文章

端侧AI七雄备案 · 英特尔代工破冰

#1128 · 2026-07-21 · 21ZHAO Blog
权威要闻时政调查思想深度解读财经
Reading Path / ARTICLE 先抓主张，再转成行动 #1128 · 21ZHAO Blog · 读完进入产品或下一篇

端侧AI七雄备案 · 英特尔代工破冰

一、 权威必看

EN: On July 15, the Cyberspace Administration of China published its latest batch of generative AI service filings, with seven on-device mobile AI products completing registration simultaneously: Apple Intelligence, Huawei Xiaoyi, OPPO Andes GPT, vivo BlueLM, Xiaomi HyperAI, Samsung Galaxy AI, and Nubia Doubao. Samsung’s Galaxy AI secured approval after partnering with ModelBest, whose self-developed MiniCPM series will ship on several flagship Samsung phones. ModelBest CEO Li Dahai projected that 70% of the future on-device AI market will be supplied by third-party model vendors, drawing a parallel to the autonomous driving industry’s division of labor. Google, BlackRock, Carhartt, and Ford jointly launched the Alliance for America’s Skilled Trades, channeling corporate resources into vocational training pipelines to address the skilled labor gap in the U.S. manufacturing and infrastructure sectors.

中： 7月15日，国家互联网信息办公室公布最新一批生成式人工智能服务备案，Apple智能、华为小艺、OPPO Andes GPT、vivo蓝心端侧大模型、小米澎湃AI、三星盖乐世AI及努比亚豆包手机大模型共7款产品同日获批。三星的备案背后是与面壁智能的深度合作，其MiniCPM系列端侧模型将搭载于数款三星旗舰机型上市。面壁智能CEO李大海在2026 WAIC期间给出明确判断：参考智能汽车自动驾驶的分工格局，未来端侧AI的70%市场将由第三方模型厂商提供，手机厂全栈自研不会是长期主流。面壁智能副总裁刘豪杰拆解了三星的核心需求——“端侧主内、云端主外”的架构，端侧掌握所有用户上下文和交互主导权，处理隐私数据，仅在需要外部实时信息时将脱敏任务拆给云端。另一侧，Google联合贝莱德、Carhartt与福特汽车，正式发起“美国技能行业联盟”，直接向制造业和基建领域的技能工人短缺问题投入企业资源，构建职业培训输送管道。

二、 深度与多元

EN: Intel secured its first external foundry customer under CEO Lip-Bu Tan, who took office in March 2025. The company will manufacture Fortinet’s next-generation security chip, the SP6, using the Intel 4 process, a 7nm-class node. Fortinet, whose stock has surged nearly 100% year-to-date, previously sourced its FortiSP5 from TSMC’s 7nm line, making this a direct customer win from the industry leader. The deal signals Intel’s foundry business is gaining operational traction after two fiscal years of revenue stagnation. Separately, Moonshot AI’s B端 business head Huang Zhenxin addressed the post-K3 landscape during a site visit by reporters to its Beijing Haidian office. He attributed the K3 model’s performance leap to foundational architecture innovation and a systematic rewrite of AI base mechanisms that had remained unchanged for 8 to 11 years, rather than mere engineering efficiency gains. Huang confirmed the company is evaluating an open-source strategy and acknowledged the compute capacity strain following the July 17 launch, while welcoming Elon Musk’s competitive challenge.

中： 英特尔代工业务迎来关键破冰。现任掌门陈立武2025年3月上任后，首次确认签约外部企业客户——网络安全公司飞塔信息。双方将合作开发飞塔下一代安全芯片SP6，采用英特尔4工艺制造，本质为7纳米节点。飞塔信息今年以来股价上涨近100%，其上一代FortiSP5芯片大概率由台积电7nm工艺制造，英特尔此次直接从行业老大手中抢下客户。这笔订单是英特尔代工业务走上正轨的重要信号，此前两个财年代工业务营收均陷于停滞。另一维度，月之暗面在7月17日发布Kimi K3引发全球震动后，B端企业业务负责人黄震昕首次系统回应外界焦点。他将K3的能力提升归因于参数量和底层架构创新，而非工程效率优化，并透露团队对沿用8到11年“无人更新”的AI基础机制进行了系统性重写。黄震昕同时确认公司正在评估开源策略，坦言算力紧张现状，并对马斯克的挑战表示“欢迎竞争，拭目以待”。

