Global AI Radar
AI Industry Tracker
Capital flows, valuation rankings and sector heat across global AI, ranked by investment relevance. The part of our research radar we make public.
📡 Data updated 2026-10-04 · refreshed daily
Key Moves
Ranked by investment value × recency被称为史上最大IPO的Anthropic传来最新上市进度…
被称为史上最大IPO的Anthropic传来最新上市进度… 手机新浪网
小马智行回应乘客被夹手事件：已配合交警调查，属意外而非交通事故|无人驾驶|新浪科技|就医|车门|持续_手机新浪网
小马智行回应乘客被夹手事件：已配合交警调查，属意外而非交通事故|无人驾驶|新浪科技|就医|车门|持续_手机新浪网 新浪财经
估值2万亿！Anthropic招股文件曝光：420亿亏损、46亿营收、5180亿算力承诺
估值2万亿！Anthropic招股文件曝光：420亿亏损、46亿营收、5180亿算力承诺 搜狐网
开源双语视频学习扩展 Glean 拾句：支持夸克/YouTube/B站，可接入自有大模型 API
开源双语视频学习扩展 Glean 拾句：支持夸克/YouTube/B站，可接入自有大模型 API 80aj.com
Claude Sonnet 5.5发布：Terminal-Bench得分70.6%，超过旗舰Opus 5.5且价格减半
Claude Sonnet 5.5发布：Terminal-Bench得分70.6%，超过旗舰Opus 5.5且价格减半 搜狐网
云深处科技更新IPO招股书：2026年上半年转亏，人形机器人商业化仍处早期
云深处科技更新IPO招股书：2026年上半年转亏，人形机器人商业化仍处早期 虎嗅网
美债收益率飙升致美股IPO市场“熄火”，Anthropic 2万亿美元估值存疑；美联储10月加息概率降至22%；澳大利亚又一政府机构遭OpenAI模型入侵；迪拜航空驾驶舱冲突袭击者身份披露| 一周国际财经
美债收益率飙升致美股IPO市场“熄火”，Anthropic 2万亿美元估值存疑；美联储10月加息概率降至22%；澳大利亚又一政府机构遭OpenAI模型入侵；迪拜航空驾驶舱冲突袭击者身份披露| 一周国际财经 每日经济新闻
NVIDIA DGX Spark 64GB 将于 10 月 23 日上市，售价 4,999 美元，两台设备可通过 ConnectX-7 集群扩展至 128GB。
NVIDIA DGX Spark 64GB 将于 10 月 23 日上市，售价 4,999 美元，两台设备可通过 ConnectX-7 集群扩展至 128GB。 storagereview.com
Mistral 支持方 Headline 完成 4 亿美元 EU VIII 基金募集，面向欧洲 AI 种子轮和 A 轮
Mistral 支持方 Headline 完成 4 亿美元 EU VIII 基金募集，面向欧洲 AI 种子轮和 A 轮 CryptoRank
对标英伟达平台，DeepSeek开源华为昇腾全套基础设施组件
对标英伟达平台，DeepSeek开源华为昇腾全套基础设施组件 21财经
宇树科技上市6周股价腰斩，摩根大通首覆即予“减持”：目标价300元
宇树科技上市6周股价腰斩，摩根大通首覆即予“减持”：目标价300元 环球网
DeepSeek开源昇腾全栈工具，国产AI生态换了底层逻辑
DeepSeek开源昇腾全栈工具，国产AI生态换了底层逻辑 手机网易网
Anthropic招股书，全招了
Anthropic招股书，全招了 hub.baai.ac.cn
Anthropic下一代旗舰Fable 5.5上线真超级智能来了- AI 人工智能
Anthropic下一代旗舰Fable 5.5上线真超级智能来了- AI 人工智能 cnbeta.com.tw
Claude Code Mods 上线，但自定义 UI 拥有完整机器访问权限
Claude Code Mods 上线，但自定义 UI 拥有完整机器访问权限 remio
法官驳回Chegg与Penske针对谷歌AI搜索诉讼
法官驳回Chegg与Penske针对谷歌AI搜索诉讼 QQ News
AMD拟82亿美元收购World Labs，英特尔与英伟达合作AI基础设施
AMD拟82亿美元收购World Labs，英特尔与英伟达合作AI基础设施 虎嗅网
寒武纪：原副总经理梁军变更诉讼请求，索赔金额由42.87亿元调整为278.32亿元
寒武纪：原副总经理梁军变更诉讼请求，索赔金额由42.87亿元调整为278.32亿元 搜狐网
宇树科技上市6周，股价从1100元跌到450.40，摩根大通落井下首次cover就是"减持"评级：目标300元！
宇树科技上市6周，股价从1100元跌到450.40，摩根大通落井下首次cover就是"减持"评级：目标300元！ 财富号
Anthropic上市文件披露：博通将提供最高420亿美元融资
Anthropic上市文件披露：博通将提供最高420亿美元融资 凤凰网科技 Anthropic上市再进一步：被曝冲刺感恩节前2万亿估值IPO，10月14日办投资者日 华尔街见闻 估值2万亿美元有望超越SpaceX！史上最大IPO要来|Anthropic|Claude|大模型|云平台|Google_手机新浪网 新浪财经
阿里：通义千问3.8全能版发布
阿里：通义千问3.8全能版发布 blog.csdn.net
AI语音需求升温：ElevenLabs完成要约出售 估值升至220亿美元
AI语音需求升温：ElevenLabs完成要约出售 估值升至220亿美元 东方财富
马斯克的AI维基百科Grokipedia推出v0.3版本：界面更新，恢复内容编辑
