AI科技热点日报 | 2026年7月7日

文章目录

AI科技热点日报 | 2026年7月7日

综合整理自TechCrunch、The Verge、Reuters Technology、VentureBeat、MIT Technology Review、钛媒体、36氪、虎嗅、新华网科技、网易科技、雷锋网等国内外主流科技媒体当日报道,

聚焦人工智能领域核心动态,附资深媒体人综合解读。


📌 今日摘要

今日,全球AI行业的注意力被即将于7月17日---20日召开的2026世界人工智能大会(WAIC 2026)锁定。7月7日上午上海市政府新闻发布会上披露的数据显示,WAIC 2026展览总面积首次突破10万平方米,1100余家企业参展,300余款产品将实现全球首发。华为Atlas 950超节点、阶跃Agent操作系统、全球首款AI智能体手机等硬核国产AI新品确定首秀,将"智算"与"具身"两大赛道推上千亿规模门槛。与此同时,国家发展改革委宣布将发布《中国智·惠世界》案例集与《人工智能合作发展行动计划》,将大会升格为中国主导的全球AI治理议程平台。海外方面,OpenRouter数据显示中国AI大模型周调用量达23.45万亿Token,连续十周领跑全球;微软宣布25亿美元投入新设AI实施部门;Amazon宣布停止接受Mechanical Turk新客户,AI全面替代人类标注员的趋势进一步具象化。

Global AI industry attention has locked on to the 2026 World Artificial Intelligence Conference (WAIC 2026), scheduled for July 17--20. At the Shanghai municipal government press conference held on the morning of July 7, organizers revealed that WAIC 2026 will break the 100,000-square-meter mark for exhibition scale for the first time, hosting more than 1,100 exhibiting companies and featuring over 300 global product debuts. Confirmed headline hardware includes Huawei's Atlas 950 SuperPoD, StepFun's Agent operating system, and the world's first AI-agent smartphone, signaling that the "intelligent computing" and "embodied intelligence" tracks have crossed the 100-billion-yuan threshold. At the same time, China's National Development and Reform Commission announced it will release the "China Intelligence Benefiting the World" case collection and the AI Cooperation and Development Action Plan, upgrading the conference into a China-led platform for global AI governance. Overseas, OpenRouter data shows Chinese AI large-model weekly invocation reached 23.45 trillion tokens, leading the world for ten consecutive weeks; Microsoft announced a US$2.5 billion investment to set up a new AI implementation division; and Amazon confirmed it will stop accepting new Mechanical Turk customers, further concreting AI's substitution of human annotators.


一、WAIC 2026倒计时:展览规模首破10万平方米,300余款产品全球首发

事件概要

7月7日上午,上海市政府召开WAIC 2026新闻发布会宣布,本届大会将于7月17日---20日在世博、张江、西岸"三地四馆"同步启幕,主题为"智能伙伴,共创未来"。展览总面积首次突破10万平方米,1100余家企业参展,3000余项展品集中亮相,其中超过300款产品将实现全球首发。智算、具身两大赛道各汇聚超过200家企业,规模均创历届之最。本次大会策划140余场论坛,联动12个国家部委、8个国家级重点实验室、10余个国际组织,集聚1400余位中外嘉宾,形成覆盖"会前、会中、会后"的产业服务闭环,会前已累计达成162亿元意向合作。

At the WAIC 2026 press conference held on the morning of July 7 in Shanghai, organizers confirmed that the conference will run from July 17 to July 20 across the "three zones and four venues" of World Expo, Zhangjiang and West Bund, under the theme "Intelligent Partners, Creating the Future Together." Exhibition floor area has exceeded 100,000 square meters for the first time; more than 1,100 companies will exhibit; 3,000+ exhibits will be on display; and over 300 products will achieve their global debuts. Both the intelligent-computing and embodied-intelligence tracks will each gather more than 200 companies, setting a record for WAIC's history. The conference plans 140+ forums, linked to 12 national ministries, 8 national key laboratories, and 10+ international organizations, gathering 1,400+ Chinese and international guests, forming a closed-loop service that covers pre-, during-, and post-conference stages, with 16.2 billion yuan in pre-conference cooperation intent already secured.

