At the 17th Caixin Summit roundtable on 12 11, 2026, SiliconFlow founder and CEO Yuan Jinhui said that in 2026 inference surpassed training for the first time as the dominant consumer of compute, while supply across chips, memory, data center facilities, and power remains severely insufficient. He believes the distributed nature and heterogeneous compatibility of inference bring reshuffling opportunities for new players, and open-source models will become a strategic direction for enterprises in terms of capability, cost, and data sovereignty.
At the 2026 Mobile Network Conference, Hygon Information showcased the achievements of its CPU and DCU dual-chip solutions in edge computing, token-based operations, and trusted computing, and was named a strategic supplier of China Mobile. The article introduces the 1000 Series CPU for low-power edge scenarios (typical TDP as low as 10W, supporting wide-temperature operation from -40°C to 105°C), as well as scheduling and security measures such as NUMA affinity, PD disaggregation optimization, and encrypted transmission.
The NBD Digital Intelligence Morning Briefing for 20261110 covers three developments: on 119, the "2026 Space Cloud Industry Ecosystem Conference" unveiled the "Nebula Constellation Plan," advancing unified scheduling of space-based and ground computing; Musk said Tesla may offer the Optimus humanoid robot for lease rather than sale, has begun early mass production, and expects delivery to external customers around the end of 2027; Apple's research team disclosed on arXiv HALO, an autonomous control system for humanoid robots, focusing on zero-shot simulation-to-real-world transfer.
On 11 2, the 2026 World Internet Conference Leading Science and Technology Award was presented in Wuzhen, and Huawei's Ascend 950 SuperNode project won. Based on the UnifiedBus interconnect protocol, each node is configured with 1024 NPU cards, over 1300 UnifiedBus interconnect chips, and over 4000 optical modules. Officially, the 1024-card SuperNode's training compute utilization can reach 45%, with inference latency as low as 5 ms. Over 1000 Ascend SuperNode systems have been deployed, serving internet, telecom, finance and other industries.
On 11 9, Google released TPU 8tx and TPU 8i Max, targeting confidential inference and ultra-long context processing respectively. TPU 8tx reuses TPU 8t chips and integrates hardware-level confidential computing, with inference latency reduced by up to 50%; TPU 8i Max integrates two TPU 8i chips in a single package, doubling compute, SRAM, and HBM bandwidth. Google also expanded its software stack with TorchTPU, the agent toolset, Agent Compute Engine, TPU Sync, and Cloud Pathways. Four eighth-generation TPU systems have now been released this year.
On 2026 of 11, 2, the Ascend 950 SuperPoD (Atlas 950 SuperPoD) won the Leading Tech Award at the 2026 World Internet Conference Wuzhen Summit, standing out from more than 400 entries. The supernode adopts the Lingju interconnect protocol and supports configurations of 128 or 1024 NPU units; officially claimed figures include training compute utilization of 45% and inference latency as low as 5 milliseconds, targeting training and inference scenarios for trillion-parameter-scale large models.
The 22nd CCF National High-Performance Computing Conference opened in Shenzhen on 18 10, 2026. Sugon8000, a fully domestic 100,000-card AI supercluster, made its first public appearance as a real system, sharing the stage with Lingsheng—which topped TOP500 with an all-CPU architecture—and the Kunpeng supernode. The report argues that the benchmark for evaluating compute is shifting from peak compute toward scaling efficiency, heat load per unit area, and ecosystem adaptation, and the conference released the New Era Supercomputing Industry Collaborative Development Action Plan along with the Superintelligence Collaboration Open Community.
On 21 10, 2026, NVIDIA introduced the DSX POD modular AI factory reference design, integrating the Vera Rubin computing system with power, liquid cooling, and networking infrastructure into shipping containers, with integration and testing completed before leaving the factory. Partners plan to begin global deliveries in early 2027; deployment time is said to be reduced from about six months to as fast as 14 days, though that figure excludes site construction.
Google and Meta open-sourced TorchTPU at PyTorch Conference North America on 2026 10, 20. Developers can run PyTorch models on TPUs by changing .to("cuda") to .to("tpu") in device initialization. The project is released under the Apache 2.0 license, built on the PrivateUse1 backend, compatible with torch.compile and XLA/StableHLO, and officially validated on 5000 Hugging Face models with numerical consistency of 97%. However, the closed TPU instruction set, single-process limitation, and communication overhead still constrain peak performance.
Enflame Technology made its exhibition debut at the 2026 China International Industry Fair Integrated Circuit Expo (NEXT-GEN IC EXPO), held at the National Exhibition and Convention Center in Shanghai from 10 12 to 16, showcasing the Yunsui ESL64-O high-performance supernode. The solution, jointly developed with ZTE based on Enflame's self-developed accelerator modules, adopts an orthogonal architecture and zero-cable design, supporting networking of intelligent computing clusters at the 10,000-card level. Its self-developed domestic-chip-and-domestic-model high-efficiency adaptation and optimization system won the 2026 NEXT-GEN IC EXPO Innovation Achievement Award.