China counted 4.95 million active AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. this year, a total the 2026 China Artificial Intelligence Computing PowerComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. Because it is expensive, physical and concentrated in a few suppliers, compute is the main lever governments pull to shape who can build frontier AI. Development Assessment Report projects will reach 197 million by 2030, per TechNode. Agents are systems that string together reasoning, software tools and other programs to finish a job on their own, a step past producing text and images. The same projection puts China's market for AI computing at $150.1 billion by 2030, four times the 2025 level. Worldwide, the totals are 2.22 billion active agents and a $1.25 trillion market for AI computing by the same year.
Read at TechNode ↗
Alibaba's chip unit T-Head unveiled the Zhenwu V900 AI accelerator, a chip built for AI workloads, reported to run three times as fast as the Zhenwu M890 it replaces, per TechNode. Chief Executive Wu Yongming told Alibaba's annual Apsara cloud conference in Hangzhou that the company will keep investing in three areas: its models; its own chips; and its cloud infrastructure. Alibaba Cloud said it wants to be running more than 20 gigawatts of data center capacity around the world by 2032.
Read at TechNode ↗
DeepSeek founder Liang Wenfeng told investors at a closed door meeting that getting Huawei and other Chinese accelerators to train his company's coming models is one of its biggest bets and has to succeed, per The Information. Earlier accounts of DeepSeek's large Huawei order had centered on inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. clusters, the hardware that serves finished models to users. The remarks move domestic chips into the training stage. People familiar with the remarks said DeepSeek expects another shipment of Huawei training chips between the last quarter of this year and the first quarter of 2027.
Read at The Information ↗ • Read at Pandaily ↗
Xiaomi released and open sourced its MiMo-V2.6 models overnight, five months after its last model, per Caixin. The series holds two models, Pro and Flash, that handle more than text. Xiaomi's model team said the gains came from scaling up the computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. Because it is expensive, physical and concentrated in a few suppliers, compute is the main lever governments pull to shape who can build frontier AI. spent on reinforcement learning, which trains a model by scoring its own attempts. Xiaomi said MiMo-V2.6-Pro scored 46 on the Artificial Analysis composite intelligence index, a third party benchmark, ahead of Moonshot AI's Kimi K3 and Alibaba's Qwen3.8 Max. Xiaomi priced the models at a twentieth to a sixtieth of comparable models sold outside China.
Read at Caixin ↗
Hygon Information Technology is due to unveil a new processor in its CPU1000 line at a Shenzhen event, taking the listed chipmaker beyond the data center CPUs that built its business, per the South China Morning Post and Chinese financial outlets. Its three existing processor families are sold for data centers, mainstream corporate buyers and cost sensitive deployments. Hygon has billed the event under the slogan that computing powerComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. Because it is expensive, physical and concentrated in a few suppliers, compute is the main lever governments pull to shape who can build frontier AI. should reach the physical world. The pre-event reports describe the chip as aimed at low power work inside robots, vision inspection systems and factory equipment. Pre-launch notices carried no core counts, clock speeds or power figures.
Read at Pandaily ↗ • Read at South China Morning Post ↗