CHINA AI DATA / INDUSTRY INTELLIGENCE
The latest developments, connected coverage and a daily reading list.
Tipster SmartPikachu said the ROBOT PHONE would begin mass production in the first half of 2026. Honor said it would reveal more details at MWC 2026; the device features a movable camera arm that folds back into place.
ByteDance's Doubao, Alibaba's Qwen and DeepSeek rank first, second and third, respectively, based on user scale, API usage and open-source ecosystems. The report forecasts China's AI foundation model market will exceed RMB 70 billion in 2026 as competition shifts toward efficiency and ROI.
The MI325X combines CDNA 3 compute with 256GB of HBM3E memory. Reserved capacity offers rates well below on-demand pricing and guarantees scheduled availability for memory-bound inference and very-large-context LLM workloads.
As of the 2026-07-28 07:00 update, the OCPI H100 rental rate was $2.60/hr, up +39.0% over 3 months but falling on the week.
GPU compute became cheaper in 2025, but low prices can lead to higher overall costs when performance is unstable; effective cost depends on stable throughput and effective training steps. The article expects further market segmentation, with more complex supply, delivery increasingly priced into services and compliance becoming a firm requirement.
The measures encourage data centers to connect to Fujian’s integrated public computing resource service platform under unified standards and promote interconnection with telecommunications operators’ computing resource platforms. By 2027, the platform aims to serve over 1000 enterprises.
Liquid-based cooling is expected to be the fastest-growing cooling technique in the coming years. Hyperscale data centers held the largest market share by data center type in 2025.
In 2025, liquid cooling as a service (LCaaS) began deployment at scale, allowing customers to pay monthly per kilowatt of cooling capacity without purchasing liquid-cooling equipment. Adoption among small and medium-sized data center users reached 19%.
In January 2026, Alibaba formed a 250 million yuan ($35.9 million) joint venture with state-owned China National Nuclear Power Co. to secure a stable and reliable electricity supply for its expanding AI data centers.
Data centers are complex facilities for centralized data processing, storage, transmission, exchange and management. Favorable developments in AI and cloud computing are driving growth in data center market demand.
Released in December 2025 by YHResearch, the report analyzes the global and Chinese data center power equipment markets, covering market size, growth trends, major manufacturers, geographic distribution, product types and applications.
The report lists the data center power market size for 2026 at 2.56 billion US dollars and forecasts 7.17 billion US dollars for 2034, with a compound annual growth rate of 15.6% for 2026-2034.
Digitalization and EcoStruxure solutions helped EcoDataCenter become one of the most sustainable data centers in the Nordics and worldwide.
The 1700V PowiGaN switch integrated into the InnoMux™2-EP IC supports input voltages up to 1000VDC. Its synchronous rectification zero-voltage switching (SR ZVS) operating mode delivers over 90.3% efficiency for 12V systems in a liquid-cooled, fanless 800VDC architecture.
Neither company is currently prohibited from selling optical transceivers in the United States, and the FCC has not announced a vote, published a final rule or set a formal implementation date. Reuters reported that officials hoped to complete the proposed restriction before the end of 2026.
Optical module makers are raising funds for upstream chips, system solutions and global production capacity rather than accepting a component-supplier valuation. Whether the financing gives them the power to set industry rules depends on genuine chip-level breakthroughs, not merely packaging adaptations.
The report values the global AI server market at USD 194.62 billion in 2025 and projects growth from USD 262.22 billion in 2026 to USD 2847.32 billion by 2034, with a forecast CAGR of 34.73%. North America held a 38% revenue share in 2025.
The report describes over RMB 300 billion ($42 billion) invested in China's AI startups across 113 deals during H1 2026. It highlights regional funding clusters, technical priorities and evolving business models.
According to MIR Rui Gongye, Langchao Xinxi held approximately 32.0% of China's AI server market in the first half of 2025, ranking first. Industry concentration was high, with CR2 at 47% and CR5 at 81.8%, as market resources continued to concentrate among leading manufacturers.
