文章目录AI科技热点日报 | 2026年07月17日 今日摘要一、WAIC 2026 今日启幕108 款芯片、261 款大模型、300 首发产品集中亮相事件概要来源 / Sources二、月之暗面发布 Kimi K32.8 万亿参数成为全球最大开源模型事件概要来源 / Sources三、DeepSeek 启动 4800 亿元估值第二轮融资筹备 2027 年 IPO事件概要来源 / Sources四、具身智能产业进入量产拐点今年前 5 个月销售收入同比增长 13.9%事件概要来源 / Sources五、谷歌 Gemini 3.5 Pro 再度延期Google DeepMind 内部承压事件概要来源 / Sources六、Meta 自研 AI 芯片 9 月量产AI 云业务引发芯片股 2000 亿美元抛售事件概要来源 / Sources行业数据速览媒体人解读参考来源英文来源中文来源AI科技热点日报 | 2026年07月17日综合整理自 36氪、钛媒体、虎嗅、澎湃新闻、第一财经、The Verge、TechCrunch、MIT Technology Review 等科技财经媒体当日报道聚焦 AI、机器人、芯片、前沿技术、行业金融动态。 今日摘要2026 世界人工智能大会WAIC 2026今天在上海三地四馆同步启幕1100 余家企业参展、300 余款产品全球首发国产芯片与大模型集体亮相。月之暗面正式发布 2.8 万亿参数的 Kimi K3开源模型参数规模再创新高。DeepSeek 启动 4800 亿元估值的第二轮融资并筹备 IPO国产大模型赛道融资潮持续升温。具身智能产业今年前 5 个月销售收入在去年高基数上继续增长人形机器人量产进入关键拐点。国际端谷歌 Gemini 3.5 Pro 因质量未达内部目标再次延期Meta 自研 AI 芯片将于 9 月量产并冲击传统芯片巨头股价。The 2026 World Artificial Intelligence Conference (WAIC 2026) opened today across four venues in Shanghai, with 1,100 exhibitors and 300 global product debuts, showcasing China’s domestic AI chips and foundation models. Moonshot AI released Kimi K3 with 2.8 trillion parameters, setting a new benchmark for open-source models. DeepSeek has kicked off a second funding round at a 480-billion-yuan valuation and is preparing for an IPO. The embodied intelligence industry continues to grow on top of last year’s high base, with humanoid robotics entering a critical mass-production window. Internationally, Google’s Gemini 3.5 Pro was delayed again due to internal quality concerns, while Meta’s in-house AI chip will enter mass production in September, rattling legacy chipmakers.一、WAIC 2026 今日启幕108 款芯片、261 款大模型、300 首发产品集中亮相事件概要2026 世界人工智能大会暨人工智能全球治理高级别会议WAIC 2026于 7 月 17 日在上海世博、张江、西岸三地四馆同步启幕会期四天。本届大会展览总面积首次突破 10 万平方米1100 余家企业参展3000 余项展品集中亮相超 300 款产品将实现全球首发规模创九届以来新高。展品涵盖 108 款芯片、261 款大模型、208 款具身智能产品。9 位图灵奖得主、6 位诺贝尔奖得主将出席并做学术分享。中兴通讯旗下努比亚将在本次大会首发全球首款 AI 智能体手机华为、阿里、百度、商汤、阶跃星辰、星动纪元等头部厂商均布展核心产品线。本届大会的焦点从模型能力转向产业落地超节点算力、人形机器人、AI 智能体被列为三大主线。WAIC 2026 kicked off today at four venues across Shanghai’s Expo, Zhangjiang, and West Bund districts, running through July 20. Exhibition space exceeded 100,000 square meters for the first time, with 1,100 exhibitors and 3,000 exhibits on display, of which over 300 products will make their global debut—the largest edition in the conference’s nine-year history. The showcase covers 108 AI chips, 261 foundation models, and 208 embodied-intelligence products. Nine Turing Award laureates and six Nobel laureates will deliver talks. ZTE’s Nubia will debut the world’s first AI-agent smartphone at the conference; Huawei, Alibaba, Baidu, SenseTime, Stepfun, and Robotera will also showcase their flagship product lines. The focus this year has shifted from pure model capability to industrial deployment, with super-node compute, humanoid robots, and AI agents as the three main tracks.来源 / Sources36氪108款芯片、261款大模型亮相9位图灵奖得主开讲丨WAIC26全剧透钛媒体WAIC 2026即将开幕这届AI亮点有哪些新浪财经2026世界人工智能大会今日启幕 300余款首发产品展现AI产业全景东方财富2026世界人工智能大会治理、学术、产业多线并进二、月之暗面发布 Kimi K32.8 万亿参数成为全球最大开源模型事件概要2026 年 7 月 17 日凌晨月之暗面Moonshot AI正式宣布推出迄今能力最强的大模型 Kimi K3。该模型总参数达 2.8 万亿是目前全球参数规模最大的开源模型也是首个突破 2 万亿级别的开源模型原生支持图像与视频理解提供最高 100 万 Token 上下文窗口采用自研 KDAKimi Delta Attention混合线性注意力架构。