AI科技热点日报 | 2026年8月12日

AI科技热点日报 | 2026年8月12日

综合整理自 TechCrunch、The Information、财联社、界面新闻、IT之家、36氪、钛媒体、第一财经、新华财经等科技财经媒体当日报道，聚焦国产大模型商业化拐点、AI芯片与人形机器人量产进展、企业级融资动态。

📌 今日摘要

2026 年 8 月 12 日，AI 行业出现三条主线：一是 DeepSeek V4-Plus 涨价方案正式落地，输入价 3.5 元/百万 tokens、输出价 12 元/百万 tokens，整体涨幅约 60%，首日调用量仍逆势增长 18%，被视为国产大模型"按价值定价"进入实质兑现阶段；二是端侧大模型独角兽面壁智能正式向北京证监局提交 IPO 辅导备案，半年内累计融资超 50 亿元、估值突破 200 亿元，成为端侧智能赛道估值最大独角兽；三是特斯拉弗里蒙特工厂原 Model S/X 产线已停用、首代 Optimus 专用产线进场安装，目标年产能 100 万台；高盛同步上调人形机器人出货预期，2027 年全球预计 7.6 万台、2032 年升至 50.2 万台。同期，微软计划 9 月发布下一代 Maia 300 AI 芯片并与台积电洽谈 30 万颗产能，Stripe 联合创始人创立的 River AI 完成 11 亿美元融资。

On August 12, 2026, the AI sector saw three headline threads. First, DeepSeek’s V4-Plus pricing officially took effect — input at 3.5 yuan and output at 12 yuan per million tokens, roughly a 60% increase — yet day-one call volume still grew 18%, signaling that Chinese foundation models have entered a real “price-for-value” phase. Second, edge-AI unicorn ModelBest (面壁智能) filed its IPO tutoring paperwork with the Beijing CSRC bureau; with over 5 billion yuan raised in H1 2026 and a valuation above 20 billion yuan, it is now the largest edge-AI unicorn. Third, Tesla’s Fremont factory has idled its Model S/X line and begun installing the first dedicated Optimus production line, with a target of 1 million units per year; Goldman Sachs raised its humanoid outlook, projecting 76,000 global shipments in 2027 and 502,000 by 2032. In parallel, Microsoft plans to launch its next-generation Maia 300 AI chip in September and is negotiating 300,000-unit capacity with TSMC, while Stripe co-founder–backed River AI closed a 1.1 billion dollar funding round.

1、面壁智能启动 A 股上市辅导，端侧大模型独角兽估值突破 200 亿元

事件概要

2026 年 8 月 11 日，端侧大模型企业北京面壁智能科技股份有限公司向北京证监局提交 IPO 辅导备案，辅导券商为中信证券，8 月 12 日正式出现在证监会公开发行辅导公示平台。备案报告显示，面壁智能无控股股东，第一大股东为北京清语启航科技中心（有限合伙），持股 16.45%；公司 2022 年孵化自清华大学 NLP 实验室，专注高效大模型全栈研发与端侧落地，旗下 MiniCPM 系列端侧模型全球下载量超 4300 万次，已在汽车、手机、具身智能等场景实现产业落地。融资方面，2026 年上半年公司累计融资金额超 50 亿元，估值突破 200 亿元，正式成为端侧智能赛道公开估值最大的独角兽；本轮涉及京国瑞、茅台基金、龙芯创投、中金保时捷等机构。本轮辅导备案的提交，意味着这家成立仅四年的清华系 AI 公司正式踏上 A 股资本市场征程。

On August 11, 2026, Beijing ModelBest (面壁智能) — an edge-AI foundation-model company — filed IPO tutoring paperwork with the Beijing CSRC bureau; CITIC Securities is the tutor, and the filing appeared on the regulator’s public tutoring platform on August 12. The filing shows no controlling shareholder, with the largest holder being Beijing Qingyu Qihang Technology Center (LP) at 16.45%. ModelBest was incubated in 2022 from Tsinghua University’s NLP lab and focuses on efficient foundation models and edge deployment; its MiniCPM edge models have exceeded 43 million downloads globally and shipped into automotive, smartphone and embodied-AI scenarios. In H1 2026 alone the company raised over 5 billion yuan across multiple rounds at a valuation above 20 billion yuan, making it the largest publicly valued edge-AI unicorn. Investors include Beijing Guorui, Moutai Fund, Loongson Venture and CICC Porsche. The filing marks the Beijing-based AI startup’s formal entry into the A-share IPO pipeline just four years after founding.