Media reports say Zhongguo Jushi raised electronic fabric prices for month 9 by 15% for thick fabric and 20% for thin fabric amid growing AI server demand. Guoji Fucai announced plans to raise no more than 5 billion yuan for projects including high-frequency, high-speed electronic fiber fabric for high-performance substrates.
Citing IDC, the report says China's intelligent computing capacity (FP16) reached 259.9 EFLOPS in 2022 and was forecast to rise to 1117.4 EFLOPS in 2027, with a CAGR of 34%. It also identifies higher processor performance, memory capacity and bandwidth as hardware requirements for AI phones.
The report forecasts 2026 growth of 44.6% for CSP-developed ASIC AI servers, versus 16.1% for GPU servers. It expects ASIC servers' share of AI server shipments to rise from 27.8% in 2026 to nearly 40% in 2030.
Zhongshang Industry Research Institute estimates China's AI server market at approximately RMB 160.7 billion in 2024 and RMB 250 billion in 2025, forecasting RMB 350 billion in 2026. The article links growing demand to accelerating commercialization of generative AI and large models.
The article says domestic AI chip substitution exceeded 40% in 2025, while Nvidia's share of China's cloud AI chip market fell from 95% to 55%. The report highlights uneven infrastructure use: average GPU utilization at operating commercial AI computing centers is only 30%, versus 75%-98% in leading cloud providers' dedicated clusters.
The article identifies thermal management upgrades as a key 2026 AI chip supply-chain trend, saying diamond cooling is entering major-company validation and initial mass production, with diamond-copper composites advancing fastest. It also says HBM supply expansion is constrained by a 12-18 month cycle, requiring customers to lock in capacity through long-term agreements and substantial prepayments.
According to Gaogu Capital, Jiuwanli Future Technology, founded by former HiSilicon Kunpeng general manager and Horizon Robotics chip president Chen Peng, completed seed and angel rounds totaling over US$100 million, approximately RMB 671 million. The company plans to develop high-compute edge and device inference chips capable of efficiently running very large models on a single chip.
The article compares major Chinese AI models as of August 2026, including their developers and capabilities. It describes a field that expanded from a single viral release into numerous credible contenders in about eighteen months, with frequent new versions making coverage quickly outdated.
The article says five firms delivered six near-frontier releases in roughly two months, spanning different architectures and modalities. These include Kimi K3, enterprise-focused Qwen3.8-Max, DeepSeek V4-Flash and V4 Pro, GLM-5.2, and ByteDance's video model Seedance.
The report forecasts Asia-Pacific's GPU rental platform market to post the highest CAGR, exceeding 20% from 2026 to 2034.
The report covers GPU computing rental suppliers, time-based and monthly or annual billing models, and applications for medium-to-large and small companies. Its outline includes market size and share by product type for 2020-2025 and forecasts for 2026-2031.
The article describes Saiyi Information's move into computing rental amid rising demand. It reports that the company's revenue decline narrowed and net profit attributable to shareholders rose in the first quarter of 2026, while its internally developed Shanmou GPT supports AI applications including after-sales service and recruitment.
The office is a provincial department-level administrative agency dispatched by the National Energy Administration to Yunnan. Its responsibilities include electricity market oversight, access to power grids and oil and gas pipelines, administrative enforcement, electricity safety and licensing.
The page lists GB/T 31168-2023 on cloud service security capability requirements and GB/T 31167-2023 on cloud service security guidance. It also lists GB/T 34942-2025 on security capability assessment methods and GB/T 37972-2019 on the framework for cloud service operational supervision.
The entry describes China Telecom's cloud ecosystem cooperation plan, which recruits partners in technology, applications, services and channels. It also reports that Tianyi Cloud issued a security advisory after detecting increased attacks on customer cloud hosts, websites and cloud platform services.
The article argues that cloud services delivered over public networks are likely to be classified as value-added telecommunications services. It states that foreign providers must establish a commercial presence in China to offer committed telecommunications services and cannot supply telecommunications services to China on a cross-border basis.
