|
|
|
|
Bottom Line Up Front SMIC reported Q1 revenue growth of 11.5% to $2.505 billion, with Co-CEO Zhao Haijun telling the May 15 earnings call that AI driven pricing effects will expand through Q3 and Q4 amid a continuing memory shortage, though net profit fell 28.6% to $231 million on capacity expansion depreciation and gross margin landed at 20.2%. MIIT, SAMR and MOFCOM launched conformance testing under the new GB/Z 177-2026 AI terminal grading standard on May 13 via CAICT, applying a 2-plus-N architecture across seven product categories including mobile terminals, glasses and car cockpits, with L1 to L4 grading and a testing window open through June 30, 2026. Guangdong issued a 2026-2030 cultivation plan on May 15 naming smart chips, basic software and industrial vision as priority targets, requiring proof-of-concept centers, pilot platforms and open source communities while supporting data product registration for qualifying enterprises. |
 |
| SMIC says AI demand will expand its price hike cycle through Q4 after Q1 revenue rose 11.5% Mainland China's largest wafer foundry posted Q1 2026 revenue of $2.505 billion, up 11.5% year-over-year, after AI server, memory and power management chip demand left mature node capacity short, Caixin reported. Co-CEO Zhao Haijun said on the May 15 earnings call that the pricing effect will become more evident in Q2 and is expected to expand further in Q3 and Q4, and that there are no signs of easing in the memory shortage. Net profit fell 28.6% to $231 million on depreciation from continued capacity expansion, and Q1 gross margin came in at 20.2%. |
| MIIT, SAMR and MOFCOM open first-round testing under a new AI terminal grading standard The three economic ministries jointly published the GB/Z 177-2026 series national standard for grading AI capabilities in consumer terminals, and the China Academy of Information and Communications Technology opened first round conformance testing on May 13, ITHome reported citing CAICT. The standard uses a 2-plus-N architecture covering a reference framework, general requirements, and seven product categories: mobile terminals, microcomputers, TVs, glasses, car cockpits, speakers and headphones. The grading runs from L1 response level to L4 collaborative level, and the open testing window runs through June 30, 2026. |
| Guangdong names smart chips, basic software and industrial vision in 2026-2030 enterprise cultivation plan The General Office of the Guangdong Provincial People's Government issued the "Action Plan for Accelerating the Construction of a Key Enterprise Gradient Cultivation System (2026-2030)" on May 15, naming smart chips, basic software and industrial vision as priority targets for key core technology research, Yicai reported. The plan mandates construction of proof-of-concept centers, pilot platforms and open source communities, and instructs the province to support data product registration for qualifying enterprises. The cultivation system runs through 2030. |
| FiberHome mass-produces a 13,824-core optical cable framed for 100,000-GPU AI data center connectivity FiberHome said it has developed and mass-produced a 13,824-core ultra-high-count optical cable, framing it as compatible with ultra-large-scale AI data center construction, ITHome reported. The cable is targeted at inter-building optical connectivity for 100,000-GPU-class compute clusters. FiberHome said no domestic vendor has previously produced a higher count fiber cable. |
|
Ant Group's Bailing team open-sources Ring-2.6-1T, a trillion-parameter reasoning model with adjustable effort modes Ant Bailing released Ring-2.6-1T with weights on Hugging Face and ModelScope, branding it the team's flagship reasoning model with an adjustable reasoning effort control offering high and xhigh modes, InfoQ China reported. The release adds a Chinese open-weights trillion-parameter-class reasoning system alongside DeepSeek and Qwen lines. Ant Bailing said the model is also available on OpenRouter with a limited time free API trial. |
| Baidu forms an internal Model Committee to centralize large language model R&D strategy Baidu announced the establishment of the Baidu Model Committee, tasked with integrating internal AI resources behind a single large language model R&D strategy, PingWest reported citing Jiemian News. The committee is described as primarily composed of younger researchers. Baidu has not publicly disclosed the BMC's named members, reporting line or budget authority over the Apollo and Kunluncore workstreams. |
| Huawei Ascend ships Fully Async feature for open source VeRL, claims 1.6 to 2.0 times throughput uplift Huawei's Ascend team integrated a Fully Async feature into VeRL, the open source reinforcement learning framework that supports PPO, GRPO and DAPO training, CSDN reported. The feature decouples training from inference rollout to address long-tail latency in synchronous RL pipelines, with a claimed 1.6 to 2.0 times performance uplift over the synchronous baseline. CSDN did not publish the benchmark model size, Ascend SKU or whether the uplift holds equally across PPO, GRPO and DAPO. |
|
| III | CHIPS & SEMICONDUCTORS | |
| Intellifusion says its next-generation DeepVerse cloud inference chip will use a domestic process node AI chip designer Intellifusion told institutional investors that its newest DeepVerse series chip, aimed at large model cloud inference, will be built on a domestic process node, and that the company has laid out a three-generation roadmap for the line, Yicai reported. The disclosure did not name the domestic foundry or the specific process node. Intellifusion plans three chip generations over the next three years. |
| Yunnan Tin says affiliate holds 7N and 8N high-purity indium production capacity with expansion under way Yunnan Tin disclosed in an investor briefing that its affiliate Yunnan Tin New Materials Co., Ltd. currently has 7N and 8N high-purity indium production capacity, and that capacity expansion is being carried out simultaneously, Yicai reported. Indium is on China's export control list for technology relevant materials and is used downstream in semiconductor and display manufacturing. Yunnan Tin did not disclose the tonnage of 7N and 8N capacity or the timing and scale of the expansion. |
|
| IV | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Ant Group's LingBo open-sources LingBot-VLA real-robot post-training toolchain, claims 150 episodes suffice for adaptation Ant Group affiliate LingBo open-sourced the full real-robot post-training toolchain for its LingBot-VLA foundation model on May 15, with a claim that 150 demonstration episodes suffice to adapt the model to a new robot, InfoQ China reported. The toolchain supports multi-LeRobot data flow and is designed to handle differences in arm geometry, end-effectors, sensors and control interfaces across heterogeneous robots. InfoQ did not link to independent validation of the 150-episode adaptation claim. |
| BYD driver-assistance fleet tops 2.99 million units, DiPilot logs more than 190 million km of data daily BYD said its driver-assistance fleet, including Huawei Qiankun variants, has crossed 2.99 million units, ITHome reported. The company said DiPilot generates more than 190 million kilometers of driving data per day. April additions to the driver-assistance fleet totaled 140,006 units. |
|
|
|