来源 / Sources

• 面壁智能启动上市辅导，中国最大端侧 AI 独角兽估值突破 200 亿 - 36氪
• AI 公司面壁智能启动 IPO - IT之家
• 面壁智能启动 A 股上市辅导 - 财联社

2、DeepSeek V4-Plus 涨价方案正式落地，国产大模型"按价值定价"进入兑现期

事件概要

2026 年 8 月 12 日，DeepSeek V4-Plus 新一轮价格调整正式落地：输入价格调整至 3.5 元/百万 tokens、输出价格调整至 12 元/百万 tokens，整体涨幅约 60%。值得注意的是，涨价首日 DeepSeek 平台调用量仍较前一日增长 18%，显示市场对头部国产模型的"提价承受力"明显抬升。同日，阿里 Qwen 正式实施"重度商用收入分成"机制，年 API 调用金额超过 500 万美元的大客户需进入新签约框架；月之暗面 Kimi K3.1 发布视频原生多模态能力，1M 上下文推理成本进一步下降 22%；智源社区 FlagOS 2.0 同步上线。本次同日集中亮相的"涨价 + 收费 + 能力升级"组合，意味着国产大模型正式从"以价换量"流量阶段，迈向"价值定价 + 商业化分层"阶段；对于 Agent 应用与企业 IT 采购方而言，2026 年 Q4 之前的 TCO 模型需要重新校准。

On August 12, 2026, DeepSeek’s V4-Plus pricing adjustment officially took effect, with input at 3.5 yuan and output at 12 yuan per million tokens — roughly a 60% increase. Despite the hike, day-one call volume grew 18% versus the previous day, indicating that the market’s tolerance for price increases at top-tier Chinese models has materially risen. On the same day, Alibaba’s Qwen formally launched a “heavy-use commercial revenue sharing” mechanism, requiring enterprise customers with annual API spend above 5 million USD to enter a new contract framework; Moonshot AI released Kimi K3.1 with native video multimodality and a 22% further reduction in 1M-context inference cost; BAAI’s FlagOS 2.0 also went live. The simultaneous landing of “price hikes + paid tiers + capability upgrades” marks China’s foundation-model sector’s formal transition from a price-for-share growth phase to a value-pricing and tiered-commercialization phase. For agent applications and enterprise IT procurement, TCO models will need to be rebaselined before Q4 2026.

来源 / Sources

• 国产大模型从价格竞争走向价值变现 - 南方基金
• DeepSeek V4-Plus 涨价方案落地，首日调用量仍增长 18% - 财联社
• Kimi K3.1 视频原生多模态发布，1M 上下文推理成本下降 22% - IT之家

3、特斯拉弗里蒙特工厂首条 Optimus 专用产线进场，目标年产能 100 万台

事件概要

2026 年 8 月 12 日，特斯拉 CEO 马斯克在最新财报会议上确认，加州弗里蒙特工厂原 Model S/X 产线已停用，第一代 Optimus 专用产线正在安装，预计很快启动生产，目标年产能 100 万台；得州超级工厂的 Optimus 机器人基地已进入主体建设阶段，远期目标年产能 1000 万台，预计 2027 年夏季投产。供应链端，特斯拉已于近期下发具体零件采购指引，8 月已现数百台订单，EDI 订单进入自动释放状态，正式采购订单（PO）已下发；供应链预期 9-10 月周产水平可达 1000-2000 台级别。7 月底至 8 月初，弗里蒙特工厂已启动小批量 SOP 量产，采用按月滚动订单模式。特斯拉 Optimus 量产从"指引阶段"正式进入"订单验证阶段"，三季度有望成为量产数据集中验证窗口。

