AI科技热点日报 | 2026年08月10日

本文基于 The Verge、TechCrunch、MIT Technology Review、界面新闻、36氪、钛媒体、财联社等国内外科技财经媒体的公开报道，对当日 AI 领域热点事件进行梳理与解读。

📌 今日摘要

今日 AI 行业聚焦六大热点：宇树科技科创板申购正式开启，"人形机器人第一股"以 150.80 元发行价登陆资本市场，发行市值 609.93 亿元；寒武纪、摩尔线程同日发布半年报，国产 AI 芯片进入业绩兑现期；OpenRouter 平台 DeepSeek-V4-Flash 登顶周榜，中国大模型周调用量连续 15 周超越美国；L3/L4 自动驾驶强制国标 GB 44721—2026 正式落地，2027 年 7 月 1 日实施；特斯拉 AI 芯片工厂 Terafab 在得州落成，年产能规划达 1 太瓦算力；首个全国产 10 万卡 AI 超集群在郑州投入运行。

Today’s AI industry spotlight falls on six stories: Unitree Robotics officially opens its STAR Market subscription at RMB 150.80 per share with a post-issuance market cap of RMB 60.99 billion; Cambricon and Moore Threads released H1 reports simultaneously, marking domestic AI chip makers’ earnings season; DeepSeek-V4-Flash tops the OpenRouter weekly token ranking, with Chinese large models surpassing US peers in token calls for 15 consecutive weeks; China’s mandatory L3/L4 autonomous-driving standard GB 44721-2026 was officially released, taking effect on July 1, 2027; Tesla’s AI chip fab Terafab broke ground in Texas, targeting 1 terawatt of annual compute output; the first fully-domestic 100,000-card AI supercluster entered operation in Zhengzhou.

一、宇树科技科创板申购正式开启：发行价 150.80 元，募资 60.99 亿元

事件概要

8月10日，被市场称为"A股人形机器人第一股"的宇树科技在科创板同步开启网上、网下申购。公告显示，本次发行价格为 150.80 元/股，拟公开发行 4044.6434 万股，占发行后总股本 10%，预计募资总额 60.99 亿元，发行后总市值约 609.93 亿元。网上申购代码为 787836，证券代码为 688836，缴款日为 8 月 12 日。市场分析预计中签率在 0.02%–0.03% 之间，每中一签需缴款约 7.54 万元。本次发行从 3 月 20 日 IPO 申请获受理到 8 月 10 日申购，用时不到 5 个月，创下年内最快 IPO 审核纪录。募集到的资金将投入智能机器人模型研发、机器人本体研发、新产品开发和智能机器人制造基地建设四个项目。

On August 10, Unitree Robotics, often dubbed the “first humanoid robot stock” on China’s A-share market, launched its STAR Market subscription both online and offline. The issuance price is set at RMB 150.80 per share, with 40.4464 million new shares — 10% of post-issuance total — for an aggregate raise of RMB 6.099 billion and a post-issuance market cap of roughly RMB 60.993 billion. The online subscription code is 787836 and the ticker is 688836, with payment due on August 12. Analysts project a 0.02%-0.03% allotment rate, with each winning lot requiring about RMB 75,400 in subscription funds. From IPO acceptance on March 20 to subscription on August 10, the process took less than five months — the fastest IPO review timeline of the year. The proceeds will fund four projects: intelligent robot model R&D, robot body R&D, new product development, and an intelligent-robot manufacturing base.

来源 / Sources

• 财联社（东方财富网转载）
• 钛媒体

二、国产 AI 芯片业绩兑现：寒武纪、摩尔线程同日发布半年报

事件概要

国产 AI 芯片龙头寒武纪（688256.SH）8 月 7 日晚发布 2026 年半年报，上半年实现营业收入 59.96 亿元，同比增长 108.13%；归母净利润 23.11 亿元，同比增长 122.61%；扣非归母净利润 21.66 亿元，同比增长 137.30%。公司新一代旗舰云端训练芯片思元 690 于 2026 年初正式量产，采用双 die 封装设计，FP16 算力超 700 TFLOPS，配备 196GB HBM3 高带宽内存，已进入互联网大厂、政务云等客户的批量部署阶段。同日，国产 GPU 企业摩尔线程发布 2026 年半年报，上半年营收 17.36 亿元，同比增长 147.42%，已超 2025 年全年营收；归母净利润亏损 1156 万元，亏损同比收窄 95.73%；毛利总额 9.89 亿元，整体毛利率达 56.95%；公司同步公告正在筹划港股上市。

