Ascend
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 7
- China's compute network faces 4 trillion yuan buildout with 36.8% utilization
China's intelligent compute supply was about 3.8 billion card-hours in 2025 against roughly 1.4 billion card-hours of actual use, a 36.8% utilization rate, according to CAICT figures cited in the article, while the National Data Administration put nationwide daily average token calls at 140 trillion in March 2026, up more than a thousandfold from about 100 billion in early 2024. The piece says self-built clusters at big internet companies and the three state carriers keep high loads, whereas compute centers funded by local government financing vehicles and state platforms still have to find customers, with some reporting under-20% rack occupancy and less than 30% actual utilization on installed servers. It cites an NDRC-quoted estimate of 4 trillion yuan in new direct investment in the compute network during the 15th Five-Year Plan, and Gartner projections that global inference-related AI infrastructure spending will reach about $23.3 billion in 2026, surpassing training-related spending of $190 billion...
36氪 人工智能 · 🔥 0 - Huawei details Kirin 9050, logic folding and a 10-year chip roadmap
Zhidx reported on Oct. 15 that Huawei recently held a two-day technical briefing in Shanghai, led by terminal BG CTO Li Xiaolong and Xu Songyu, who heads wireless terminal chip product management in Huawei's semiconductor business; it was Huawei's first detailed public walkthrough of Kirin chips in six years. Huawei said the Kirin 9050 series is the first logic-folding chip under its "Tao" approach, claiming hybrid-bonding pitch of 1.5 microns with more than 50 million bonds, and it disclosed a 10-year chip roadmap plus a claim that Kirin 2027 test metrics all met expectations. Huawei also said a 30B on-device model runs on the platform and cited figures including 23% higher multi-core CPU performance, 40% higher Maleoon GPU performance, 45% lower GPU power and 78% better ray tracing, along with an ISP with a pixel-level NPU that processes in under 16.7 milliseconds. All performance and process figures are Huawei's own and were not independently verified.
智东西 · 🔥 0 - China's non-GPU AI accelerator shipments overtake GPUs as big tech turns chip seller
IDC data shows China's data center AI accelerator shipments reached 2.008 million units in the first half of 2026, with non-GPU accelerators at 1.016 million — surpassing GPUs for the first time. Nvidia remained No. 1 with 34% share but its shipments fell 44% year-on-year, narrowing its lead over Huawei's Ascend (28%) to 6 percentage points; Alibaba's T-Head (7%) and Baidu's Kunlunxin (4%) joined Huawei and Cambricon in accounting for about 95% of the non-GPU segment. The report argues that with model profits still elusive, full-stack vendors are monetizing their own silicon: Alibaba plans Q1 2027 mass production for Zhenwu V900 and Panjiu super-node servers carrying it, Baidu is spinning off Kunlunxin (reported to seek a ~$50 billion valuation), and Huawei is bundling Ascend super-nodes while opening up CANN.
36氪 人工智能 · 🔥 0 - Intern Lumina U2: multi-codebook diffusion LLM for unified vision and image generation
The paper introduces Intern Lumina U2, which the authors describe as the first diffusion language model to represent multi-codebook visual tokens in a fully discrete space, where each visual position is encoded by a group of codes drawn from complementary codebooks to raise representational capacity while staying compatible with discrete text modeling. At the input the embeddings are concatenated per position, while at the output a lightweight multi-codebook autoregressive head predicts codes within each position sequentially and the diffusion backbone processes all positions in parallel. Built on AToken's unified 4D representation, it handles image, video and 3D inputs in a single backbone and, according to the authors, achieves leading results among unified models in visual understanding, image generation and image editing; a stability-aware sampling strategy combining confidence-plateau-guided unmasking with drift-triggered cache refreshing is reported to speed up image generation by 4.45x. The authors also say they built a full-stack training, inference and evaluation system on both Ascend and NVIDIA platforms to verify cross-platform effectiveness.
Hugging Face · Papers · 🔥 0 - Huawei-ECNU time-series model tops TIME-leaderboard and GIFT-Eval
According to Huawei Cloud, a time-series pretrained model called Qiyao, developed jointly by Huawei teams (Huawei Cloud industry large-model team, the 2012 Applied Scenario Innovation Lab and others) and East China Normal University, has topped two international time-series benchmarks: TIME-leaderboard and GIFT-Eval (Pretrained category), as reported by ITHome. Huawei Cloud said the team deeply optimized training and inference on the Ascend ecosystem, yielding a model it describes as highly accurate, robust, general-purpose and easy to engineer. The company says the model can be used directly for zero-shot time-series forecasting in areas such as grid load assessment, industrial equipment anomaly detection and financial trend analysis, and supports low-cost fine-tuning for vertical industries.
IT之家 AI · 🔥 0 - HiSilicon's HiSpark launches XDAS framework, digital FAE beta
Shanghai HiSilicon used an upgrade of its HiSpark developer community to introduce the XDAS four-pillar framework — AI for Development, AI for Application, AI for Support and EXperience for AI — arguing that the developer ecosystem for edge AI needs to be rebuilt around agents. The community has opened a first set of AI Skills spanning five domains, started a 24/7 digital FAE beta, and offers a model/algorithm marketplace plus quantization and compression tools, with an online training platform planned for year-end. Claims such as roughly 76% faster AI-assisted embedded development, over 85% of support questions being repeats, and a 10x download increase as of September 2026 are vendor statements and are not independently verified.
智东西 · 🔥 0 - MoWorld world model reaches Huawei Mate 90 phones
Moxin Technology is reported to have deployed its MoWorld-2.0-Flash world-model variant in the HarmonyOS-exclusive Moxiao Pet app for Huawei’s Mate 90 series. The source says cloud infrastructure generates 3D Gaussian assets while the approximately 600-million-parameter model runs on the phone for rendering, but the performance and cost claims are promotional and unverified.
新智元 · 🔥 0
Experience and discussion from the community
Share my Ascend experienceAsk about Ascend
Nobody has shared their experience with Ascend yet.