来源 / Sources


二、华为Atlas 950真机首秀:国产AI算力巅峰亮相展会

事件概要

上海市经信委主任汤文侃在7月7日的发布会上披露,业界最大规模超节点华为Atlas 950 SuperPoD真机将在WAIC 2026首秀。Atlas 950基于灵衢互联,最大可支持8192张昇腾NPU卡高速互联,提供8 EFLOPS(FP8)算力与16.3 PB/s互联带宽,是面向万亿参数大模型训练与推理部署的核心底座。同期亮相的还包括阶跃Agent操作系统、近存计算3D芯片、全球首款AI智能体手机,以及多款人形机器人和AI灵巧手。这些产品集中体现了国产AI从底层算力到上层智能体的端到端能力跃迁,是华为过去一年联合生态伙伴推动"灵衢+开源"双轮战略的阶段性成果。

At the July 7 press briefing, Shanghai Municipal Economic Information Committee director Tang Wenkai revealed that Huawei's Atlas 950 SuperPoD --- the industry's largest-scale supernode --- will make its real-machine debut at WAIC 2026. Atlas 950, built on the LingQu interconnect, supports up to 8,192 Ascend NPUs in high-speed fabric, delivering 8 EFLOPS of FP8 compute and 16.3 PB/s of interconnect bandwidth, forming the core substrate for trillion-parameter model training and inference. Also debuting alongside it are StepFun's Agent operating system, near-memory 3D compute chips, the world's first AI-agent smartphone, and multiple humanoid robots and AI dexterous hands. Together these products demonstrate Chinese AI's end-to-end capability leap --- from underlying compute to upper-layer agents --- and represent the phased results of Huawei's joint ecosystem "LingQu + open source" dual-engine strategy over the past year.

来源 / Sources


三、国家发改委两项重要成果亮相:构建全球AI治理中国方案

事件概要

国家发展改革委创新和高技术发展司副司长王若蒙在WAIC 2026 7月7日新闻发布会上宣布,组委会将在本届大会期间正式发布两项重要成果:一是《中国智·惠世界》案例集,自2024年起已连续发布三年,本年度案例覆盖20多个国家,涉及农业、工业、能源、科研等领域,集中展现"人工智能+"行动的产品、技术与实践经验;二是《人工智能合作发展行动计划》,覆盖智能算力普惠、开源生态共享、人工智能赋能、安全治理协作等8个方面。外交部同期披露,将邀请几十个国家和国际组织派高级别代表出席大会,并将为代表团安排一次集体巡展,让各国与会代表亲身体验中国AI发展与治理成就。这是中国首次以大会为平台,系统性输出AI治理方案与中国实践样本。

At the July 7 WAIC 2026 press conference, Wang Ruomeng, deputy director of the Innovation and High-Tech Development Department of the National Development and Reform Commission, announced two major outcomes to be officially released during the conference: first, the "China Intelligence Benefiting the World" case collection, now in its third consecutive year of publication since 2024 --- this year's edition covers more than 20 countries across agriculture, industry, energy, and scientific research, showcasing the products, technologies, and practices accumulated under the "AI Plus" initiative; second, the AI Cooperation and Development Action Plan, covering eight dimensions including inclusive intelligent computing, open-source ecosystem sharing, AI empowerment, and safety-governance collaboration. The Ministry of Foreign Affairs also disclosed it will invite senior representatives from dozens of countries and international organizations to attend, and will arrange a collective delegation tour, allowing participants to experience firsthand China's AI development and governance achievements. This will be the first time China uses the conference as a platform to systematically export its AI governance framework and practical sample cases.

来源 / Sources


四、中国AI大模型周调用量23.45万亿Token,连续十周领跑全球

事件概要

据OpenRouter最新数据测算,6月29日至7月5日一周内,全球AI大模型总调用量为46.7万亿Token。中国AI大模型周调用量达23.45万亿Token,环比增长15.01%,连续十周实现增长;同期美国AI大模型周调用量为4.28万亿Token,环比增长4.28%。DeepSeek-V4-Flash连续七周位居榜首,周调用量达5.34万亿Token,环比增长15%。这表明在OpenAI、Anthropic等海外厂商主导高端闭源模型的同时,中国大模型已经在Token消耗量这一反映真实生产力的指标上,建立起近5.5倍于美国的领先优势,且差距仍在持续扩大。

According to the latest OpenRouter data, in the week from June 29 to July 5, global AI large-model invocation reached 46.7 trillion tokens. Chinese AI large models recorded 23.45 trillion tokens in weekly invocation, up 15.01% week-over-week, growing for ten consecutive weeks; US AI large models recorded 4.28 trillion tokens in the same period, up 4.28% week-over-week. DeepSeek-V4-Flash remained at the top of the chart for the seventh consecutive week, with weekly invocation of 5.34 trillion tokens, up 15%. These numbers reveal that while OpenAI, Anthropic and other overseas vendors dominate the high-end closed-source model segment, Chinese models have built up roughly a 5.5x lead over US models on this productivity-reflecting metric --- and the gap continues to widen.