马斯克的AI维基百科Grokipedia推出v0.3版本：界面更新，恢复内容编辑 新浪财经
【SMM快讯】第三代Ray-Ban Meta智能眼镜正式进入印度市场，音频款同步预告上市
【SMM快讯】第三代Ray-Ban Meta智能眼镜正式进入印度市场，音频款同步预告上市 上海有色金属
2.中颖电子定增获批！拟募资不超10亿元加码车规芯片与补充流动资金；
2.中颖电子定增获批！拟募资不超10亿元加码车规芯片与补充流动资金； ijiwei.com
美国法院驳回针对谷歌 AI Overviews 的诉讼——The Verge
美国法院驳回针对谷歌 AI Overviews 的诉讼——The Verge UA.NEWS
“大空头”：为了人类的利益，市场应当大幅下跌，阻止OpenAI和Anthropic IPO
“大空头”：为了人类的利益，市场应当大幅下跌，阻止OpenAI和Anthropic IPO 虎嗅网
“大空头”：为了人类的利益，市场应当大幅下跌，阻止OpenAI和Anthropic IPO
“大空头”：为了人类的利益，市场应当大幅下跌，阻止OpenAI和Anthropic IPO 虎嗅网
Anthropic抢跑OpenAI：最早11月IPO，目标估值达2万亿美元
Anthropic抢跑OpenAI：最早11月IPO，目标估值达2万亿美元 环球网
离职高管向寒武纪索赔278亿元：一纸上市前的约定，如何变成一场豪赌
离职高管向寒武纪索赔278亿元：一纸上市前的约定，如何变成一场豪赌 搜狐网
DeepSeek Harness v0.2 发布桌面版：装插件靠npm 包名，还能让AI 自己写插件- OSCHINA - 开源 × AI · 开发者生态社区
DeepSeek Harness v0.2 发布桌面版：装插件靠npm 包名，还能让AI 自己写插件- OSCHINA - 开源 × AI · 开发者生态社区 OSCHINA
Anthropic宣布Claude for Government正式上线为美国联邦和州政府机构特别打造- AI 人工智能
Anthropic宣布Claude for Government正式上线为美国联邦和州政府机构特别打造- AI 人工智能 cnBeta.COM
Anthropic宣布Claude for Government正式上线为美国联邦和州政府机构特别打造- AI 人工智能
Anthropic宣布Claude for Government正式上线为美国联邦和州政府机构特别打造- AI 人工智能 cnBeta.COM
NVIDIA 被调查
NVIDIA 被调查 timeline.sohu.com
AMD以82亿美元收购李飞飞创立的World Labs，进军世界模型赛道
AMD以82亿美元收购李飞飞创立的World Labs，进军世界模型赛道 虎嗅网
语音AI公司ElevenLabs完成3亿美元员工股份回购 估值220亿美元
语音AI公司ElevenLabs完成3亿美元员工股份回购 估值220亿美元 搜狐网
Anthropic与SpaceX签署最高845亿美元算力协议，可提前90天发通知解除
Anthropic与SpaceX签署最高845亿美元算力协议，可提前90天发通知解除 新浪财经 Anthropic上市再进一步：被曝冲刺感恩节前2万亿估值IPO，10月14日办投资者日 华尔街见闻 一年营收46亿，Anthropic怎么敢花出去5180亿美元？ thepaper.cn
昉擎科技CEO梁军向老东家寒武纪索要赔偿278亿 自身公司估值百亿
昉擎科技CEO梁军向老东家寒武纪索要赔偿278亿 自身公司估值百亿 搜狐网
硬科技早报|美团上线万亿参数模型；晶澳科技被起诉；“人造太阳”新进展
硬科技早报|美团上线万亿参数模型；晶澳科技被起诉；“人造太阳”新进展 观察者网
寒武纪：离职高管诉讼索赔金额调整至278.32亿元 此前已6次败诉
寒武纪：离职高管诉讼索赔金额调整至278.32亿元 此前已6次败诉 东方财富
英伟达洽谈投资Perplexity，企业估值或超300亿美元
英伟达洽谈投资Perplexity，企业估值或超300亿美元 timeline.sohu.com
寒武纪：离职高管诉讼索赔金额调整至278.32亿元，此前已6次败诉
寒武纪：离职高管诉讼索赔金额调整至278.32亿元，此前已6次败诉 thepaper.cn
寒武纪离职高管梁军变更诉讼 索赔金由42亿改为278亿 此前6起案件均败诉
寒武纪离职高管梁军变更诉讼 索赔金由42亿改为278亿 此前6起案件均败诉 驱动之家
ElevenLabs估值翻倍至220亿美元，完成3亿美元员工要约收购
ElevenLabs估值翻倍至220亿美元，完成3亿美元员工要约收购 finance.sina.com.cn
谷歌发布 Gemini 4 Argon，宣称在关键基准测试上领先 OpenAI 与 Anthropic
谷歌发布 Gemini 4 Argon，宣称在关键基准测试上领先 OpenAI 与 Anthropic Yellow.com
受人工智能语音助手需求激增推动，ElevenLabs的估值翻了一番，达到220亿美元
受人工智能语音助手需求激增推动，ElevenLabs的估值翻了一番，达到220亿美元 TradingView
美FTC被曝加大对Anthropic、OpenAI等AI巨头调查力度，官员称不会妨碍AI主导地位
美FTC被曝加大对Anthropic、OpenAI等AI巨头调查力度，官员称不会妨碍AI主导地位 华尔街见闻
美FTC被曝加大对Anthropic、OpenAI等AI巨头调查力度，官员称不会妨碍AI主导地位
美FTC被曝加大对Anthropic、OpenAI等AI巨头调查力度，官员称不会妨碍AI主导地位 华尔街见闻
ElevenLabs估值翻倍至220亿美元，员工可出售最多3亿美元股票| 乌克兰新闻
ElevenLabs估值翻倍至220亿美元，员工可出售最多3亿美元股票| 乌克兰新闻 Межа. Новини України.