Cambricon (688256.SH) released its H1 2026 report on the evening of August 7. Revenue reached RMB 5.996 billion, up 108.13% YoY; net profit attributable to shareholders was RMB 2.311 billion, up 122.61% YoY; non-GAAP net profit grew 137.30% YoY to RMB 2.166 billion. The company’s new flagship cloud training chip Siyuan 690 entered mass production in early 2026, adopting a dual-die package with FP16 throughput exceeding 700 TFLOPS and 196GB HBM3 high-bandwidth memory, and is now in batch deployment at major internet companies and government clouds. On the same day, domestic GPU maker Moore Threads reported H1 revenue of RMB 1.736 billion, up 147.42% YoY — already exceeding full-year 2025 revenue — with net loss narrowing 95.73% YoY to RMB 11.56 million; gross profit reached RMB 989 million with a 56.95% gross margin, and the company disclosed plans for a Hong Kong listing.

来源 / Sources

• 财联社
• 财联社
• 澎湃新闻

三、OpenRouter 周榜：DeepSeek-V4-Flash 登顶，中国大模型周调用量连续 15 周超越美国

事件概要

根据 OpenRouter 平台最新数据测算，上周（8 月 3 日至 8 月 9 日）全球 AI 大模型总调用量为 69 万亿 Token，环比增长 21.48%。中国 AI 大模型周调用量达 34.25 万亿 Token，环比增长 21.76%；同期美国 AI 大模型周调用量为 9.17 万亿 Token，环比增长 109.36%。中国大模型周调用量连续 15 周超越美国，稳居全球首位。全球调用量排名前四均为中国 AI 大模型：DeepSeek-V4-Flash-0731 正式版以 8.83 万亿 Token 居首，环比增长 570%；腾讯 Hy3 以 8.05 万亿 Token 排第二；DeepSeek-V4-Flash-0423 预览版排第三；小米 MiMo V2.5、智谱 GLM-5.2、DeepSeek-V4-Pro 分别位列四至六位。

According to OpenRouter’s latest weekly data (Aug 3–9), global large-model token calls reached 69 trillion, up 21.48% week-over-week. Chinese large models accounted for 34.25 trillion tokens, up 21.76%, while US models logged 9.17 trillion, up 109.36% — China has now surpassed the US for 15 consecutive weeks. The top four by call volume are all Chinese: DeepSeek-V4-Flash-0731 led with 8.83 trillion tokens (up 570% WoW), followed by Tencent Hy3 at 8.05 trillion, DeepSeek-V4-Flash-0423 in third, then Xiaomi MiMo V2.5, Zhipu GLM-5.2 and DeepSeek-V4-Pro rounding out positions four through six.

来源 / Sources

• 财联社
• 钛媒体

四、自动驾驶新国标落地：GB 44721—2026 强制要求发布，2027 年 7 月 1 日实施

事件概要

工业和信息化部组织制定并归口的《智能网联汽车 自动驾驶系统安全要求》（GB 44721—2026）强制性国家标准近日由国家市场监督管理总局、国家标准化管理委员会批准发布，将于 2027 年 7 月 1 日正式实施。这是我国首部针对 L3、L4 级自动驾驶的强制性国家标准。标准规范对象为搭载 L3 级（有条件自动驾驶）和 L4 级（高度自动驾驶）系统的 M 类（载客）和 N 类（载货）车辆，自动泊车系统不在适用范围内。标准要求涉及功能的新车必须达标，否则无法销售，旨在建立自动驾驶全生命周期安全规范，对事故责任认定、夸大宣传、强制 AEB 等给出统一标尺。

The mandatory national standard “Intelligent Connected Vehicles — Automated Driving System Safety Requirements” (GB 44721-2026), developed and stewarded by the Ministry of Industry and Information Technology, has been approved by the State Administration for Market Regulation and the Standardization Administration of China, and will take effect on July 1, 2027. This is China’s first mandatory national standard targeting L3 and L4 autonomous driving. It applies to M-class (passenger) and N-class (cargo) vehicles equipped with L3 (conditional automation) or L4 (high automation) systems — automated parking is excluded. Vehicles with covered functions must meet the standard before they can be sold, providing a single reference point across accident attribution, marketing claims, and mandatory AEB requirements.