On August 12, 2026, Tesla CEO Elon Musk confirmed on the latest earnings call that the Fremont factory’s Model S/X line has been idled and the first dedicated Optimus line is being installed, with production targeted to start soon at 1 million units per year. The Texas Gigafactory’s Optimus base is now in main-structure construction, with a long-term target of 10 million units per year, expected to come online in summer 2027. On the supply-chain side, Tesla has issued detailed parts procurement guidance, with hundreds of unit orders appearing in August; EDI orders are auto-releasing and formal POs have been issued; supply-chain partners expect 1,000-2,000 units per week by September-October. Late July to early August saw small-batch SOP production begin at Fremont on a monthly rolling-PO model. Tesla has moved from “guidance” to “order verification,” with Q3 shaping up as the concentrated verification window.

来源 / Sources

• Optimus 产线进场，高盛聚焦高技术壁垒核心部件 - 财联社
• 特斯拉弗里蒙特工厂首条 Optimus 专用产线正在安装 - 路透社
• 特斯拉 Optimus 量产进入订单验证阶段 - 第一财经

4、微软计划 9 月发布 Maia 300 AI 芯片，与台积电洽谈 30 万颗产能

事件概要

2026 年 8 月 10 日-11 日，据 The Information 援引两位知情人士消息，微软计划最早于 2026 年 9 月公开发布其下一代 Maia 300 人工智能加速器，并已与台积电就超过 30 万颗的产能达成协议，预计 2027 年开始交付。微软同时希望说服 Anthropic 等大型云服务客户采用 Maia 系列芯片，作为对英伟达昂贵 GPU 替代方案的延展。微软首批自研 AI 芯片 Maia 100 于 2023 年 11 月推出，但在大规模部署上持续落后于竞争对手；新一代 Maia 300 被视为微软自研 AI 芯片战略的关键拐点。受该消息影响，微软股价在消息确认日盘前交易中上涨约 0.7%。台积电是此次供应链的直接受益方。微软 Azure Maia 总经理 Andrew 的表态是：从长期看，公司目标是将 Maia 系列芯片的部署规模扩大至吉瓦（GW）级别。

Between August 10-11, 2026, The Information, citing two people with direct knowledge, reported that Microsoft plans to publicly launch its next-generation Maia 300 AI accelerator as early as September 2026, and has agreed with TSMC on capacity for more than 300,000 units, with deliveries expected to start in 2027. Microsoft is also working to convince large cloud customers like Anthropic to adopt Maia chips as an alternative to expensive Nvidia GPUs. The original Maia 100 chip launched in November 2023 but lagged at scale; the new Maia 300 is seen as a critical inflection point for Microsoft’s in-house AI chip strategy. The news pushed Microsoft’s stock up about 0.7% in pre-market trading on the day confirmation surfaced. TSMC is the direct beneficiary of the supply chain. Per Microsoft Azure Maia general manager Andrew, the long-term goal is to scale Maia deployment to the gigawatt level.

来源 / Sources

• Microsoft plans to release next-gen Maia 300 AI chip in September - The Information
• 微软最快 9 月发布 Maia 300 芯片，与台积电洽谈 30 万颗产能 - IT之家
• 微软 Maia 300 今秋亮相，目标产能 30 万片 - 财联社

5、爱芯元智下一代大算力 AI 芯片完成流片，支持 2/4 芯级联满血大模型推理

事件概要

2026 年 8 月 12 日，界面新闻从爱芯元智 2026 年中期业绩电话会上获悉，公司下一代高性能大算力 AI 芯片已完成流片，算力规格较上一代大幅提升，同时配备高带宽内存（HBM），并支持两芯或四芯级联，可实现满血大模型在边缘侧的高性能推理。该系列被命名为"元曦"系列大算力加速卡，以单芯片或多芯组合形态提供算力，能够实现满血大模型高性能推理，尤其擅长超长上下文处理、高清图像视频解析等任务，可在边缘侧实时、就近数据源完成推理。爱芯元智同期公布的上半年业绩显示，公司 2026 年上半年实现营业收入 4.02 亿元，同比增长 181.8%，毛利率提升至 29%；其中边缘 AI 业务同比增长 251.9%，已成为业绩增长引擎；新兴业务（智能汽车 + 边缘 AI）收入占比由 11.3% 升至 14.1%。爱芯元智也成为 8 月 9 日披露上市后首份半年报的港股 AI 芯片企业。