刚刚，Tibo承诺「连续28天发布或重置」，OpenAI遭遇信任危机
刚刚，Tibo承诺「连续28天发布或重置」，OpenAI遭遇信任危机 finance.sina.com.cn
离职员工爆料OpenAI风险意识不足
离职员工爆料OpenAI风险意识不足 中国经济网
业界首个：AI音乐制作平台Suno推出Speech语音生成功能|AI音频生成|音频|模型|音乐风格|工作流_手机新浪网
业界首个：AI音乐制作平台Suno推出Speech语音生成功能|AI音频生成|音频|模型|音乐风格|工作流_手机新浪网 新浪财经
Black Forest Labs 7B开源机器人模型登顶英伟达排行榜，成功率42.9%
Black Forest Labs 7B开源机器人模型登顶英伟达排行榜，成功率42.9% 搜狐网
巨头竞速AI4S：微软推出生物学世界模型 Anthropic、OpenAI等也都下场了
巨头竞速AI4S：微软推出生物学世界模型 Anthropic、OpenAI等也都下场了 财联社
巨头竞速AI4S：微软推出生物学世界模型 Anthropic、OpenAI等也都下场了
巨头竞速AI4S：微软推出生物学世界模型 Anthropic、OpenAI等也都下场了 财联社
稀有货！AMD从未发布的锐龙9 5900X3D工程样品现身：128MB缓存+12核
稀有货！AMD从未发布的锐龙9 5900X3D工程样品现身：128MB缓存+12核 电子工程专辑
刚刚，Meta开源Muse Gadgets，让全球开发者自己「造AI外设」
刚刚，Meta开源Muse Gadgets，让全球开发者自己「造AI外设」 投资界
月之暗面Kimi K3上线！系全球最大规模开源模型支持视觉理解、100万token上下文
月之暗面Kimi K3上线！系全球最大规模开源模型支持视觉理解、100万token上下文 content.foshanplus.com
月之暗面Kimi K3上线！系全球最大规模开源模型支持视觉理解、100万token上下文
月之暗面Kimi K3上线！系全球最大规模开源模型支持视觉理解、100万token上下文 content.foshanplus.com
Cursor弃用向量检索转向grep，Kapa发布企业知识检索基准测试
Cursor弃用向量检索转向grep，Kapa发布企业知识检索基准测试 80aj.com
MSI发布配备Nvidia GeForce RTX 5070 12GB和240 Hz OLED显示屏的16英寸游戏笔记本电脑
MSI发布配备Nvidia GeForce RTX 5070 12GB和240 Hz OLED显示屏的16英寸游戏笔记本电脑 Notebookcheck CN
生图AI模型FLUX 3 Image发布：支持4K生成|IT之家|网格|德国|基座|卖点_手机新浪网
生图AI模型FLUX 3 Image发布：支持4K生成|IT之家|网格|德国|基座|卖点_手机新浪网 新浪财经
AI音乐制作平台Suno推出Speech语音生成功能测试版 可同步制作配乐与旁白
AI音乐制作平台Suno推出Speech语音生成功能测试版 可同步制作配乐与旁白 cnBeta.COM
AI音乐制作平台Suno推出Speech语音生成功能测试版可同步制作配乐与旁白- AI 人工智能
AI音乐制作平台Suno推出Speech语音生成功能测试版可同步制作配乐与旁白- AI 人工智能 cnbeta.com.tw
美团发布国庆假期首日出行数据：十大热门目的地出炉
美团发布国庆假期首日出行数据：十大热门目的地出炉 新浪财经
硬科技早报|快手任命可灵CEO；鸿蒙终端破9000万；苹果大举进军智能家居
硬科技早报|快手任命可灵CEO；鸿蒙终端破9000万；苹果大举进军智能家居 观察者网
OpenAI上线“虚拟试穿”功能：ChatGPT能帮你试衣，还能找“明星同款”
OpenAI上线“虚拟试穿”功能：ChatGPT能帮你试衣，还能找“明星同款” i.ifeng.com
波士顿动力为Atlas推出四指机械手
波士顿动力为Atlas推出四指机械手 UA.NEWS
中国开源模型首次进入OpenAI企业客户采购体系
中国开源模型首次进入OpenAI企业客户采购体系 新浪财经
从OpenAI核心高管到死敌 Anthropic创办内幕曝光
从OpenAI核心高管到死敌 Anthropic创办内幕曝光 文学城
Meta Platforms推出Ray-Ban Meta Audio，并在印度发布第三代Ray-Ban Meta智能眼镜
Meta Platforms推出Ray-Ban Meta Audio，并在印度发布第三代Ray-Ban Meta智能眼镜 Moomoo
美网安公司指疑似中国黑客冒充美AI前高官及Anthropic高管以获信息
美网安公司指疑似中国黑客冒充美AI前高官及Anthropic高管以获信息 RFI
惠普战99锐龙版AI移动工作站发布：AMD R9-9955HX3D + RTX PRO 2000，21999元
惠普战99锐龙版AI移动工作站发布：AMD R9-9955HX3D + RTX PRO 2000，21999元 搜狐网
Perplexity 为美国运通卡会员推出商务技能
Perplexity 为美国运通卡会员推出商务技能 Unite.AI
美团发布国庆假期首日出行数据，成都居热门目的地第一位
美团发布国庆假期首日出行数据，成都居热门目的地第一位 凤凰网财经
离职高管向寒武纪索赔278亿
离职高管向寒武纪索赔278亿 搜狐网
传苹果Smart Home Hub10月13日发布 双款设计并深度整合Siri AI
传苹果Smart Home Hub10月13日发布 双款设计并深度整合Siri AI cnBeta.COM
推出22天下载突破500万！Muse“火爆”超过当年ChatGPT
推出22天下载突破500万！Muse“火爆”超过当年ChatGPT 华尔街见闻
小马智行：董事兼高管Jun Peng拟出售300万份ADS，市值约1962万美元
小马智行：董事兼高管Jun Peng拟出售300万份ADS，市值约1962万美元 东方财富
寒武纪离职高管索赔278.32亿元
寒武纪离职高管索赔278.32亿元 凤凰网财经
突发，OpenAI 24小时在线dots正式上岗！GPT悍然推出500美元最贵套餐
突发，OpenAI 24小时在线dots正式上岗！GPT悍然推出500美元最贵套餐 智源社区
Manus携2.0和「Cue」加入战斗！个人Agent混战彻底热闹咯｜附实测
Manus携2.0和「Cue」加入战斗！个人Agent混战彻底热闹咯｜附实测 品玩
DoorDash推出短信订餐AI代理可在Apple Messages中下单- 电子商务
DoorDash推出短信订餐AI代理可在Apple Messages中下单- 电子商务 cnBeta.COM
Meta Muse引爆个人Agent后，OpenAI发布dots正面迎战
Meta Muse引爆个人Agent后，OpenAI发布dots正面迎战 新浪财经
打破安全性壁垒，英矽智能提名新一代小分子GIPR拮抗剂ISM1354
打破安全性壁垒，英矽智能提名新一代小分子GIPR拮抗剂ISM1354 EurekAlert! Science News Releases
前瞻AI药研与长寿科学｜英矽智能10月首周亮相三大国际前沿峰会
前瞻AI药研与长寿科学｜英矽智能10月首周亮相三大国际前沿峰会 EurekAlert! Science News Releases
英矽智能联合创始人兼总裁 Alex Aliper 博士将受邀出席在阿布扎比举办的 AI Everything 峰会（AI Everything Abu Dhabi）。
英矽智能联合创始人兼总裁 Alex Aliper 博士将受邀出席在阿布扎比举办的 AI Everything 峰会（AI Everything Abu Dhabi）。 EurekAlert! Science News Releases
Héctor Sierra (Sony): "En IBC no presentamos productos aislados, sino flujos de trabajo conectados"
Héctor Sierra (Sony): "En IBC no presentamos productos aislados, sino flujos de trabajo conectados" Panorama Audiovisual
美团LongCat-2.5-Preview加入图像理解能力
美团LongCat-2.5-Preview加入图像理解能力 动点科技
快手任命于越为可灵AI CEO 程一笑兼任社区科学线负责人
快手任命于越为可灵AI CEO 程一笑兼任社区科学线负责人 手机新浪网
Abridge Closes $300 Million Series E, Valuation Rises to $5.3 Billion
Abridge, a medical ambient voice-transcription AI company, closed a $300 million Series E on June 25, 2026, lifting its valuation to $5.3 billion, led by a16z with participation from Khosla Ventures and others—roughly doubling its prior valuation. The company reportedly reached about $117 million in annual recurring contract revenue in the first quarter, demonstrating the commercialization velocity of clinical documentation AI and making it one of the most capital-coveted application-layer targets in the space.