来源 / Sources

• 财联社
• 澎湃新闻
• 界面新闻

五、特斯拉 AI 芯片工厂 Terafab 落成得州，年产能规划超 1 太瓦算力

事件概要

8月10日，特斯拉官方宣布史上最大 AI 芯片工厂 Terafab 将在美国得州格莱姆斯县落成，今年 4 月已在得州超级工厂北园区的芯片研发基地破土动工。Terafab 目标每年生产超过 1 太瓦算力的 AI 芯片，远超当前全球产能。特斯拉方面表示，人工智能将来的发展需要远超当前全球产能的芯片供应，Terafab 是为此而生的下一代制造设施。这一布局与特斯拉自动驾驶 FSD 芯片、Dojo 训练集群及 Optimus 人形机器人的算力底座直接关联。

On August 10, Tesla officially announced that Terafab — its largest AI chip fab to date — will be located in Grimes County, Texas; site preparation at the chip R&D base in the northern park of the Texas Gigafactory commenced in April. Terafab targets more than 1 terawatt of annual AI-chip output, far exceeding today’s global capacity. Tesla noted that future AI development will require chip supply far beyond current global output, and Terafab was built for that purpose. The fab directly supports the compute backbone of Tesla’s FSD chip, the Dojo training cluster, and the Optimus humanoid robot program.

来源 / Sources

• 钛媒体
• 财联社

六、消费级具身智能大模型公司诺因智能完成 5 亿元天使++轮融资

事件概要

8月10日，消费级具身智能大模型公司诺因智能宣布完成单笔 5 亿元人民币天使++轮融资。本轮融资由经纬创投领投，红点创投、商汤国香资本、华登科技、L2F 光源创业者基金跟投，光源资本担任独家财务顾问。本轮融资后，诺因将继续投入 GLOW 生成式具身大模型、家庭任务泛化以及数据与计算基础设施建设，并推进 KNOWIN-X1 的产品迭代、工程验证和量产准备。这是近期具身智能赛道单笔规模最大的早期融资之一，反映出资本市场对"具身大脑 + 通用本体"路径的持续押注。

On August 10, Nuoyin Intelligence, a consumer-grade embodied AI large-model company, announced the close of a RMB 500 million angel++ round — among the largest early-stage checks in the embodied-AI track recently. Matrix Partners China led the round, with Redpoint China, SenseTime Guoxiang Capital, HuaDing Investment and L2F Light Source Entrepreneur Fund co-investing; Light Source Capital served as exclusive financial advisor. Proceeds will fund the GLOW generative embodied model, household task generalization, data and compute infrastructure, and product iteration, engineering validation and mass-production preparation for the KNOWIN-X1. The deal reflects continued capital conviction in the “embodied brain + general-purpose body” architecture.

来源 / Sources

• 钛媒体
• 投中网

七、智能驾驶量产提速：千里智驾 ASD 累计装车突破 53 万辆

事件概要

8月8日，千里智驾宣布搭载千里智驾 ASD 4.0 辅助驾驶方案的 smart 精灵 #1 量产上市，这是千里智驾 ASD 第二次在 smart 车型上量产搭载，实现跨车型快速复用。自 2026 年 1 月首发以来，千里智驾 ASD 已实现 30 多万辆量产搭载，从 ASD 3.0 迭代到 ASD 4.0 仅用时 3 个月。截至目前，千里智驾 ASD 累计装车量突破 53 万辆，覆盖极氪、领克、银河、smart 等 19 款车型，辅助驾驶激活率达 93.8%。同日，东风风行旗下星海 V6 举行全国预售发布会，搭载华为乾崑智驾系统，配备 25 个高灵敏度传感器与华为自研 MDC 计算平台，可实现 360 度环境感知，华为智驾方案下探至 10 万级大六座市场。