On August 12, 2026, Jiemian News learned from Aixin’s 2026 interim earnings call that the company’s next-generation high-performance AI chip has completed tape-out, with substantially higher compute specifications, HBM support, and 2/4-chip cascading to enable full-scale foundation-model inference at the edge. Branded the “Yuanxi” accelerator card series, it comes in single- or multi-chip form factors and is suited to ultra-long-context and high-resolution video-image understanding workloads, completing inference at the edge in real time. Aixin’s H1 2026 results showed revenue of 402 million yuan, up 181.8% year-on-year, with gross margin improving to 29%; the edge-AI segment grew 251.9%, becoming the growth engine, while emerging business (smart auto + edge AI) rose from 11.3% to 14.1% of revenue. Aixin disclosed its first post-listing interim report on the Hong Kong Stock Exchange on August 9.

来源 / Sources

• 爱芯元智下一代大算力 AI 芯片完成流片 - 界面新闻
• 爱芯元智 2026 上半年营收同比增长 181.8% - 21 世纪经济报道
• 边缘 AI 业务增长 251.9%，爱芯元智构建"边-车-端"全栈算力生态 - 财联社

6、高盛上调人形机器人出货预期：2027 年 7.6 万台、2032 年 50.2 万台

事件概要

2026 年 8 月 12 日，高盛最新研究报告显示，随着 AI 计算基础设施投资持续扩张，AI 投资主题正从算力竞赛转向终端应用，人形机器人有望成为下一阶段增长引擎。高盛预计全球人形机器人出货量将在 2027 年达到 7.6 万台，并于 2032 年升至 50.2 万台，大规模部署将在 2027-2029 年逐步展开。报告认为，劳动力短缺和自动化需求将为机器人产业提供长期结构性支撑。从主题驱动看，AI 投资主线的切换信号已经出现：算力板块经过近两年的高速扩张，进入"订单验证"阶段；而人形机器人赛道从"故事"走向"业绩兑现"。近期特斯拉 Optimus 量产进入订单验证阶段、谷歌 DeepMind 发布新一代具身智能大模型、得州超级工厂 Optimus 基地进入主体建设，叠加国内宇树科技科创板 IPO 已于 8 月 10 日启动申购，构成本轮人形机器人主题的核心催化组合。

On August 12, 2026, a Goldman Sachs research report noted that as AI compute investment continues to expand, the AI investment theme is shifting from compute race to endpoint applications, with humanoid robots positioned as the next growth engine. Goldman projects global humanoid shipments of 76,000 units in 2027, rising to 502,000 by 2032, with mass deployment taking shape over 2027-2029. The report cites labor shortages and automation demand as long-term structural tailwinds. The narrative shift is now visible: after nearly two years of rapid expansion, the compute segment is moving into “order verification,” while the humanoid segment is moving from “story” to “performance delivery.” Recent catalysts include Tesla Optimus entering order verification, Google DeepMind’s release of a new embodied-AI foundation model, the Texas Gigafactory Optimus base entering main-structure construction, and Unitree’s STAR Market IPO subscription opening on August 10.

来源 / Sources

• 高盛：人形机器人出货量 2027 年 7.6 万台、2032 年 50.2 万台 - 财联社
• Humanoid robot shipments to accelerate, Goldman Sachs projects - Bloomberg
• AI 主题切换，高盛唱多 Optimus 量产 - 第一财经