NVIDIA 2026 Annual Shareholder Meeting: Vera Rubin in Full Production, Pledge to Return Over 50% of Free Cash Flow to Shareholders
At the June 24 shareholder meeting, Jensen Huang said customers are buying "AI factories" that directly generate revenue, that the new Vera Rubin platform is in full production with orders continuing to pour in, and that the Vera CPU marks NVIDIA's first entry into the general-purpose CPU market. The company pledged to return more than 50% of free cash flow to shareholders, on top of a prior $80 billion buyback authorization and a quarterly dividend raised to $0.25 per share. Its latest fiscal-Q1 2027 revenue reached $81.6 billion, up 85% year over year, with data center up 92%, though continued uncertainty over exports to China remains the key variable.
SpaceX to Acquire Cursor Parent Anysphere for USD 60 Billion in All-Stock Deal
On June 16, SpaceX and Anysphere signed a definitive merger agreement under which SpaceX will acquire Cursor in an all-stock deal at an implied valuation of about USD 60 billion, with closing expected in Q3 2026. Anysphere had just completed a Thrive Capital-led round at roughly USD 90 billion (up more than 3x within the year), and the premium acquisition by a major player marks the AI coding gateway as a strategic-grade asset while reshaping exit expectations for independent AI coding startups.
DeepSeek's First External Round Tops RMB 50 Billion, Post-Money Valuation Rises to RMB 350-400 Billion
On June 15, DeepSeek announced the completion of its first external funding round exceeding RMB 50 billion, setting a new global record for a single AI round, at a post-money valuation of about RMB 350-400 billion that places it in the global top tier. Tencent, CATL, and the national AI industry fund led the round, with founder Liang Wenfeng personally co-investing about RMB 20 billion; the entry of state-backed capital underscores the strategic weight of domestic large models and a self-controlled chip ecosystem. For the primary market, this resets the valuation anchor for leading domestic model companies.
Pulled Just Four Days After Launch: Fable 5 and Mythos 5 Hit by US Export Controls; Anthropic Complies but Disagrees
Anthropic's two most capable models, Fable 5 and Mythos 5, released on June 9, were ordered offline for all foreign users (including foreign nationals in the US and Anthropic's own non-US employees) just days after launch, after Commerce Secretary Lutnick invoked export-control authority on national-security grounds. Anthropic chose to comply but publicly stated its disagreement, arguing that the jailbreak vulnerabilities cited are not universal methods and apply equally to publicly available models such as GPT-5.5. The episode marks the first time frontier AI models have been brought under export controls, signaling that regulatory intervention in large models' dual-use capabilities will become a key variable affecting leading AI firms' commercialization and global availability.
Anthropic Officially Releases Claude Fable 5
On June 9 (US Eastern Time), Anthropic officially launched its next-generation general-purpose large model Claude Fable 5, which it says exceeds all previously public models and sets new benchmark records across software engineering, vision, and scientific research, though priced at double the prior generation. In private testing, payments giant Stripe used it to complete in a single day a migration of about 50 million lines of Ruby code that would otherwise have taken two months. Fable 5 shares a common lineage with the simultaneously released Mythos 5, priced at USD 10 per million input tokens and USD 50 per million output tokens, reflecting how leaps in frontier-model capability are pushing unit prices higher, with implications for compute costs and commercialization pricing.