模型面向软件工程、知识工作、深度研究、多模态理解等复杂任务场景进行深度优化。在官方综合测评中Kimi K3 在参与对比的模型里仅次于 Anthropic Claude 系列旗舰版前端编程能力超过 Fable 5被业内称为达到Fable 级体感。Kimi API 开放平台已同步上线 K3 模型的 API 接口支持 Tool Calling 与多模态推理。Moonshot AI officially unveiled Kimi K3 in the early hours of July 17, 2026, the company’s most capable large model to date. With 2.8 trillion total parameters, K3 is the largest open-source model in the world and the first open-source model to surpass the 2-trillion-parameter threshold. It natively supports image and video understanding, provides up to a 1-million-token context window, and is built on Moonshot’s proprietary KDA (Kimi Delta Attention) hybrid linear attention architecture. The model is optimized for software engineering, knowledge work, deep research, and multimodal understanding. In Moonshot’s official benchmarks, K3 ranks just behind Anthropic’s flagship Claude lineup, with front-end coding ability surpassing Fable 5. The Kimi API platform has simultaneously launched K3 API endpoints with Tool Calling and multimodal reasoning support.来源 / Sources新浪财经2.8万亿月之暗面发布最新开源模型华尔街见闻全球第三强“Fable级体感”2.8万亿参数Kimi K3刷新开源记录品玩月之暗面发布开源模型Kimi K3每经网月之暗面发布Kimi K32.8万亿参数全球最大开源模型AIHub月之暗面发布Kimi K32.8万亿参数支持百万Token上下文三、DeepSeek 启动 4800 亿元估值第二轮融资筹备 2027 年 IPO事件概要DeepSeek 近期启动第二轮融资目标估值推高至 4800 亿元人民币约 710 亿美元较 6 月中旬刚刚完成的首轮 500 亿元约 70 亿美元投后估值约 520 亿美元融资上涨近 4 成。彭博社援引知情人士消息称DeepSeek 已启动 IPO 准备工作最快将于 2026 年内提交上市申请目标瞄准境内资本市场争取 2027 年完成上市。开润股份 7 月 16 日披露的公告显示其间接投资 DeepSeek 的首轮融资对应估值约 3510 亿元新一轮融资的投资方阵容仍在洽谈中。DeepSeek V4 系列以极低定价席卷企业级 API 市场被业内视为推动本轮估值上调的核心产品力支撑。DeepSeek has kicked off a second funding round at a target valuation of 480 billion yuan (about US$71 billion), up nearly 40% from the first round closed in mid-June at roughly 50 billion yuan (US$7 billion raised, US$52 billion post-money valuation). Bloomberg reported that DeepSeek has begun IPO preparation and could file as early as 2026, aiming to list in the domestic market in 2027. According to a disclosure by Kairun Co. on July 16, its indirect investment in DeepSeek’s first round values the company at around 351 billion yuan. Investor lineup for the new round is still being finalized. DeepSeek V4 series, with its aggressive API pricing, is widely seen as the product force underpinning the valuation jump.来源 / Sources虎嗅DeepSeek首轮融资完成估值3510亿澎湃新闻DeepSeek都要IPO了豆包还不分拆吗OFweekDeepSeek或冲刺A股IPO4800亿估值引行业热议新浪财经DeepSeek冲刺IPOAI四小强的估值都要重算了四、具身智能产业进入量产拐点今年前 5 个月销售收入同比增长 13.9%事件概要国家税务总局最新数据显示2026 年前 5 个月具身智能产业企业销售收入在 2025 年全年增长 13.9% 的高基数上继续增长。IT 桔子统计显示2026 年上半年国内具身智能赛道融资总金额已突破 935 亿元较 2025 年上半年提升 5 倍融资事件达 322 笔。工业和信息化部此前披露今年我国人形机器人整机产量有望突破 10 万台。长三角地区集聚了智元机器人、傅利叶智能、开普勒、宇树科技、云深处、海康机器人等头部整机企业企业数量与融资规模占全国 50% 以上。WAIC 2026 现场专门设立 6000 平方米的具身智能孵化与投资专区宇树科技、银河通用、星动纪元等独角兽集中亮相。本月长三角具身智能企业 IPO 进程加速宇树科技科创板 IPO 已于 6 月 1 日过会拟募资 42.02 亿元。