7、Stripe 联合创始人加码 AI：River AI 完成 11 亿美元融资，英伟达、AMD 战略投资

事件概要

2026 年 8 月 11 日，由 xAI 联合创始人 Igor Babuschkin（此前亦为 Stripe 主要工程师）创办的全栈 AI 公司 River AI 宣布完成 11 亿美元融资。本轮由 General Catalyst 和 AMP PBC 领投，英伟达和 AMD Ventures 作为战略投资方参与，Y Combinator、淡马锡等基金跟投。融资资金将用于建设开放 AI 技术栈，并为开发者和企业提供训练、微调及部署自有 AI 模型的工具。River AI 称，其 API 可在无需专门基础设施团队的情况下，于 15-20 分钟内完成复杂的强化学习训练，成本较闭源替代方案低 2-4 倍。从投资人结构看，本轮标志着 GPU 龙头（英伟达、AMD）首次以"同时战略投资"的方式押注同一家全栈 AI 创业公司，意味着 GPU 厂商对开放生态的争夺正在升温。

On August 11, 2026, full-stack AI company River AI — founded by Igor Babuschkin, xAI co-founder and former Stripe principal engineer — announced a 1.1 billion USD funding round. General Catalyst and AMP PBC co-led the round, with NVIDIA and AMD Ventures participating as strategic investors, and Y Combinator and Temasek joining. Proceeds will fund an open AI stack and tools for developers and enterprises to train, fine-tune, and deploy their own models. River AI claims its API can complete complex RL training in 15-20 minutes without dedicated infrastructure teams, at 2-4x lower cost than closed-source alternatives. The investor lineup marks the first time the two GPU leaders (NVIDIA and AMD) have simultaneously made strategic bets on the same full-stack AI startup, signaling an intensifying race for open AI ecosystems.

来源 / Sources

• River AI raises 1.1 billion USD with NVIDIA and AMD backing - BusinessWire
• Stripe 联合创始人加码 AI，River AI 完成 11 亿美元融资 - 财联社
• 英伟达、AMD 罕见同时出手，押注 xAI 联合创始人新项目 - TechCrunch

8、上海市经信委印发《"十五五"软件和信息服务业规划》，明确 GPU/NPU/QPU/HBM 协同方向

事件概要

2026 年 8 月 11 日，上海市经济和信息化委员会印发《上海市软件和信息服务业发展"十五五"规划》，提出攻坚下一代模型架构，推动基于状态空间模型、循环神经网络变体、液态神经网络等非 Transformer 架构的多技术路线探索；加快布局物理智能、世界模型、量子智能、类脑智能等前沿基础模型技术体系。规划同时明确，攻关超大规模智能算力集群组网技术，围绕高性能计算芯片（GPU/NPU）、量子芯片（QPU）、高速光互连（CPO）、高带宽内存（HBM）及异构服务器等核心环节，提升智算硬件的供给能力，推动自主芯片与主流大模型深度融合。从产业落地看，该规划延续了上海"模塑申城"工程方向；上海数据港、网宿科技、云赛智联、宝信软件等本地算力链上市公司被券商视为规划落地的直接受益方。

On August 11, 2026, the Shanghai Municipal Economic and Informatization Commission issued the “15th Five-Year Plan for the Software and Information Services Industry,” outlining plans to tackle next-generation model architectures, including exploration of non-Transformer architectures such as state-space models, recurrent neural network variants, and liquid neural networks; accelerate the layout of physical intelligence, world models, quantum intelligence, and brain-inspired intelligence frontier foundation models. The plan also sets out targets to tackle ultra-large-scale intelligent compute cluster networking technology, focusing on high-performance compute chips (GPU/NPU), quantum chips (QPU), co-packaged optics (CPO), high-bandwidth memory (HBM), and heterogeneous servers to enhance domestic supply and integrate domestic chips with mainstream foundation models. The plan extends the “Modelling Shanghai” initiative; Shanghai-listed compute supply chain companies including Shuju Data, Wangsu Science & Technology, Yunshai Intelligent, and Baosight Software are seen by sell-side analysts as direct beneficiaries.

