology leadership" and "manufacturing excellence" rather than aggressive jumps. The remarks follow TSMC Chairman and CEO C.C. Wei's prior shareholder meeting statement that he "hopes" to raise prices because competitors have already moved. [Wallstreetcn](https://wallstreetcn.com/articles/3774309)

# 4. Digital Infrastructure

- **China Merchants Bank documents DeepSeek-V4 Flash deployment on domestic AI chips with SGLang RBG** — China Merchants Bank's IT department architecture team documented a production deployment of DeepSeek-V4 Flash large expert-parallel inference services on domestic AI chips using the SGLang RBG component, InfoQ China reported June 10. The deployment uses a Prefill-Decode disaggregated architecture orchestrated through Kubernetes, with multi-role topology spanning Router, Prefill and Decode instances and Worker pods spread across 16 AI chips per Prefill instance. Engineering work addressed dynamic port allocation under hostNetwork mode, service discovery timing under collective communication libraries, multi-level fault self-healing for KV cache RDMA transfers and in-place upgrade flows, in what the team describes as the industry's first such deployment. [InfoQ China](https://www.infoq.cn/article/FDIT4N6S583uNKGmUm8F)

- **Ant Group OS team names "Datacenter Use" as next agent computing approach beyond multi-agent collaboration** — Ant Group engineers Tan Jianfeng and Zhou Tianyu told the QCon Beijing conference in April that the next step in agent computing after "Multi-Agent" is "Datacenter Use," in which AI agents directly command data center computing power rather than coordinate with other agents, per a talk transcript InfoQ China published June 10. Ant's AKernel system is its bet on the approach. The company says fewer than three people handle AKernel's development and operations, versus the 30 to 50 staff a conventional infrastructure stack typically requires, with AI-assisted workflows letting a single engineer cover everything from design through testing and operations. [InfoQ China](https://www.infoq.cn/article/iSlYBH5XfQ6RCJrbaqOU)

- **Huawei HDC 2026 agenda lists June 13 JetBrains IDE session on HarmonyOS PC** — The agenda for Huawei's HDC 2026 developer conference includes a June 13 session titled "JetBrains IDEs on HarmonyOS," presented by the head of JetBrains' IntelliJ division, ITHome reported June 10. JetBrains' widely used programming tools run on Windows, macOS and Linux but not on Huawei's HarmonyOS PCs, and ITHome said the session suggests that may be about to change. The conference runs June 12 to 14 in Dongguan, with the keynote June 12 at 2:30 p.m. [ITHome](https://www.ithome.com/0/962/408.htm)

# 5. Talent & Capital

- **Alibaba in talks with Chinese nuclear SOE on small reactors for Hangzhou Renhe data center** — An insider at a state-owned nuclear company confirmed to China Business Journal that Alibaba approached it about building small nuclear reactors to power its Hangzhou Renhe data center, Sina Finance reported June 9. The site is Zhejiang's first cloud computing data center and serves Chinese AI firms including BrainCo and LingBan Technology. The sticking point is price: at China's grid baseline of about 0.58 yuan ($0.08) per kilowatt-hour, the source said, direct nuclear supply loses its cost advantage. A May 2026 action plan from four Chinese agencies proposed for the first time letting nuclear and hydrogen power feed computing facilities directly, and in January an Alibaba subsidiary joined China Nuclear Power in a venture with 250 million yuan ($35 million) in registered capital. [Sina Finance](https://finance.sina.com.cn/tech/roll/2026-06-09/doc-iniauzkt5102403.shtml)

- **Zhongke Wenge, a CAS Institute of Automation spinout, passes HKEX hearing for governance LLM listing** — Beijing Zhongke Wenge Technology, a Chinese Academy of Sciences Institute of Automation spinout focused on LLM driven decision intelligence for public governance, media and commercial enterprise customers, passed the Hong Kong Stock Exchange main board listing hearing with CICC as sole sponsor on June 9, Wallstreetcn reported. The company holds 10.2% share of the domestic enterprise-level LLM driven decision intelligence market by 2025 revenue per CIC Consulting, in a segment with total 2025 revenue of approximately 3.9 billion yuan ($550 million). Revenue rose from 249.7 million yuan ($35 million) in 2023 to 405.3 million yuan ($57 million) in 2025, with gross margin stabilizing around 51.2% and operating cash flow at negative 188 million yuan ($26 million) at the end of 2025. [Wallstreetcn](https://wallstreetcn.com/articles/3774303)