Claude Mythos 5 Officially Released, Upending the Global Offense-Defense Landscape
Anthropic's Claude Mythos 5, released alongside and sharing a lineage with Fable 5, removes certain safety restrictions and is offered only to trusted professional institutions in fields such as cybersecurity defense and biological research. Anthropic says it has the world's strongest cybersecurity capabilities, excelling at discovering and exploiting software vulnerabilities and able to execute attack-chain steps such as reconnaissance and lateral movement, while in drug design it produced strong candidates for 9 of 14 protein targets at roughly tenfold speed. The model decouples capability from safety for the first time, opening the door to commercializing AI security capabilities, but its dual-use offense-defense nature directly triggered subsequent US export controls, highlighting the tension between frontier-model regulation and commercial deployment.
50 Million Lines of Code Migrated in a Day: Anthropic Unveils Claude Fable 5 and Mythos 5
36Kr reports that Anthropic on June 9 launched Claude Fable 5 and Mythos 5, two models of common lineage differentiated only by software-layer safety restrictions, decoupling capability from safety for the first time. Stripe used them to complete a migration of about 50 million lines of code in a single day, compressing months of engineering into days, while outperforming Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on benchmarks such as GDPval-AA. This demonstrates a productivity leap for AI in enterprise-grade engineering scenarios like large-scale code migration, with implications for revaluing the software outsourcing, IT services, and coding-tools segments.
Suno Closes $400M Series D, Valuation Doubles to $5.4B
AI music leader Suno announced a $400M Series D in early June, led by Bond Capital with participation from IVP, Forerunner, and USV, bringing its post-money valuation to $5.4B—a doubling from its prior round (roughly $2.45B) in just about seven months. Notably, the raise came against the backdrop of ongoing copyright litigation (now covering more than 61,000 tracks), signaling that investors' long-term conviction in the AI music space outweighs near-term legal risk. For the private market, it confirms that generative audio has become one of the fastest-recovering valuation segments within creative AI.
Kling AI Launches Pre-IPO Carve-Out Raise at ~$18-20B Pre-Money; Tencent in Talks
Kuaishou plans to spin off its Kling AI video business for independent fundraising, with this round targeting roughly $2B at a pre-money valuation of $18-20B; Tencent is among the potential investors, with a goal of filing with the Hong Kong Stock Exchange in early 2027. Kling posted Q1 2026 revenue of over RMB 650M, up more than 300% year-over-year, and has raised its ARR guidance to about $500M; if completed, it would become the world's most valuable standalone video-generation model company. This is the largest capital event in the space to date, underscoring China's lead in commercializing video models.
Tencent-Led Chinese Consortium Buys Back Meta's Stake at ~$2B Cost; Manus Returns to Chinese Control
After China's NDRC formally rejected Meta's $2B+ acquisition of Manus parent Butterfly Effect in late April, a Tencent-led Chinese consortium repurchased Meta's stake at the original price, reclaiming this benchmark general-purpose AI Agent. The case is the first publicly blocked AI M&A deal since the 2020 foreign-investment security review mechanism took effect; Manus may subsequently pivot toward a Hong Kong listing, offering early-stage institutions a secondary exit window.
Anthropic Files Confidential S-1; Claude Code Hits $2.5B Annualized Revenue
On June 1, Anthropic confidentially filed its S-1 with the SEC at a valuation of roughly $965B, with annualized revenue reportedly jumping from $14B in February to about $44B in May within three months. Claude Code reached $2.5B in annualized revenue just 14 months after launch, with enterprise customers contributing over half; its ability to handle large codebases and cross-file rewrites has made it a core engine of Anthropic's enterprise revenue.
Unitree Faces June 1 STAR Market IPO Review, Vying to Be A-Shares' First Embodied-AI Listing
The Shanghai Stock Exchange's listing committee is set to review Unitree's IPO application on June 1; the company plans to issue no fewer than 40.4464M shares to raise RMB 4.202B, implying an offering valuation of at least RMB 42B, with founder Wang Xingxing holding 23.82% and controlling nearly 70% of voting rights. With 2025 revenue of RMB 1.708B and non-GAAP net profit of RMB 590M, the company has achieved scaled profitability, and Tencent and Alibaba—which invested a year ago—are sitting on roughly 3x paper gains. If approved, Unitree would become A-shares' first humanoid robotics listing, serving as a benchmark for private-market valuation anchors and exit expectations across the embodied-AI space.
Cognition Raises Over $1B at ~$26B Post-Money Valuation
On May 27, Cognition (parent of AI software engineer Devin) closed over $1B in funding at a $25B pre-money / ~$26B post-money valuation, led by Lux Capital, General Catalyst, and 8VC. This more than doubles its ~$10.2B valuation from September 2025 in just eight months. With annualized revenue of about $492M and customers including Goldman Sachs, Mercedes-Benz, and NASA, the round confirms accelerating enterprise spending on autonomous coding Agents.
Exa Closes $250M Series C at $2.2B Valuation, Led by a16z
AI search infrastructure company Exa announced on May 20 the close of a $250M Series C led by Andreessen Horowitz, bringing its post-money valuation to $2.2B—a threefold jump from $700M last fall. Exa focuses on a search API for AI Agents, with over 5,000 customers (including Cursor, Cognition, and HubSpot). Proceeds will fund training its next-generation model and building infrastructure capable of supporting hundreds of thousands of queries per second, confirming the rapid scaling of the 'search built for AI' niche.
Cerebras Pulls Off 2026's Largest US IPO at ~$67B Market Cap, Holding Major OpenAI Order
The 'NVIDIA challenger' Cerebras debuted on Nasdaq (CBRS) in mid-May at an IPO price of $185, closed up 68% on its first day at a market cap of roughly $67B, and raised about $5.55B including the overallotment—a record for a US semiconductor IPO and the largest US IPO of 2026 so far. Behind it is a multi-year, 750-megawatt compute contract with OpenAI worth over $20B (deployed in batches from 2026-2028). Its latest quarter showed core revenue of $191.3M, up 92% year-over-year and well ahead of expectations, though whether the lofty valuation can be sustained remains to be seen.