三、 科技与财经

EN: Google released Tunix, a JAX-native post-training library engineered to eliminate TPU idling during multi-turn, tool-using LLM agent training. Tunix decouples rollout generation from training via an asynchronous producer-consumer pipeline, maximizing hardware throughput while agents wait on network I/O or environment steps. It ships with plug-and-play abstractions and continuous macro-level profiling. On the governance front, the “Global Dialogue and Collaborative Development on AI Rules” forum convened during WAIC 2026, co-hosted by China University of Political Science and Law and UNESCO under the guidance of China’s Ministry of Justice. UNIDO Deputy Director-General Zou Ciyong stressed that developing countries and SMEs face gaps in computing power, data, infrastructure, skills, and standards, requiring responsible innovation and international cooperation to ensure equitable access to AI-driven industrialization.

中： Google发布JAX原生的后训练库Tunix，专门解决多轮次、工具调用型LLM推理智能体训练中的TPU闲置瓶颈。Tunix通过异步生产者-消费者管道将rollout生成与训练过程解耦，在智能体等待网络I/O或环境步进时持续向训练器输送数据，最大化硬件吞吐量，并内置即插即用抽象层与连续宏观性能剖析工具。同一日，由司法部指导、中国政法大学与联合国教科文组织联合主办的“人工智能规则全球对话与协同发展”论坛在2026 WAIC期间举行。联合国工业发展组织副总干事邹刺勇指出，发展中国家和中小企业在算力、数据、基础设施、技能和标准等方面存在显著差距，需通过负责任创新、能力建设和国际合作弥合鸿沟。中国联合国教科文组织全国委员会副秘书长葛思颖提出，当前AI全球治理呈现多组织、多机制并行推进的格局，面临规则碎片化和职能重叠问题，需要更加协调的国际回应。

四、 国际视野

EN: Chinese Foreign Minister Wang Yi met with ASEAN Secretary-General Kao Kim Hourn in Manila on July 21, 2026. Wang underscored that 2026 marks the 50th anniversary of the Treaty of Amity and Cooperation in Southeast Asia and the 5th anniversary of the China-ASEAN Comprehensive Strategic Partnership. He positioned China as the region’s foremost stability anchor and called for deeper cooperation to safeguard shared interests and advance a more just international order. Wang referenced the newly established World AI Cooperation Organization, launched during WAIC in Shanghai, and invited more ASEAN nations to join. On the South China Sea, he stated that provocative actions by certain Philippine military and police elements serve external forces and must not become an obstacle in China-ASEAN relations. In parallel, Google launched Gemini 3.5 Flash Cyber, a cost-efficient AI security model purpose-built for rapid vulnerability discovery and patching, positioned as a cheaper alternative to larger systems like Anthropic’s Mythos.

中： 2026年7月21日，中共中央政治局委员、外交部长王毅在马尼拉会见东盟秘书长高金洪。王毅明确两个时间锚点：今年是《东南亚友好合作条约》签署50周年，中国是最早加入该条约的大国；今年也是中国东盟建立全面战略伙伴关系5周年，定位为承前启后的重要年份。王毅将中国定义为“本地区最重要的稳定锚”，提出愿同东盟深化互信，维护全球南方正当权益，推动国际秩序朝更加公正合理方向发展。他特别提及中方刚刚在上海WAIC期间倡导成立的世界人工智能合作组织，旨在提升南方国家AI能力建设，欢迎更多东盟国家加入。在南海议题上，王毅直言菲律宾军警部分势力蓄意挑衅、制造海上事端，服务的是外部势力，南海问题不应成为中国东盟关系的障碍。科技侧，Google推出Gemini 3.5 Flash Cyber安全模型，定位为Anthropic Mythos等大型安全AI系统的低成本替代方案，专攻快速漏洞发现与修补。

五、 青年与生活

EN: Microsoft agreed to invest billions of dollars in French AI startup Mistral to build computing infrastructure across Europe and integrate Mistral’s models into Azure, Foundry, and Copilot Studio. Mistral’s Medium 3.5 and OCR 4 models are now available on Microsoft’s developer platform, and enterprises running Azure Local can opt to deploy Mistral’s open models. Gartner analyst Arun Chandrasekaran framed the deal as a response to the “sovereign AI” trend, giving European customers a regulatory-compliant alternative to U.S.-dominated infrastructure. On the consumer side, the 2027 Kia Seltos was unveiled with a starting price just under $25,000. The compact SUV adopts design cues from the larger Telluride, offering a spacious interior and updated tech features at an entry-level price point, though the hybrid variant will not arrive until next year.