来源 / Sources


五、芯片与企业AI:微软25亿美元成立新部门、谷歌TPU转向英特尔EMIB-T封装

事件概要

7月6日国际芯片与AI企业领域多家头部公司密集披露重要动态。微软公司宣布拟投资25亿美元,派遣6000名员工成立新的人工智能实施部门,主攻企业级AI落地服务。谷歌决定下一代张量处理器(TPU)先进封装采用英特尔EMIB-T封装技术,标志其供应链策略向多元化转变。美光科技宣布斥资1.5万亿日元扩建日本广岛工厂,重点量产高带宽存储器(HBM)等先进存储芯片,用于满足AI算力需求。三星电子DRAM在第一季度涨价90%、第二季度涨价60%的基础上,第三季度将再涨价20%,AI驱动的存储涨价潮持续。国内方面,深圳具身智能企业智平方完成近50亿元新融资,公司估值突破200亿元,成为2026年具身智能赛道最大单笔融资之一。

On July 6, multiple leading chip and AI-enterprise players disclosed important moves. Microsoft announced a US$2.5 billion investment to dispatch 6,000 employees to set up a new AI implementation division, focused on enterprise AI landing services. Google decided that its next-generation Tensor Processing Unit (TPU) advanced packaging will adopt Intel's EMIB-T packaging technology, marking a diversification of its supply-chain strategy. Micron announced an investment of 1.5 trillion yen to expand its Hiroshima, Japan facility, focusing on mass-producing advanced memory chips such as HBM to meet AI compute demand. Samsung Electronics DRAM will raise prices by another 20% in Q3 after Q1 increases of 90% and Q2 increases of 60%, with the AI-driven memory price-up cycle continuing. Domestically, Shenzhen embodied-AI company Zhipingfang completed a new financing round of nearly 5 billion yuan, pushing its valuation past 20 billion yuan and making it one of the largest single-round financings on the embodied-AI track in 2026.

来源 / Sources


六、Amazon停止接受Mechanical Turk新客户:AI全面替代人类标注员进入倒计时

事件概要

TechCrunch 7月6日报道,Amazon宣布其Mechanical Turk众包市场平台将从2026年7月起停止接受新客户。该平台长期作为人机协同数据标注的核心基础设施,曾支撑无数AI模型的训练数据生产。然而,随着LLM被用于完成标注任务、机器人账号欺诈频发以及AI自身标注能力快速提升,该平台需求结构性萎缩。Amazon的这一决定被业界视为行业风向标事件:传统人类微任务众包模式已经走到了尽头,AI模型训练数据生产正全面转向自动化流水线与合成数据生成。这标志着"人在回路"的数据标注模式正在被AI原生数据生产模式所取代,也意味着数百万全球众包劳动者的生计正在遭遇根本性冲击。

TechCrunch reported on July 6 that Amazon announced its Mechanical Turk crowdsourcing marketplace will stop accepting new customers starting in July 2026. The platform had long served as a core infrastructure for human-in-the-loop data annotation, supporting the training data pipelines of countless AI models. However, with LLMs being used to complete annotation tasks, surging bot fraud, and the rapid improvement of AI's own annotation capabilities, platform demand has structurally shrunk. Amazon's decision is seen as an industry bellwether: the traditional human-microtask crowdsourcing model has reached its end, and training-data production for AI is shifting decisively to automated pipelines and synthetic-data generation. This marks the moment "human-in-the-loop" annotation is being replaced by AI-native data production, with millions of global crowdworkers facing a fundamental livelihood shock.

来源 / Sources


行业数据速览

指标 数据 备注
WAIC 2026展览总面积 10万+ 平方米 首次突破
WAIC 2026参展企业 1,100+ 家 创历史新高
WAIC 2026全球首发产品 300+ 款 智算+具身两大主赛道
WAIC 2026会前意向合作金额 162亿元 已覆盖57个核心场景
WAIC 2026中外嘉宾 1,400+ 位 联动12个国家部委
华为Atlas 950 NPU规模 8,192 卡 灵衢互联
华为Atlas 950 FP8算力 8 EFLOPS 单超节点
中国大模型周调用量 23.45 万亿 Token 环比+15.01%
美国大模型周调用量 4.28 万亿 Token 环比+4.28%
DeepSeek-V4-Flash周调用量 5.34 万亿 Token 连续7周第一
微软AI实施部门投资 25 亿美元 配套6,000员工
美光广岛工厂扩建投资 1.5 万亿日元 重点量产HBM
深圳智平方融资 近50 亿元 估值破200亿元
三星DRAM Q3涨价 20% 连续三季度涨价

媒体人解读

今日的核心信号高度收敛在WAIC 2026这枚"主锚点"上。从政府层面的"两项成果发布",到产业层面的"10万平米+300款首发",再到企业层面的"Atlas 950真机+Agent OS+AI手机三连击",中国AI正在从"模型能力竞赛"过渡到"系统级产业基座"的较量。这种转变的背后,是中国在OpenRouter周调用量上保持十周领跑所代表的真实生产力增长------不是炫参数,而是切实跑出了5.5倍于美国的Token消耗量,这背后是开发者生态、应用场景和Agent渗透率的全面领先。