Isomorphic Labs Closes $2.1B Series B, the Largest Single Raise in AI Pharma History
DeepMind's AI drug-discovery company Isomorphic Labs closed a $2.1B Series B on May 12, 2026, led by Thrive Capital with follow-on participation from Alphabet and GV, and brought in sovereign and strategic capital including MGX, Temasek, and CapitalG, setting a new global record for AI drug-discovery funding. Proceeds will primarily support preclinical development; the company pushed its first-in-human dosing target to late 2026, underscoring that moving AI-designed drugs from algorithm to clinical validation remains the key milestone for realizing its valuation.
Isomorphic Labs Closes $2.1B Series B, the Largest Single Raise in AI Pharma History
In May 2026, Alphabet's Isomorphic Labs closed a $2.1B Series B led by Thrive Capital, with follow-on participation from Alphabet, GV, Temasek, CapitalG, and MGX, bringing total funding to about $2.7B. The company is concurrently advancing its IsoDDE drug-design engine and partnerships with Eli Lilly and Novartis, and plans to launch its first clinical trials by the end of 2026. The round sets a new private-market record for AI pharma funding and marks a landmark event in the capital concentration of the space.
Sierra Raises $950M, Valuation Climbs to ~$15.8B
Sierra, the customer-service AI Agent company founded by Bret Taylor, announced in early May the close of roughly $950M in new funding led by Tiger Global and Google's GV, with follow-on participation from existing backers Benchmark, Sequoia, and Greenoaks, bringing its post-money valuation to about $15.8B and total funding to over $1B. The company's ARR has surpassed $150M, serving over 40% of the Fortune 50, and it is expanding from customer service toward a broader enterprise Agent platform. This is one of 2026's largest single raises in the enterprise AI Agent space, confirming investor bets on application-layer leaders.
Blackwell Ultra (B300) Enters Mass Production, Delivering 5x Single-GPU Inference
NVIDIA's B300 (GB300 platform), based on the Blackwell Ultra architecture, has officially entered mass production and shipment, positioned as the most powerful single-GPU compute platform currently available, with its companion server systems launching simultaneously. NVIDIA disclosed an order-of-magnitude throughput gain over the prior-generation Hopper in inference scenarios, with major upgrades to per-card memory and bandwidth enabling larger-parameter models on a single card. For the supply chain, B300 volume will drive orders for foundry partners such as Foxconn and Quanta and for the HBM supply chain, further cementing NVIDIA's lead in the AI inference compute market.
Cybercab Enters Production, Reportedly with 3M Orders
On Tesla's Q1 2026 earnings call, Elon Musk confirmed that the steering-wheel-free, pedal-free Cybercab has officially entered production at the Texas Gigafactory in Austin, with exponential capacity ramp expected by year-end; the plant targets annual capacity of 2M-4M vehicles, potentially exceeding the combined output of the Model 3 and Model Y. Positioned as a Robotaxi with native L5 autonomy support, the vehicle is officially priced at no more than $30,000. Caveat: public reporting centers on the production start and capacity targets, while the specific '3M orders' figure has not been confirmed by authoritative sources and should be treated with caution; the real variable for Cybercab lies in regulatory approval for driverless FSD and whether the per-vehicle unit economics can be realized.
Cursor Reaches $30B Valuation, Closes Series E
AI coding tool Cursor (parent company Anysphere) closed a new round of about $2.3B, with its valuation rising nearly threefold in roughly five months to about $29.3B (approaching the $30B mark), led by Coatue and Accel, with new investors including NVIDIA and Google participating. The company's annualized revenue has surpassed $1B, with sales-driven revenue up a hundredfold since the start of the year, and the team has expanded from just over 60 people to more than 300 within a few months. This valuation surge confirms AI coding's breakout from a tool to an enterprise standard, but the controversy over its 'thin wrapper' on underlying models and lack of a moat leaves the high valuation facing ongoing pressure from foundation-model vendors entering the fray.
GPT-5.1 Released, Touted as First to Support 2M-Token Ultra-Long Context
OpenAI released the GPT-5.1 series on November 13, 2025, an upgrade to GPT-5 whose core selling points are being 'warmer and more conversational' along with stronger self-capability awareness and safety guardrails; GPT-5.1 Chat is specifically aligned as a model specialized for everyday chat and lightweight tasks. Important correction: this version's actual positioning is conversational experience and alignment optimization, not 'the first to support 2M-token ultra-long context'—in public materials the 2M-token context corresponds to GPT-6 reporting, while GPT-5.5, released in April 2026, has a 1M-token context. From an investment perspective, OpenAI's iteration cadence is shifting from simply stacking parameters/context toward usability and safety alignment, with model differentiation entering a phase of refined operations.
π0.5 Released: First Commercial General-Purpose Robotics Model
Physical Intelligence released its vision-language-action (VLA) model π0.5 in April 2025, with the core breakthrough being 'open-world generalization'—the ability to perform long-horizon tasks such as cleaning and tidying in real home environments never seen in training data, with cross-scenario transfer significantly improved through co-training on multi-source heterogeneous data. As a landmark advance on the general-purpose robotics 'foundation model' path, π0.5 reinforces the embodied-AI paradigm of 'one model driving multiple bodies,' serving as a bellwether for the commercialization potential and private-market valuation anchors of the humanoid/general-purpose robotics space.