The report states that China's data center liquid cooling market reached 98.6 billion yuan in 2025, supported by AI computing infrastructure deployment and stricter PUE controls. It describes cold-plate cooling as the mainstream approach and immersion cooling as gaining penetration in ultra-high-density computing applications.
AI development and computing demand are driving demand for liquid-cooled servers. Cold-plate cooling accounts for approximately 65% of the market, immersion cooling approximately 34%, and spray cooling approximately 1%.
Shenzhen Xiongtao identifies UPS power supplies for computing centers and data centers as its core development strategy. In the first half of 2025, this business accounted for 47% of company revenue, more than 80% of R&D spending, and 75% of profit.
Trend Force describes 2025 as the first year of AIDC storage demand and 2026 as a major delivery year. It projects that 13GW of new US data centers will drive 10.7-25GWh of storage demand, assuming 20%-50% storage capacity and 4-hour duration, with duration extending toward 6-8 hours.
The source describes a duopoly trend led by Innolight and Eoptolink. In the third quarter of 2025, shipments of 400G and higher-speed data center optical modules exceeded 10 million units for the first time, generating over $5 billion in revenue.
Huawei claims it has the industry’s first 7.2-terabit-per-second NPO module, as companies compete to establish next-generation AI infrastructure standards.
The classification lists SFP, XFP, GBIC and X2 optical modules, alongside electrical-interface variants. XFP optical modules are listed at 10Gbit/s, while GBIC optical modules are listed at 1.25Gbit/s.
The source states that some Chinese industry standards for 200G/400G wavelength-division multiplexing and 25Gb/s preceded international standards. IEEE and OIF are advancing standardization for optical modules operating at 800G and higher rates.
The source gives conflicting market estimates for 2025: USD 4861.5 million in its FAQ and $174.6 billion in its base-year data. Its analysis forecasts a CAGR of 38.632% during 2026 to 2033, while the FAQ separately projects 35.20%.
The report says large models and generative AI applications are increasing demand for high-performance computing resources and expanding the AI server market. Citing IDC, it forecasts that inference workloads will account for 73% by 2028.
The article characterizes the US approach as prioritizing peak hardware performance and China's approach as prioritizing networked system efficiency through optical communications and national scheduling platforms. It expects domestically produced computing capacity to exceed 65% of government and central state-owned enterprise procurement by the end of 2026.
The article says AI demand and capacity shortages drove up memory and storage prices, doubling server costs and constraining computing expansion. It describes an industry consensus for 2026 that inference will dominate computing consumption, agents will become mainstream, and competition will shift toward efficient, stable, low-cost model services and AI applications.
According to Huakun Zhenyu's official account, an IDC report shows that China's non-IA chip server shipments grew 24.3% in 2025, versus 4.4% for the overall market, with the company leading at 13.7%. General-purpose servers accounted for over 80% of units, and edge inference demand was a core growth driver.
The note examines China's export-control framework and its reach beyond the border, tracing policy from the 2020 Export Control Law through rare-earth measures in 2025 and adjustments in late 2025 and early 2026. It also reviews effects on downstream manufacturers, trade flows and efforts to diversify mineral supplies away from China.
The article states that the advanced-computing temporary general license expires on December 31, 2025 and does not override Entity List or Military End User List restrictions. Under the NAC exception, exports or reexports to D:5 destinations or Macao require notification to BIS at least 25 calendar days in advance; after 25 days, BIS indicates whether the exception may be used or a license is required.
The article describes rebutting a presumption through approved or authorized IC designers' datasheets or performance declarations showing that an IC falls outside 3A090.a. It states that after April 13, 2026, companies meeting the specified criteria and applying for approved IC designer status are treated as authorized for 180 days.
The report estimates that China's humanoid robot shipments grew 614.29% in 2025 and represented 80%-90% of global shipments, with education and research the largest application segment. Yushu Keji and Zhiyuan Jiqiren together held nearly 50% of the market, while the industry remained at an early stage with technology paths and business models still being explored.
The cited data show that China's industrial robot exports grew 18.6% in the first half of 2026, reaching 141 countries and regions. Surgical robot export markets expanded from 23 in the same period a year earlier to 49.
