Signal Brief

中国AI芯片自给提速:进口依赖降至60%以下

据WSJ报道,中国AI芯片进口依赖从2021年的90%降至2025年的60%以下,摩根士丹利预计2030年降至25%。华为的Ascend芯片已可替代部分Nvidia任务,但其最强芯片算力仅为Nvidia最佳芯片的四分之一,且仍依赖外国设备。

twitter关注列表 Chubby♨️ (@kimmonismus) 发布 2026-07-24 收录 2026-07-24 观察

一句话判断

提供了中国AI芯片进口依赖度下降的具体数据,低于市场预期,值得关注其供应链变化。

核心信息

据WSJ报道,中国AI芯片进口依赖从2021年的90%降至2025年的60%以下,摩根士丹利预计2030年降至25%。华为的Ascend芯片已可替代部分Nvidia任务,但其最强芯片算力仅为Nvidia最佳芯片的四分之一,且仍依赖外国设备。

原始内容

China is doing what many thought export controls would prevent: building a domestic AI-chip stack that can start replacing Nvidia. Vis WSJ In 2021, around 90% of China’s AI chips came from abroad. By 2025, that figure had fallen below 60%. Morgan Stanley estimates it could drop to 25% by 2030. Without access to cutting-edge EUV machines, Huawei is combining older production tools with advanced packaging, chip stacking and massive clusters. Its Ascend chips can already replace Nvidia hardware for some AI tasks. China is still behind: Nvidia’s best chip offers roughly four times the compute of Huawei’s strongest chip. China also remains dependent on foreign equipment. But complete chip sovereignty no longer looks impossible. And that is a development almost nobody saw coming this quickly. To be perfectly clear: the biggest moat for the US right now remains chips and the embargoes. ASML is not allowed to sell EUV chips to China. Despite this, China is increasingly managing to develop its own chips, which are constantly improving. ![photo](https://pbs.twimg.com/media/HN_v8_xWgAA3G2k.jpg) ![photo](https://pbs.twimg.com/media/HN_wYKlW4AAkNpr.jpg) Chubby♨️ (@kimmonismus): Source: https://t.co/SkiEfvwTra

相关动态

01

NVIDIA CEO 呼吁政策制定者关注模型蒸馏技术

NVIDIA CEO Jensen Huang 在推特上引用并支持《Interconnects AI》发布的公开信,强调开源模型对AI安全性、创新加速和技术扩散至关重要,同时指出模型蒸馏等技术是合法的模型改进手段,应由精准的法律和商业框架加以规范,而非广泛限制。

twitter关注列表2026-07-24#AI产业#政策#技术
值得跟进
02

3D 自动拆解

Hyper3D 的 BANG to Parts 功能可将任意完整的 3D 模型自动拆解为多个完整、可编辑的零部件,支持材质替换、形状修改、3D 打印等操作,并可递归拆分,如整车拆出轮毂、轮胎、螺丝。该技术采用生成式分解,AI 理解部件拼接的自然接缝而非使用几何切割,实现更精确的结构解体。未来将在零件级模型上生成交互式动画资产,提升动画、游...

twitter关注列表2026-07-24#模型发布#技术突破#行业动态
观察
04

Offloop多智能体超越Claude Code

Offloop 4 人团队公布其多智能体 harness 在 GDPval 基准上以 84.9 分、单任务成本 $1.65 超越 Claude Code(Opus 4.8 得 82.4 分、$14.38) 与 Codex(GPT 5.6 Sol 得 83.3 分、$5.20),并声称在 GDP.pdf(44 分) 与 JobBench(6...

twitter关注列表2026-07-23#评测#技术突破#行业动态
观察
05

SANA-Video 2.0 发布

Enze Xie 团队发布 SANA-Video 2.0,采用混合线性-softmax 注意力、块注意力残差和 Sol-Engine 加速,统一 5B 和 14B 模型,在 16 节点 H100 上训练 5B 模型,VBench 总分 84.30,单 H100 生成 720p/5s 仅需 13.06s,速度比 Wan 2.2-A14B 快...

twitter关注列表2026-07-24#模型发布#技术突破#多模态
观察