Perplexity Acquires News Assets, Valuation Reaches $18 Billion
AI search company Perplexity closed a new $100 million funding round in July 2025, aligning its valuation with the previously expected $18 billion internal target, roughly double the ~$9 billion mark at the start of the year. To address news copyright disputes, Perplexity simultaneously launched an advertising revenue-sharing mechanism for publishers, seeking to bind itself to media content assets via a 'cite content and share revenue' model. This valuation leap and copyright monetization push reflect capital's high expectations for AI search's challenge to Google, but content-sourcing compliance and the monetization model remain the key risks to whether the valuation can be realized.
Kling 2.2 Launches and Kling AI Surpasses 100 Million Global App Downloads
Kuaishou's Kling AI has scaled rapidly in overseas markets on the back of continuously iterated video-generation models and new features, driving a significant rise in Kuaishou's share price and becoming a flagship case of AI application monetization. Over 80% of Kling's web traffic comes from overseas, with rapidly growing MAU and downloads and steadily climbing ARR, making it one of the most certain bets for a Chinese AI video model monetizing abroad. From an investment perspective, Kling's user and revenue growth curve, alongside market expectations of a potential carve-out for independent fundraising and a higher valuation, makes it a core lens for the re-rating of Kuaishou's AI value.
Nvidia Becomes First Company Ever to Top $5 Trillion Market Cap
Nvidia's market capitalization broke through $5 trillion for the first time in history, making it the first listed company to reach that threshold, and intraday it climbed further to $5.5 trillion, far surpassing tech giants such as Alphabet, Apple, and Microsoft, with a market cap exceeding even Germany's GDP. The company took only about 113 days to go from $4 trillion to $5 trillion, reflecting strong expansion in global AI compute demand and data center capital expenditure. The milestone cements Nvidia's absolute central position in the AI infrastructure value chain, with Jensen Huang noting roughly $500 billion in orders still to be delivered.
Claude Opus 4.7 Launches: 85% on SWE-bench, Setting a New Coding High
Anthropic released Claude Opus 4.7 on April 16, 2026, with markedly improved coding and agentic capabilities, lifting SWE-bench Verified from the prior generation's ~80.8% to 87.6% and clearly leading GPT-5.4 and Gemini 3.1 Pro on autonomous coding and long-horizon workflows. The new model shows large gains in visual understanding (supporting up to 2576px) and real-world GitHub task resolution, while pricing stays consistent with Opus 4.6. This signals that frontier models are moving from demos to coding agents that can run unattended in production, strengthening Anthropic's competitiveness in the enterprise coding agent market.
Figure 03 Unveiled, With Over 100 Units Deployed at BMW Plant
Figure unveiled Figure 03, its third-generation general-purpose humanoid robot, powered by its in-house vision-language-action (VLA) model Helix, with major upgrades in perception, fingertip tactile precision, and mass-production manufacturability, marking humanoid robots' move from the lab toward scaled deployment. Figure 02 had previously been deployed at BMW's Spartanburg plant in the US, helping produce tens of thousands of vehicles and validating practical utility in industrial settings. For the industry, this reinforces the 'physical AI + auto manufacturing' commercialization path, but BMW's robot pilot at its European plants has also brought in other suppliers, reflecting that humanoid mass deployment remains in an early, multi-party competitive stage.
Waymo Operates Robotaxis in 10 US Cities, Surpassing 100 Million Annual Rides
Waymo expanded its Robotaxi commercial operations to a tenth city, recently launching service in Dallas, Houston, San Antonio, Orlando and elsewhere, and plans to expand to overseas cities such as London and Tokyo in 2026. Its fleet of roughly 2,500 vehicles now exceeds 450,000 weekly rides, with about 15 million rides over full-year 2025 and cumulative trips surpassing 20 million. Scaled expansion combined with a new $126 billion high-valuation funding round makes Waymo the most capital-favored bet in autonomous driving, with the competitive focus shifting to scale, unit costs, and global operating efficiency.
AMD Advances Instinct MI400 Series to Challenge NVIDIA Blackwell
AMD is pushing its Instinct MI400 series (including MI455X) built on CDNA 5 with HBM4 memory and TSMC CoWoS-L advanced packaging, landing in 2026. It surpasses NVIDIA's B200/B300 on memory capacity and bandwidth and, via its rack-scale Helios design, competes with Blackwell at the system level while emphasizing price. For investors this signals a loosening AI-chip duopoly, though AMD still faces CUDA's moat, with ROCm software maturity the decisive factor.
DeepSeek Open-Sources V4: 1.6T-Parameter MoE at a Fraction of Rivals' Inference Cost
DeepSeek open-sourced V4, a 1.6-trillion-parameter MoE model with native million-token context, pricing its API at $0.14/M input and $0.28/M output—far below leading Western models. It ships both V4-Pro and a lightweight V4-Flash that beats the prior generation on benchmarks with fewer active parameters. The industry significance is a reshaped cost curve for frontier models, accelerating domestic-compute (e.g., Ascend) adoption and inference monetization while pressuring expensive closed models.
Apple-OpenAI Rumors: Partnership Sours as OpenAI Weighs Lawsuit and Builds Its Own Hardware
The purported 'Apple acquires OpenAI' story did not happen; in reality the partnership soured, with OpenAI weighing a lawsuit over Apple's limited use of its tech and buried feature placement. Meanwhile OpenAI poached Apple hardware executives (including from Vision Pro) and is building its own AI device with Jony Ive, while Apple Intelligence and the Siri rebuild slipped repeatedly with near-zero consumer adoption. For investors the risk is not an acquisition but Apple's stalled AI execution and a talent/narrative shift toward OpenAI.
No 'Sora 3': OpenAI's Flagship Is Sora 2, With 4K and Multi-Shot Consistency Amid Rising Chinese Rivals
There is no official 'Sora 3'; OpenAI's current flagship is Sora 2, emphasizing greater physical realism, controllability and synced audio, and supporting 4K video with multi-shot character consistency. Per 2026 reporting the standalone Sora product was wound down, while the AI-video field tilted toward Chinese tools—Kuaishou's Kling 3.0 topped global leaderboards with several Chinese models near the top. The takeaway: a technical lead in video generation does not guarantee a commercial or ecosystem lead, as first movers are being caught by cost-efficient Chinese supply.
