NVIDIA announced a joint capacity expansion with six supply chain partners—IBIDEN, Samsung Electro-Mechanics, Unimicron, Amkor, KYEC, and SPIL—across four areas: packaging substrates, advanced packaging, chip testing, and CPO optical engines. Substrate capacity is targeted to reach 2.5 times current levels within three years, interposer wafer capacity to increase by more than 300% by the end of 2028, chip testing capacity to double by 2027 (backed by a USD 11 million/billion prepayment), and optical engine capacity to expand to 15 times current levels by mid-2027. The expansion is primarily located in Taiwan (China), Japan, South Korea, and the United States, with capacity coming online in phases starting 2027.
On 10 18, 2026, at the CCF HPC China2026 conference, the National Supercomputing Internet Core Node (Sugon 8000) released ten frontier applications spanning cosmological simulation, machine learning force fields, digital twin brain, unified weather–climate forecasting, and more. Large-scale running card counts concentrate between 40,000 and 90,000 cards, with multiple 10,000-card-class tasks maintaining weak scaling efficiency above 90%. Since coming online in Zhengzhou in 7 2026, the platform has completed nearly one thousand deep adaptations of integrated supercomputing and intelligent computing applications.
According to Securities Daily, Sugon centrally released ten frontier application results for Sugon 8000 during the 2026 CCF National Conference on High Performance Computing. The system adopts a natively integrated supercomputing-AI architecture, uniformly scheduling high-precision scientific computing and AI workloads, and has completed large-scale validation in meteorology, astronomy, semiconductor materials, and other fields; over 40% of its deployed workloads are integrated supercomputing-AI composite tasks.
MiniMax announced on 10 (23) the open-sourcing of the MiniMax M3.1 Flash model weights, with 428B total parameters and 23B activated parameters. It delivers comprehensive capability improvements over the previous-generation M3 model, with inference efficiency enhanced through model-wide sparse attention, quantization-aware training (QAT), and Power-aware scheduling. After launching on 10 (19), the model processed over 100 trillion tokens from real users within 72 hours, and secured day-one adaptation from multiple domestic and international chip vendors and inference cloud platforms at open-source release.
On 10 23, MiniMax open-sourced its multimodal large model MiniMax M3.1 Flash. On the same day, Moore Threads' flagship AI training-and-inference compute card MTT S5000 and its proprietary MUSA software stack completed Day-0 adaptation of the model. With the adaptation timing synchronized with the model's open-source release, the model can run on Moore Threads' domestic GPU platform as soon as it is open-sourced, though the source provides no details on performance, deployment scale, or ecosystem.
According to CCTV News, at the 2026 National High Performance Computing Academic Conference held in Shenzhen on 10 month 18, China's first fully domestically produced 100,000-card AI super cluster made its real-machine debut, named Sugon 8000. Its biggest feature is "native integration of supercomputing and intelligent computing": the same underlying platform can run both high-precision scientific computing and large model training and inference, with computing power capable of supporting digital twin brains and simulations of the evolution of the universe across tens of billions of light-years, and ten major application achievements will be released.
According to CCTV News, citing DOIT, at the 2026 National High Performance Computing Academic Conference, China's first fully domestically developed 100,000-card AI supercluster was showcased as a real machine. The system is hosted by Sugon 8000, interconnecting 100,000 AI accelerator cards at high speed, adopting a natively fused supercomputing and AI computing approach that supports high-precision scientific computing as well as large model training and inference on a single computing foundation. Ten major application achievements were released, supporting research tasks such as digital twin brains and cosmic evolution simulation at the scale of tens of billions of light-years.
According to a Cailian Press report on 10 21, NVIDIA stated that cloud service providers in Europe, the Nordics, Africa, and Central Asia are expanding AI factories based on NVIDIA technology, covering model training, inference, and post-training. Nscale plans to deploy Vera Rubin infrastructure, Crusoe is developing liquid-cooled modular AI factories, and Nebius is using the Groq 3 LPX integrated with Vera Rubin NVL72; multiple cloud providers also support the OpenShell reference architecture, providing access control and secure operation mechanisms for AI agents.
Unitree Technology released the humanoid robot “H2-X Kung Fu Legend Guardian” on 10 21. Official demos show it lifting a car bare-handed and denting a car door with a kick; its positioning includes a “private bodyguard” scenario, with a claimed maximum battery life of about 8 hours—roughly 2.7 times the 3-hour battery life of the H2 platform. Unitree has not yet disclosed H2-X's full specifications, pricing, or commercialization plans, nor the test conditions behind the 8-hour battery life claim. Earlier, in 9, Unitree released the world action model UnifoLM-X2-1.0 for fully autonomous combat.
Zhiyuan (AGIBOT) announced on 10/21 that its "AI WAVE" event is scheduled to launch on 10/22, as a prelude and extension of the 2026 AGIBOT INNOVATION DAY. From 10/22, Zhiyuan will release seven milestones in succession—datasets, an operation platform, robot bodies, an Agent framework, a simulation platform, a foundation model, and a Collective Experience Learning paradigm—showcasing its full-stack embodied AI technology portfolio.
Ling-3.1-flash, open-sourced by Ant Group's inclusionAI, is a hybrid-architecture MoE language model with 560B total parameters and 25B activated per token, with FP8 quantized weights. According to a CSDN blog compilation, as of 2026 10, 15, the SGLang official deployment page verified two recipes: 8×H200 and 4×GB300; the official 54 safetensors shards total approximately 573.8 GB. Deployment requires multi-GPU resources and a dedicated image; the context length has three definitions — 1M nominal, 256K on the official service, and 262,144 in the weight config — and first deployments should start with a short context.
On 20 10, DeepSeek announced it would fully open-source its internal inference framework DeepInfer under the Apache-2.0 license, along with the MIT-licensed token encode/decode component deepseek-recipe. DeepInfer has already supported DeepSeek V4.1-Flash large-scale production inference and currently supports NVIDIA SM100 and Huawei Ascend 950 platforms, but it is an offline inference engine without online serving or a standalone KV cache system, and supports only the DeepSeek series of models.