Yunshenchu has developed a solution that lets quadruped robots transfer instrument-reading capabilities between factory sites, shortening commissioning. Liu Chen said real-world operation requires coordinated task understanding, navigation, manipulation and safety decisions, rather than simply adding a large model.
An International Federation of Robotics report released on September 24 said China's new industrial robot installations reached a record in 2025, accounting for about 59% of the global total.
A report says the Yangtze River Delta accounts for more than 50% of China's embodied-intelligence companies and financing, supporting the shift toward factory deployment. Yushu Keji founder Wang Xingxing cautioned that humanoid robots remain at an early application stage because of technical bottlenecks, including embodied-intelligence brains.
The article describes new products spanning AI companion devices, offline computing terminals and embodied robotics. It says Haochen Keji's riding robot, developed with support from AGIQUAD, has opened global preorders, with small-scale customer deliveries planned for September this year.
Built on the Seed mixture-of-experts large language model architecture, Doubao's speech recognition model targets difficult cases including proper nouns, personal names, place names and ambiguous pronunciations. The source says stronger contextual reasoning enables multimodal understanding and accurate mixed-language recognition.
The University of Hong Kong's business school benchmarked 36 mainstream large language models on reasoning in Chinese contexts. Its report ranked GPT-o3 first in basic logic and Gemini 2.5 Flash first in contextual reasoning.
Zhang Lei said his company pursues 3 to 5-year contracts rather than short-term rental arbitrage. He said high-end computing rental prices rose 20% to 30% in the first quarter of 2026, with long-term orders locking in higher gross profit and customer deposits of 30% to 50% supporting cash-flow certainty.
China has increased policy support for financial leasing, particularly for green financing, equipment financing and financing for small and medium-sized enterprises. It has also introduced regulatory policies to standardize leasing companies' operations and prevent financial risks.
The report states that more than 4200 GPU computing rental patent applications were published globally in 2024-2025, focusing on virtualization scheduling, multi-tenant isolation and heterogeneous computing resource pooling. It says Chinese companies' share rose from 22% in 2023 to 37% in 2025.
The guide divides China's cloud market into commercially driven private platforms and state-backed telecom enterprises. It says price wars since 2023 have reduced compute costs by 15% to 50% across providers such as Alibaba, Tencent and China Mobile.
PPIO launched Agentic Cloud at the 2026 World AI Conference to build an intelligent Token factory for the Agent era. According to CIC Consulting, it ranked first among independent AI cloud computing service providers in China by average daily token consumption in 2025 and the first quarter of 2026.
Green Revolution Cooling received funding from Samsung Ventures, HTS, SK Enmove and ENEOS in June 2025 and partnered with Samsung C&T to scale immersion cooling. In March 2025, Vertiv launched the CoolLoop Trim Cooler to manage water temperature changes in liquid-cooled systems.
Liquid-cooling competition increasingly centers on system-level capabilities, technical suitability and delivery at scale. In real estate, oligopoly and regional competition coexist, with competitive factors varying across residential, commercial and industrial properties.
The report estimates that rack occupancy in the first half of 2026 remained above 90% at core facilities in first-tier cities and leading third-party IDC operators. It expects many newly built racks in western hubs, third- and fourth-tier cities and some second-tier new districts to remain at an estimated 50%–70%, based on industry research.
The guide says Eoptolink prioritizes 400G+ products over the 1G/10G/25G segments relied on by many private-label distributors. It describes InnoLight as having very high minimum order quantities and limited flexibility for custom branding or EEPROM coding.
The connection table lists GB200 in 2024 with copper scale-up connections and optical scale-out at 400G and 800G. It lists GB300 in 2025 with copper scale-up connections and optical scale-out at 800G and 1.6T.
OptiNet China 2026 is a two-day event in Shanghai for network architects, researchers and engineers. It focuses on optical networking and AI-scale infrastructure, including ultra-high-speed networks, photonic computing and AI-native fabrics.