Huawei Ascend 910D Targets NVIDIA H100/H200-Class Single-Card Performance
Huawei's Ascend 910D uses a 5nm, 4-die package with FP8 support and an estimated ~1000T single-card compute (up from ~800T on the 910C), aiming to rival NVIDIA's H100/H200, with volume production expected in Q2–Q3 2026 after customer testing began in mid-2025. Against continued U.S. curbs on high-end AI chips to China, the 910D is a key step toward compute self-sufficiency; investors should watch yield, software-stack (CANN/MindSpore) maturity and large-scale networking rather than peak single-card specs alone.
Google Rolls Out AI Mode Globally, Reshaping Search Traffic
At I/O 2026 Google made AI Mode the core of Search in what it calls the biggest overhaul in 25 years, with AI Mode surpassing 1 billion monthly users, queries doubling each quarter, and personalization expanding to nearly 200 countries and 98 languages. For investors, the shift from blue links to AI-generated answers and agents rewrites ad distribution and SEO, posing structural risk to content and e-commerce ecosystems dependent on search referral traffic.
Tesla Unveils Optimus Gen 3 With Household Skills, Mass Production Ahead
Tesla announced its third-generation humanoid Optimus V3 with a mid-year debut and Q3 mass-production start, featuring fully upgraded hardware, a 22-DOF dexterous hand with sub-millimeter precision, and target household tasks like folding clothes and cleaning; at scale it aims for under $20,000 per unit and up to a million units a year. It marks the inflection point from demo to volume production, driving demand across dexterous-hand, actuator and reducer supply chains.
ByteDance Doubao Tops 50 Trillion Daily Tokens, 300M+ MAU
ByteDance's Doubao model surpassed 50 trillion daily token calls (roughly a 1,000x increase in two years), with the Doubao app at about 345 million monthly active users, as Volcano Engine pivots fully to agent deployment and monetization. The massive call volume validates domestic large-model demand at scale, but heavy consumption against low unit revenue keeps the balance of monetization and compute cost the key question.
Google Releases Gemini 3 With Native Multimodal, Video Understanding
Google released the new-generation Gemini 3, built on a sparse Mixture-of-Experts architecture with native mixed input of text, image, audio, video and PDF, handling up to 900 images per prompt and deep semantic analysis of long video, leading across reasoning, multimodal and coding benchmarks and topping LMArena at 1501. Its native multimodal capability strengthens Google's position in foundation models against OpenAI and leading Chinese labs.
Meta Orion Remains a Developer Prototype, Consumer Launch Slated for 2027
Contrary to rumors of mass production and 500,000 sold-out preorders, Meta Orion remains an internal AR prototype; in 2026 Meta plans only to widen access to a small group of developers with output of about 1,000 units, while a consumer version is expected around 2027 once costs are cut substantially. This pushes back the window for true AR glasses to scale commercially, requiring a recalibration of supply-chain timing expectations.
A compilation of public third-party AI industry information for research only. Not investment advice; does not represent FutureX Capital's holdings or performance.
LIVESTREAM DECKS
What we covered live
2026.09.03
Open Source Breaks the Deadlock: How Chinese AI Is Reshaping the Global Innovation Order
30 pages · 5 free
2026.08.28
Redemption: The Fourth Ending for Open-Source Platforms
18 pages · 5 free
2026.08.23
AI Is Diverging: Which Opportunities Are Breaking Out?
66 pages · 5 free
2026.06.21
The Tide Went Out: Ten Charts from the June 21 Livestream
10 pages · 5 free
Video Library
Qian on AI
From primary markets to frontier models, agents to embodied AI — FutureX founding partner Cynthia Zhang turns each wave of AI conviction into watchable, shareable video.
Follow 倩姐投AI on WeChat Channels for all episodes & livestreams
Podcast
🎙️ Qian on AI · The Podcast
Cynthia Zhang's AI-investing podcast — frontline conviction, Silicon Valley dispatches, founder conversations.
2026.07.28 · 123 min
The Real Opportunities in AI's Second Half: ARR, Open Models, Compute and the Five-Layer Cake
2026.06.29 · 86 min
Tearing Open the AI Traffic Illusion: In-Between SaaS Exits, and the Trillion-Dollar Physical AI Dividend
2026.06.16 · 119 min
Stop Watching Benchmarks: Who's Quietly Making the First Fortune With AI Agents?
2026.06.10 · 127 min
AI, Tariffs, and the Top Gossip We Brought Back From Big Tech
2026.05.29 · 120 min
GTC Deep Rewind: After the Compute Myth, the Make-or-Break Points for AI Startups
2026.05.20 · 110 min
Silicon Valley Line: Are We Ready for AI's Second Wave?
2026.04.28 · 130 min
100B Tokens a Day: Do Better Models Mean Fewer Chances for Apps?
2026.04.27 · 19 min
Capital Ethics and Eastern Wisdom in the AI Era (Part 5)
2026.04.26 · 30 min
Capital Ethics and Eastern Wisdom in the AI Era (Part 4)
2026.04.25 · 30 min
Capital Ethics and Eastern Wisdom in the AI Era (Part 3)
2026.04.24 · 26 min
Capital Ethics and Eastern Wisdom in the AI Era (Part 2)
2026.04.23 · 29 min
Capital Ethics and Eastern Wisdom: Reshaping the Future of Business and Investing
2026.04.14 · 145 min
Quantum Computing × AI: Bubble, or the Start of a New Era?
2026.04.07 · 123 min
Ten Days of OpenClaw Upended Three Years of Silicon Valley Order
2026.04.07 · 149 min
CES 2026: The Quiet Money-Makers Nobody Talks About
2026.03.31 · 91 min
From #1 on ClawHub to Founder in Two Months — With Zhang Haoyang
2026.01.01 · 12 min · ListenHub