Tax authority data shows that in the first five months of 2026, revenue from China’s embodied-intelligence industry continued to grow on top of last year’s 13.9% full-year increase. IT Juzi’s tally shows the domestic embodied-AI sector raised more than 93.5 billion yuan in H1 2026, five times the same period last year, across 322 deals. The Ministry of Industry and Information Technology previously stated that China’s full-year humanoid-robot output is expected to exceed 100,000 units in 2026. The Yangtze River Delta region hosts leading players such as Agibot, Fourier Intelligence, Kepler, Unitree, DeepRobotics, and Hikvision Robotics, accounting for over 50% of national industry count and financing. WAIC 2026 has set up a dedicated 6,000-square-meter embodied-intelligence incubation and investment zone. Unitree’s STAR Market IPO passed review on June 1, seeking to raise 4.202 billion yuan.来源 / Sources钛媒体2026具身智能产业销售收入数据深度解读澎湃新闻报告长三角具身智能企业数量、融资规模均占全国五成以上钛媒体明星创业者扎堆2026具身智能创业门槛有多高网易2026中国机器人融资全景54家估值曝光23家破百亿五、谷歌 Gemini 3.5 Pro 再度延期Google DeepMind 内部承压事件概要原计划于 7 月 17 日发布的谷歌 Gemini 3.5 Pro 已比原计划推迟数月。据 10 位现任及前员工透露谷歌 5 月 19 日发布 Gemini 3.5 系列时先开放了 3.5 Flash3.5 Pro 当时已在文章中预告将很快发布。但内部测试显示编程能力与多模态推理未达到内部目标团队决定延期。Gemini 3.5 Pro 此前已从 2026 年 6 月延期至 7 月 17 日本次将再次推迟。竞争对手 OpenAI 与 Anthropic 持续推出能力超越 Gemini 的模型过去半年 AI 圈风头被这两家轮流占据谷歌相对沉寂。近期 Google Gemini 联合负责人 Noam Shazeer 宣布离职加入 OpenAI进一步加大谷歌在 AI 领域的人才流失压力。Google’s Gemini 3.5 Pro, originally scheduled for release on July 17, has been delayed by several months. Ten current and former employees told reporters that when Google launched the Gemini 3.5 family on May 19, it shipped 3.5 Flash first while previewing 3.5 Pro as “coming soon.” Internal testing found that coding capability and multimodal reasoning fell short of internal targets, prompting the team to postpone. Gemini 3.5 Pro had previously been pushed from June to July 17, and will now be delayed again. Competitors OpenAI and Anthropic continue to ship models surpassing Gemini, and Google has been relatively quiet for the past six months. The recent departure of Gemini co-lead Noam Shazeer to OpenAI has intensified talent-loss pressure on Google.来源 / Sources网易Gemini 3.5 Pro延期原定6月推出编程能力未达内部目标网易谷歌Gemini发布延期技术表现未达目标Exchange4mediaGoogle Gemini co-lead Noam Shazeer to join OpenAI六、Meta 自研 AI 芯片 9 月量产AI 云业务引发芯片股 2000 亿美元抛售事件概要据路透社独家披露的内部备忘录Meta Platforms 计划于 2026 年 9 月开始量产自研 AI 芯片目标到 2027 年将整体算力翻倍至 14 吉瓦GW。该芯片由 Meta 自研团队设计将与英伟达 GPU 形成内部补充减少对外部 GPU 单一供应商的依赖。Meta 同步推进的AI 算力对外销售业务引发市场震动——英伟达、AMD、英特尔三家芯片巨头股价短线急跌整个芯片板块一日蒸发约 2000 亿美元市值。华尔街分析师指出Meta 一旦成为算力批发商将直接挑战 AWS、Azure、GCP 的云服务格局并对英伟达GPU 独占地位构成实质性冲击。According to an internal memo seen by Reuters, Meta Platforms plans to start mass-producing its in-house AI chip in September 2026, with a goal of doubling total compute capacity to 14 gigawatts by 2027. The chip is designed by Meta’s internal silicon team and will complement NVIDIA GPUs, reducing dependence on a single external supplier. Meta’s parallel push into selling AI compute externally has triggered a market shock—NVIDIA, AMD, and Intel all sold off sharply, wiping out roughly US$200 billion in market cap across the chip sector in a single session. Wall Street analysts note that once Meta becomes a wholesale compute provider, it will directly challenge AWS, Azure, and GCP, and pose a real threat to NVIDIA’s GPU dominance.