从国际视角看,三件事勾勒出AI产业正在经历的"三条结构性变化":第一是数据生产范式的人机切换------Amazon关闭Mechanical Turk新客户意味着"人在回路"模式的正式落幕,未来3年合成数据将成为训练主流;第二是硬件供应链的多元化------微软25亿美元自建AI落地能力、谷歌TPU转向英特尔EMIB-T封装、美光HBM扩产,反映出当AI算力变成像电力一样的基础设施时,任何单点依赖都成为风险敞口;第三是中美竞争的"第二战场"------从模型本身移到Token调用量、开源生态、Agent标准化、治理话语权等纵深领域,这是2026年下半年真正值得关注的暗线。

值得提醒的是,WAIC 2026是国家发改委、外交部、工信部、上海市等12个国家部委联合背书的大会,这意味着7月17日---20日将是观察中国AI产业政策方向的"最高规格窗口"。建议读者重点关注三个指标:1)Atlas 950与对标产品的实际性能对比;2)国家发改委《人工智能合作发展行动计划》的8个方面具体落点;3)几十个国家和国际组织高级别代表的实质性表态。这些都将决定未来一年中国AI产业在国际格局中的话语权重。

Today's core signal is highly convergent on the WAIC 2026 "anchor point." From the government's "two outcomes release," through the industry's "100,000 sqm + 300 product debuts," to the enterprise-level "Atlas 950 + Agent OS + AI-phone triple play," Chinese AI is transitioning from "model capability competition" to "system-level industrial substrate" battles. Behind this shift is the real productivity growth represented by Chinese AI's ten-consecutive-week lead on OpenRouter weekly invocation --- not flashy parameters, but a 5.5x higher token consumption than the US, reflecting comprehensive leadership in developer ecosystem, application scenarios, and Agent penetration.

From an international perspective, three events outline three structural shifts the AI industry is undergoing: first, the human-machine switch of the data-production paradigm --- Amazon's closure of new Mechanical Turk customers signals the formal end of the "human-in-the-loop" model, with synthetic data becoming the training mainstream within three years; second, the diversification of hardware supply chains --- Microsoft's US$2.5 billion self-built AI delivery capability, Google's TPU shift to Intel EMIB-T packaging, and Micron's HBM expansion show that when AI compute becomes infrastructure like electricity, any single-point dependency becomes a risk exposure; third, the "second front" of US--China competition --- moving from models themselves to depth-of-field metrics like token invocation, open-source ecosystem, Agent standardization, and governance discourse power, which is the underlying thread worth watching in H2 2026.

It is worth noting that WAIC 2026 is backed by 12 national ministries --- the NDRC, MFA, MIIT, and Shanghai Municipal Government --- which means July 17--20 represents the highest-level window for observing the direction of China's AI industrial policy. Readers should watch three indicators: 1) actual performance comparison between Atlas 950 and its competitors; 2) the eight specific dimensions of the NDRC's AI Cooperation and Development Action Plan; 3) substantive statements from senior representatives of dozens of countries and international organizations. All of these will determine China's AI discourse power in the international landscape over the coming year.


参考来源

英文来源

中文来源


本文内容综合整理自上述公开信息来源,媒体人解读仅代表个人观点,仅供参考。

发布日期:2026年7月7日

📚 相关阅读

相关推荐
ACP广源盛139246256731 小时前
破局 PCIe 4.0 交换瓶颈@ACP#IX8024 / ASM58024参数及应用对比
大数据·人工智能·分布式·嵌入式硬件
ishangy1 小时前
吊箱号AI识别技术:防爆摄像机实现港口集装箱自动核对
人工智能·边缘计算·智慧港口·ai视觉识别·集装箱箱号识别·港口智能化
xiaoxiaoxiaolll1 小时前
《Light: Sci Appl》封面级亮点:可训练误差卷积核如何让光学神经网络实现“逐层自治”?
人工智能
B8017913Y1 小时前
2026年实时语音转写使用场景AI转写清晰整理快,省心又高效
人工智能
sitellla2 小时前
LocalAI:本地运行多模态 AI,它全包了
人工智能·其他
明哥聊AI2 小时前
2026 AI Agent开发完全指南:从MCP协议到多Agent协作系统
人工智能
qiten_0072 小时前
Selective PEFT:BitFit方法深度解析与实践
人工智能·深度学习
秋名山码民2 小时前
拙见科技(陕西)——动态本体驱动的 超图Agentic RAG:拙见AI 面向 AI营销的企业认知系统
人工智能·科技
民乐团扒谱机2 小时前
【微实验】从迷雾中看清世界:卡尔曼滤波的数学诗意与MATLAB实践
人工智能·算法·matlab·卡尔曼滤波·kalman