On August 8, Qianli Intelligent Driving announced that the smart #1 equipped with its ASD 4.0 driver-assistance system had entered mass production — its second smart-branded model deployment and a cross-model reuse milestone. Since debut in January 2026, Qianli ASD has shipped on over 300,000 vehicles, with the upgrade from ASD 3.0 to ASD 4.0 completed in just three months. Cumulative installations now exceed 530,000 vehicles across 19 models including Zeekr, Lynk & Co, Geely Galaxy and smart, with a 93.8% driver-assist activation rate. The same day, Dongfeng Forthing’s Xinghai V6 — equipped with Huawei’s Qiankun ADS, 25 high-sensitivity sensors and Huawei’s in-house MDC compute platform — kicked off its national pre-sale, pushing Huawei’s smart-driving stack into the RMB 100,000-class large six-seat segment.

来源 / Sources

• 财联社
• 财联社

八、基础设施层：首个全国产 10 万卡 AI 超集群投用，AWS 内部 CPU 也告急

事件概要

基础设施层面，今日有两条产业级信号同步出现。其一，在国家超算互联网郑州核心节点，首个全国产 10 万卡人工智能超集群正式投入运行，科学计算加智能计算的融合算力为未来可能出现的新型计算需求做了前瞻布局，全国多个算力节点迎来新一轮算力扩容，标志我国算力基础设施建设迈入 10 万卡级部署新阶段。其二，亚马逊云（AWS）内部今年 5 月已向工程团队发出警示，要求全力节省 CPU 计算资源——过去申请 CPU 服务器几小时内即可到位，如今需等待数天。有评论指出，AI 推理需求正从 GPU 蔓延至 CPU，云计算商业模式由"卖闲置空间"转向"卖稀缺资源"。

Two industry-scale infrastructure signals surfaced today. First, at the National Supercomputing Internet’s Zhengzhou core node, China’s first fully-domestic 100,000-card AI supercluster formally entered operation, combining scientific computing and intelligent computing to plan ahead for emerging workload patterns — multiple compute nodes are entering a new capacity-expansion cycle, marking China’s compute infrastructure move into the 100,000-card deployment era. Second, Amazon Web Services internally alerted its engineering teams back in May to aggressively conserve CPU resources — provisioning that once took a few hours now takes several days, an unprecedented delay in engineers’ careers. Analysts note AI inference demand is spreading from GPU to CPU, shifting cloud’s commercial model from selling idle capacity to selling scarce resources.

来源 / Sources

• 财联社
• 快科技

九、AI 终端与算力硬件：联想 P16v AI 工作站开售，单卡 AI 算力达 572 TOPS

事件概要

8月10日，轻薄移动图形 AI 工作站 Lenovo P16v 正式开售，搭载英特尔酷睿 Ultra 7 251HX 处理器与 NVIDIA RTX PRO 2000 Blackwell 架构 GPU，整机性能释放最高可达 200W，AI 算力最高可达 572 TOPS，可应对大型 3D 模型实时预览、仿真与工程运算，30 个 ISV 认证加持下成为两万元价位兼具本地 AI 算力与专业级稳定性的移动工作站标杆。同日，NVIDIA 与 Microsoft 联合宣布推出 RTX Spark 超级芯片，重新定义 Windows PC 面向个人 AI Agent 时代的定位，将 PC 从"工具"升级为"队友"。硬件侧的密集落地与国产芯片业绩兑现、终端 AI 应用爆发形成共振。

On August 10, Lenovo’s thin-and-light mobile AI workstation P16v officially went on sale, powered by an Intel Core Ultra 7 251HX and an NVIDIA RTX PRO 2000 Blackwell GPU — peak performance 200W, peak AI throughput 572 TOPS, with 30 ISV certifications. Positioned as a RMB 20,000-class mobile workstation balancing on-device AI and pro-grade stability, it supports real-time 3D preview, simulation and engineering compute. The same day, NVIDIA and Microsoft jointly unveiled RTX Spark, a new superchip redefining Windows PCs for the personal-AI-agent era — moving the PC from “tool” to “teammate”. The hardware cadence, paired with domestic chip earnings inflection and surging AI applications, signals end-to-end AI stack maturity.