中： 微软同意投入数十亿美元，扶持法国AI初创公司Mistral在欧洲建设算力基础设施，同时将Mistral的AI技术纳入Azure云与软件生态。Mistral旗下Medium 3.5和OCR 4模型已接入微软Foundry开发者平台，Copilot Studio也引入Medium 3.5。通过Azure Local接入微软服务的企业客户，未来可选择运行Mistral的开放式模型，根据自身需求进行定制部署。Gartner分析师Arun Chandrasekaran将此合作定性为对“主权AI”趋势的直接回应——为受监管约束的欧洲行业提供不同于美国主导基础设施的替代方案。消费市场端，2027款起亚Seltos正式亮相，起售价低于25,000美元。这款紧凑型SUV汲取了Telluride的设计语言，在入门价位提供宽敞座舱和升级后的科技配置，但混动版本需等到明年才会上市。

【21ZHAO 综合判断】

EN: July 21 reveals a dual-axis convergence: AI supply chains are simultaneously localizing and globalizing. On the localization axis, China’s seven-phone AI filing batch and ModelBest’s 70% third-party market projection confirm that on-device AI is entering a regulatory-governed, multi-vendor ecosystem phase—privacy architecture (“端侧主内、云端主外”) becomes a hard requirement, not a feature. Intel’s Fortinet win and Microsoft’s Mistral investment both respond to the same sovereign-resilience logic: customers want silicon and cloud capacity that is geographically and politically diversified. On the globalization axis, Google’s Tunix and Gemini Flash Cyber target the developer toolchain and security layers that underpin all of these distributed deployments. Wang Yi’s ASEAN diplomacy and the WAIC governance forum add the institutional layer—rules and alliances are being built in parallel with the infrastructure. The Kia Seltos pricing reminds us that consumer hardware cycles still anchor the physical endpoints where all this AI lands.

  • When building on-device AI features, design for the “端侧主内” split from day one: local inference handles context and privacy, cloud handles real-time external queries only after task de-identification. This architecture is now a regulatory and OEM procurement baseline across seven major phone brands.
  • If you operate AI training infrastructure, study Tunix’s producer-consumer decoupling pattern. The core insight—separating rollout generation from training to eliminate idle GPU/TPU cycles during agent I/O waits—is directly applicable to any multi-turn agent training pipeline, not just JAX environments.

中： 7月21日的五条线索指向同一个结构判断：AI供应链正在同时经历本地化与全球化的双向收敛。本地化一侧，7款手机AI同日备案、面壁智能给出70%第三方市场份额的明确数字，标志着端侧AI已进入监管规范下的多厂商生态阶段，“端侧主内、云端主外”的隐私架构从卖点变为准入门槛。英特尔从台积电手中抢下飞塔订单、微软斥资数十亿美元帮Mistral在欧洲建数据中心，回应的是同一个主权韧性逻辑——客户要求芯片产能和云算力在地理和政治上实现多元化分布。全球化一侧，Google的Tunix和Gemini Flash Cyber分别瞄准开发者工具链和安全层，这两层恰恰是所有分布式部署的共同底座。王毅在马尼拉的外交动作与WAIC治理论坛则补上了制度层——规则与联盟正与基础设施同步构建。起亚Seltos以低于25,000美元定价入场，提醒我们消费硬件的更新周期仍是所有AI落地的物理终点。

  • 端侧AI产品设计必须从第一天起就按“端侧主内”架构分工：本地推理掌握上下文与隐私数据，云端仅在任务脱敏后处理外部实时查询。这一架构已成为7家主流手机品牌备案通过的共同基线，也是OEM采购的硬性要求。
  • AI训练基础设施团队应深入研究Tunix的生产者-消费者解耦模式。其核心思路——将rollout生成与训练过程分离，消除智能体等待I/O期间的GPU/TPU空闲周期——可直接复用到任何多轮次智能体训练管线中，不限于JAX生态。

参考来源