The report says major mobile operators collectively budgeted over $150 billion for network infrastructure capital expenditure in 2025 and 2026. It also describes internally supplied transceivers at Huawei Technologies and ZTE Corporation as creating captive demand that largely bypasses independent suppliers in Chinese carrier deployments.
The article identifies Chinese CPO leaders based on technological breakthroughs, mass-production capabilities and ecosystem integration. It claims CPO reduces optical module power consumption from 10W/Gbps to 1W/Gbps and projects that Chinese companies could capture a 50% share of the global CPO market in 2026.
The report cites Lightcounting's forecast of 14% growth in the global telecom optical module market in 2026, with an expected compound annual growth rate of 11% through 2030. It also anticipates explosive growth in domestic low-orbit communications satellite launches in 2026 as constellation deployment accelerates.
The report values China's data center server market at USD 12.95 billion in 2025. It forecasts growth from USD 15.36 billion in 2026 to USD 36.68 billion by 2031, at a CAGR of 19.02% during 2026-2031.
The analysis identifies enterprise digital transformation and the "AI+" policy as drivers of demand for large AI models, with multimodal integration and AI agents expanding applications. It also describes GPU computing applications beyond graphics rendering as a core market growth driver.
The report examines China's AI chip market structure, technological evolution, downstream demand and competitive landscape. It forecasts steady industry growth as advanced manufacturing processes evolve and demand for domestic semiconductor substitution increases.
Photo by XPENG Egypt on September 05, 2026. May be an image of car, speaker, subwoofer, stereo and text that says 'XPENG G7 AI /AUDIO DEBUTS LINGJING X 20 CUSTOM HI-FI SPEAKERS + 7.1.4 PANORAMIC SURROUND SOUND Close the doors and enter a studio-grade listening room AI AI POWERED NEXT-GEN AlPOWEREDNE
EYOU Robotics launched what it describes as the world's first automated production line for robot joint modules at Shanghai Zhangjiang Robot Valley on January 21, 2026. Leaders from Pudong New Area, Zhangjiang Group and Zhangjiang Hi-Tech attended the ceremony.
XPeng commissioned its humanoid robot production lines in Guangzhou, where the first completed IRON unit walked off the line autonomously; core processes exceed 80% automation. The company targets mass production by the end of 2026, initially deploying units in its stores and corporate campuses before a commercial launch in China and international markets in 2027.
EngineAI announced the launch of its Honghualing intelligent manufacturing base in Shenzhen and the rollout of its first batch of flagship T800 full-size humanoid robots. The company says the milestone marks its entry into a phase of 10,000-unit scalable delivery capability.
The white paper describes extreme industrial environments as testing grounds for intelligent robots and autonomous mining trucks. It argues that demanding applications could improve energy-sector safety and foster specialized robots for hazardous operations.
The article describes competition in embodied robotics expanding beyond technical breakthroughs to mass production, reliable delivery and real-world applications. It cites Tian Hua as saying that Qiyuan robots roll off the production line every 2.5 minutes using automotive manufacturing and quality-management practices.
The roundup highlights funding for foundation models, robot chips, operating systems, physical simulation and synthetic data. It interprets these investments as a shift beyond humanoid robot hardware toward upstream technologies, with mass production, delivery and practical applications becoming key measures of company value.
The guide states that generative AI services with public-opinion attributes or social-mobilization capabilities require security assessments and algorithm filing procedures. Drawing on project experience and regulations, it describes a review process involving provincial preliminary review followed by final national review and publication.
The article cites industry monitoring showing high-end Nvidia GPU rental prices rising 15%-30% month on month, with H200 and H100 leading the increases. It attributes the pressure to surging AI demand and supply constraints, while reporting industry expectations that relief may come in 2027.
The article estimates annual global compute rental market growth at 30-50% and growth in China at more than 50%. It attributes demand to the spread of large AI models, deep learning and high-performance computing, and identifies cloud providers and data-center operators as market participants.
Guoyin Jinzu announced an agreement to purchase information technology computer equipment in Hebei from Datong Qinyuan Data Technology for RMB 189 million and lease it back. The lease term is 40 months.