来源 / Sources

• 上海市经信委印发《"十五五"软件和信息服务业规划》 - 财联社
• 上海"十五五"规划：攻关 GPU/NPU/QPU/HBM 核心环节 - 第一财经
• 上海算力建设相关受益方梳理 - 界面新闻

行业数据速览

媒体人解读

2026 年 8 月 12 日的 AI 产业图景，可以用"两升一降"来概括：升的是厂商按价值定价的底气与终端硬件的出货定价，降的是模型迭代叙事的边际吸引力。两条线最值得读者关注：

第一，DeepSeek V4-Plus 涨价首日调用量仍增长 18%，配合阿里 Qwen"重度商用收入分成"机制、Kimi K3.1 多模态升级、智源 FlagOS 2.0 同日落地，国产大模型已经从"以价换量"走向"按价值定价"。对开发者而言，这意味着 Agent 化应用的 TCO 测算在 Q4 前必须重做；对企业 IT 而言，"自建 + 商用 API"组合策略需要重新校准权重。

第二，硬件端的"中国制造"开始进入兑现期。微软 Maia 300 9 月发布、与台积电 30 万颗产能，意味着北美云厂的"自研芯片替代英伟达"从 PPT 走向量产；爱芯元智下一代大算力 AI 芯片完成流片、2/4 芯级联实现边缘满血大模型推理，叠加上海"十五五"规划明确 GPU/NPU/QPU/HBM 协同，国产算力链从"政策预期"走向"产品兑现"，2026 H2-2027 H1 是关键验证窗口。

第三，人形机器人赛道的"订单验证"窗口已经开启。特斯拉弗里蒙特工厂首条 Optimus 专用产线进场、目标年产 100 万台；高盛上调 2027 年 7.6 万台、2032 年 50.2 万台出货预期；特斯拉同步下发的零件采购订单已经进入自动释放状态。三季度财报与供应链订单数据，将是该赛道"从故事到业绩"切换的关键验证点。

The August 12, 2026 AI landscape can be summarized as “two ups, one down”: rising are vendors’ confidence in value-based pricing and end-device shipment targets, while declining is the marginal appeal of model iteration narratives. Two threads stand out:

First, the simultaneous landing of “price hikes + paid tiers + capability upgrades” — DeepSeek V4-Plus’s 18% day-one call-volume growth on a 60% price hike, Alibaba Qwen’s “heavy-use commercial revenue sharing” mechanism, Kimi K3.1’s multimodal upgrade, and BAAI FlagOS 2.0 — marks Chinese foundation models’ formal shift from “price-for-share” to “price-for-value.” Developers must rebuild agentic-app TCO models before Q4; enterprise IT must rebalance “build + commercial API” mix.

Second, “Made in China” hardware is entering its delivery phase. Microsoft’s September Maia 300 launch and 300,000-unit TSMC deal move North American hyperscalers’ “in-house chip as Nvidia alternative” plan from slide deck to volume; Aixin’s next-gen high-performance AI chip completing tape-out with 2/4-chip cascade for full-scale edge inference, combined with Shanghai’s “15th Five-Year Plan” specifying GPU/NPU/QPU/HBM coordination, means the domestic compute chain is shifting from “policy expectation” to “product delivery.” H2 2026 to H1 2027 is the key verification window.

Third, the humanoid-robot “order verification” window has opened. Tesla’s Fremont first dedicated Optimus line is being installed with a 1-million-unit annual target; Goldman Sachs raised 2027 shipments to 76,000 and 2032 to 502,000; Tesla-issued parts procurement POs are auto-releasing. Q3 earnings and supply-chain order data will be the critical verification point for the segment’s “story to performance” transition.

参考来源

英文来源

• Microsoft plans to release next-gen Maia 300 AI chip in September - The Information
• River AI raises 1.1 billion USD with NVIDIA and AMD backing - BusinessWire
• Goldman Sachs raises humanoid robot shipment forecast - Bloomberg
• River AI secures 1.1 billion USD with NVIDIA and AMD backing - TechCrunch

中文来源

• 面壁智能启动 A 股上市辅导，估值突破 200 亿元 - 36氪
• 面壁智能启动 IPO - IT之家
• DeepSeek V4-Plus 涨价方案落地 - 财联社
• Kimi K3.1 多模态升级 - 钛媒体
• 特斯拉 Optimus 产线进场 - 财联社
• 微软 Maia 300 9 月发布 - 财联社
• 爱芯元智下一代大算力 AI 芯片完成流片 - 界面新闻
• 上海"十五五"软件和信息服务业规划 - 财联社
• 高盛人形机器人出货预期 - 第一财经

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8月12日