来源 / SourcesReuters via MSNExclusive-Meta to put AI chip into production in SeptemberMemeburnMeta AI Cloud Push Triggers the Biggest Chip Stocks SelloffDeepTech深科技Meta放弃芯片自研扎克伯克与英伟达签署史诗级协议行业数据速览指标数据备注WAIC 2026 展览面积10 万 平方米历届最大WAIC 2026 参展企业1,100 家创历史新高WAIC 2026 全球首发产品300 款芯片、大模型、机器人、智能体Kimi K3 参数规模2.8 万亿全球最大开源模型Kimi K3 上下文窗口100 万 TokenKDA 混合线性注意力DeepSeek 二轮估值4,800 亿元 / ~710 亿美元较首轮上涨约 40%DeepSeek 计划上市时间2027 年境内资本市场国内具身智能 H1 融资935 亿元同比 5 倍人形机器人 2026 产量预期10 万 台工信部披露芯片板块单日市值蒸发~2,000 亿美元Meta 云冲击Meta 自研芯片量产时间2026 年 9 月目标 14 GW 算力国家大基金三期首次入局具身智能银河通用国家 AI 产业基金媒体人解读今天的 AI 头条几乎全部来自 WAIC 2026——但真正值得关注的不是这场大会本身而是它所映射出的产业转向从模型能力到产业落地。1100 家展商、300 款首发、108 款芯片同台竞技意味着国产 AI 产业链已经形成芯片—模型—智能体—终端的完整拼图。Kimi K3 的 2.8 万亿参数和 DeepSeek 4800 亿估值同时出现在同一天是国产大模型进入参数军备 资本竞速双高阶段的标志性事件。但要注意参数规模和估值并不是护身符能否在企业级市场跑出 ARR、能否在 C 端做出真正的高频场景才是下半年决出胜负的关键。国际端今天最大的反差信号来自两家美国巨头Meta 自研芯片量产 算力外销直接打掉了芯片板块 2000 亿美元市值Google Gemini 3.5 Pro 再度延期意味着三足鼎立格局下谷歌暂时掉队。对中国 AI 厂商而言这既是压力也是窗口期——北美算力供给侧出现裂痕给国产芯片与模型出海留下了至少 12 个月的红利期。Almost every headline today orbits WAIC 2026—but the real story isn’t the conference itself, it’s the industry pivot it reflects:from model capability to industrial deployment. With 1,100 exhibitors, 300 global debuts, and 108 AI chips on the same floor, China’s domestic AI stack has assembled a complete “chip–model–agent–device” picture. Kimi K3’s 2.8 trillion parameters and DeepSeek’s 480-billion-yuan valuation landing on the same day mark the entry of Chinese foundation models into a “parameter arms race capital sprint” phase. But parameter size and valuation are not armor—what matters for H2 is whether these companies can generate real ARR in the enterprise market and build genuinely high-frequency consumer scenarios.Internationally, the most striking divergence came from two US giants: Meta’s in-house chip mass production plus external compute sales wiped out US$200 billion in chip-sector market cap in one session; Google’s Gemini 3.5 Pro was delayed again, meaning Google has temporarily fallen behind in the three-way race.For Chinese AI vendors, this is both pressure and a window—a crack in the North American compute supply side opens at least 12 months of dividend for Chinese chips and models going overseas.参考来源英文来源AI Weekly: AI News Today, July 17Reuters via MSN: Meta to put AI chip into production in SeptemberMemeburn: Meta AI Cloud Push Triggers the Biggest Chip Stocks SelloffExchange4media: Google Gemini co-lead Noam Shazeer to join OpenAIMERICS: Embodied AI: China’s ambitious path to transform its robotics industry中文来源36氪108款芯片、261款大模型亮相9位图灵奖得主开讲钛媒体WAIC 2026即将开幕这届AI亮点有哪些新浪财经2026世界人工智能大会今日启幕 300余款首发产品展现AI产业全景东方财富2026世界人工智能大会治理、学术、产业多线并进新浪财经2.8万亿月之暗面发布最新开源模型华尔街见闻2.8万亿参数Kimi K3刷新开源记录品玩月之暗面发布开源模型Kimi K3每经网月之暗面发布Kimi K32.8万亿参数全球最大开源模型AIHub月之暗面发布Kimi K32.8万亿参数支持百万Token上下文虎嗅DeepSeek首轮融资完成估值3510亿澎湃新闻DeepSeek都要IPO了豆包还不分拆吗OFweekDeepSeek或冲刺A股IPO4800亿估值引行业热议新浪财经DeepSeek冲刺IPOAI四小强的估值都要重算了钛媒体2026具身智能产业销售收入数据深度解读澎湃新闻报告长三角具身智能企业数量、融资规模均占全国五成以上钛媒体明星创业者扎堆2026具身智能创业门槛有多高网易2026中国机器人融资全景54家估值曝光23家破百亿网易Gemini 3.5 Pro延期原定6月推出编程能力未达内部目标网易谷歌Gemini发布延期技术表现未达目标DeepTech深科技Meta放弃芯片自研扎克伯克与英伟达签署史诗级协议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。作者署名刘一说 ️发布日期2026年07月17日