来源 / Sources

• 财联社
• NVIDIA Newsroom

行业数据速览

媒体人解读

今天的产业信号可以归纳为三条主线：第一条主线是"AI 商业化兑现"，寒武纪上半年净利润同比增长 122.61%、摩尔线程营收同比增长 147.42%，加上宇树科技以 610 亿元市值启动科创板申购，意味着国产算力与机器人本体的资本市场估值锚已经形成，AI Agent / 具身智能不再是 PPT 上的故事，而是开始进入财报。第二条主线是"基础设施扩张 + 资源紧张"——首个全国产 10 万卡 AI 超集群投用是国产算力的里程碑事件，但 AWS 内部已经要为 CPU 资源告急，这意味着云端算力供需结构性失衡，AI 推理需求正在从 GPU 蔓延到 CPU，模型调用成本下降的同时，单次推理的边际成本结构发生根本变化。第三条主线是"自动驾驶合规化"，L3/L4 强制国标的发布是中国智能驾驶从"测试运营"切换至"规模商业化"的关键节点，叠加千里智驾 53 万辆装车量、华为智驾下探 10 万级市场，意味着高阶辅助驾驶正在从选配变成标配。对开发者和创业者而言：推理 Token 成本下降 → AI 应用爆发 → 算力需求扩张 → 国产芯片放量 → 资本回流，是一个已经形成正反馈的产业循环；接下来的胜负手在于产品力、合规力和成本控制能力。

Today’s industry signals can be grouped into three threads. First, “AI commercialization inflection” — Cambricon’s H1 net profit grew 122.61% YoY and Moore Threads’ revenue grew 147.42% YoY; combined with Unitree’s RMB 61 billion STAR Market launch, the valuation anchor for domestic compute and humanoid bodies has been set. Embodied AI and AI agents are no longer slideware but line items on financials. Second, “infrastructure scale-up meets resource scarcity” — the first fully-domestic 100,000-card AI supercluster is a milestone, while AWS has internally sounded the alarm on CPU shortages: AI inference demand is spilling from GPU to CPU, and as model-call costs fall, the marginal cost structure of each inference is fundamentally changing. Third, “autonomous-driving compliance” — the L3/L4 mandatory standard marks the switch from “test operations” to “scale commercialization” in China; paired with Qianli’s 530,000 installations and Huawei ADS extending into the RMB 100,000 segment, advanced driver assistance is moving from option to standard. For developers and founders: inference cost decline → AI application explosion → compute demand surge → domestic chip ramp → capital recirculation is now a closed-loop flywheel; the next battleground is product, compliance and cost control.

参考来源与延伸阅读

声明：以下为本文所涉事件的公开报道来源，供读者延伸查阅。本文内容仅为对公开信息的梳理与行业趋势分析，不构成任何投资建议。

主要新闻来源

• 宇树科技 IPO：上海证券交易所公告、证券时报
• 寒武纪 & 摩尔线程财报：公司官方公告（巨潮资讯网）、上海证券报
• OpenRouter 数据：OpenRouter 官方周榜
• 自动驾驶国标：工业和信息化部官网、国家标准全文公开系统
• 特斯拉 Terafab：特斯拉官方新闻室、路透社
• 诺因智能融资：投资界、[公司官方声明]
• 千里智驾 & 华为乾崑：汽车之家、第一电动网
• AI 超集群与 AWS：国家超算互联网、AWS 官方博客
• 联想与 NVIDIA：联想新闻中心、NVIDIA 新闻室

行业分析参考

• Gartner, “Hype Cycle for Artificial Intelligence, 2026”
• IDC, “China AI Infrastructure Market Forecast, 2026–2030”
• 中国信息通信研究院，《人工智能白皮书（2026年）》