日?qǐng)?bào) | 2026年8月12日)
文章目錄AI科技熱點(diǎn)日?qǐng)?bào) | 2026年8月12日 今日摘要1、面壁智能啟動(dòng) A 股上市輔導(dǎo)端側(cè)大模型獨(dú)角獸估值突破 200 億元事件概要來(lái)源 / Sources2、DeepSeek V4-Plus 漲價(jià)方案正式落地國(guó)產(chǎn)大模型按價(jià)值定價(jià)進(jìn)入兌現(xiàn)期事件概要來(lái)源 / Sources3、特斯拉弗里蒙特工廠首條 Optimus 專用產(chǎn)線進(jìn)場(chǎng)目標(biāo)年產(chǎn)能 100 萬(wàn)臺(tái)事件概要來(lái)源 / Sources4、微軟計(jì)劃 9 月發(fā)布 Maia 300 AI 芯片與臺(tái)積電洽談 30 萬(wàn)顆產(chǎn)能事件概要來(lái)源 / Sources5、愛(ài)芯元智下一代大算力 AI 芯片完成流片支持 2/4 芯級(jí)聯(lián)滿血大模型推理事件概要來(lái)源 / Sources6、高盛上調(diào)人形機(jī)器人出貨預(yù)期2027 年 7.6 萬(wàn)臺(tái)、2032 年 50.2 萬(wàn)臺(tái)事件概要來(lái)源 / Sources7、Stripe 聯(lián)合創(chuàng)始人加碼 AIRiver AI 完成 11 億美元融資英偉達(dá)、AMD 戰(zhàn)略投資事件概要來(lái)源 / Sources8、上海市經(jīng)信委印發(fā)《十五五軟件和信息服務(wù)業(yè)規(guī)劃》明確 GPU/NPU/QPU/HBM 協(xié)同方向事件概要來(lái)源 / Sources行業(yè)數(shù)據(jù)速覽媒體人解讀參考來(lái)源英文來(lái)源中文來(lái)源AI科技熱點(diǎn)日?qǐng)?bào) | 2026年8月12日綜合整理自 TechCrunch、The Information、財(cái)聯(lián)社、界面新聞、IT之家、36氪、鈦媒體、第一財(cái)經(jīng)、新華財(cái)經(jīng)等科技財(cái)經(jīng)媒體當(dāng)日?qǐng)?bào)道聚焦國(guó)產(chǎn)大模型商業(yè)化拐點(diǎn)、AI芯片與人形機(jī)器人量產(chǎn)進(jìn)展、企業(yè)級(jí)融資動(dòng)態(tài)。 今日摘要2026 年 8 月 12 日AI 行業(yè)出現(xiàn)三條主線一是 DeepSeek V4-Plus 漲價(jià)方案正式落地輸入價(jià) 3.5 元/百萬(wàn) tokens、輸出價(jià) 12 元/百萬(wàn) tokens整體漲幅約 60%首日調(diào)用量仍逆勢(shì)增長(zhǎng) 18%被視為國(guó)產(chǎn)大模型按價(jià)值定價(jià)進(jìn)入實(shí)質(zhì)兌現(xiàn)階段二是端側(cè)大模型獨(dú)角獸面壁智能正式向北京證監(jiān)局提交 IPO 輔導(dǎo)備案半年內(nèi)累計(jì)融資超 50 億元、估值突破 200 億元成為端側(cè)智能賽道估值最大獨(dú)角獸三是特斯拉弗里蒙特工廠原 Model S/X 產(chǎn)線已停用、首代 Optimus 專用產(chǎn)線進(jìn)場(chǎng)安裝目標(biāo)年產(chǎn)能 100 萬(wàn)臺(tái)高盛同步上調(diào)人形機(jī)器人出貨預(yù)期2027 年全球預(yù)計(jì) 7.6 萬(wàn)臺(tái)、2032 年升至 50.2 萬(wàn)臺(tái)。同期微軟計(jì)劃 9 月發(fā)布下一代 Maia 300 AI 芯片并與臺(tái)積電洽談 30 萬(wàn)顆產(chǎn)能Stripe 聯(lián)合創(chuàng)始人創(chuàng)立的 River AI 完成 11 億美元融資。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、面壁智能啟動(dòng) A 股上市輔導(dǎo)端側(cè)大模型獨(dú)角獸估值突破 200 億元事件概要2026 年 8 月 11 日端側(cè)大模型企業(yè)北京面壁智能科技股份有限公司向北京證監(jiān)局提交 IPO 輔導(dǎo)備案輔導(dǎo)券商為中信證券8 月 12 日正式出現(xiàn)在證監(jiān)會(huì)公開(kāi)發(fā)行輔導(dǎo)公示平臺(tái)。備案報(bào)告顯示面壁智能無(wú)控股股東第一大股東為北京清語(yǔ)啟航科技中心有限合伙持股 16.45%公司 2022 年孵化自清華大學(xué) NLP 實(shí)驗(yàn)室專注高效大模型全棧研發(fā)與端側(cè)落地旗下 MiniCPM 系列端側(cè)模型全球下載量超 4300 萬(wàn)次已在汽車、手機(jī)、具身智能等場(chǎng)景實(shí)現(xiàn)產(chǎn)業(yè)落地。融資方面2026 年上半年公司累計(jì)融資金額超 50 億元估值突破 200 億元正式成為端側(cè)智能賽道公開(kāi)估值最大的獨(dú)角獸本輪涉及京國(guó)瑞、茅臺(tái)基金、龍芯創(chuàng)投、中金保時(shí)捷等機(jī)構(gòu)。本輪輔導(dǎo)備案的提交意味著這家成立僅四年的清華系 AI 公司正式踏上 A 股資本市場(chǎng)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.來(lái)源 / Sources面壁智能啟動(dòng)上市輔導(dǎo)中國(guó)最大端側(cè) AI 獨(dú)角獸估值突破 200 億 - 36氪AI 公司面壁智能啟動(dòng) IPO - IT之家面壁智能啟動(dòng) A 股上市輔導(dǎo) - 財(cái)聯(lián)社2、DeepSeek V4-Plus 漲價(jià)方案正式落地國(guó)產(chǎn)大模型按價(jià)值定價(jià)進(jìn)入兌現(xiàn)期事件概要2026 年 8 月 12 日DeepSeek V4-Plus 新一輪價(jià)格調(diào)整正式落地輸入價(jià)格調(diào)整至 3.5 元/百萬(wàn) tokens、輸出價(jià)格調(diào)整至 12 元/百萬(wàn) tokens整體漲幅約 60%。值得注意的是漲價(jià)首日 DeepSeek 平臺(tái)調(diào)用量仍較前一日增長(zhǎng) 18%顯示市場(chǎng)對(duì)頭部國(guó)產(chǎn)模型的提價(jià)承受力明顯抬升。同日阿里 Qwen 正式實(shí)施重度商用收入分成機(jī)制年 API 調(diào)用金額超過(guò) 500 萬(wàn)美元的大客戶需進(jìn)入新簽約框架月之暗面 Kimi K3.1 發(fā)布視頻原生多模態(tài)能力1M 上下文推理成本進(jìn)一步下降 22%智源社區(qū) FlagOS 2.0 同步上線。本次同日集中亮相的漲價(jià) 收費(fèi) 能力升級(jí)組合意味著國(guó)產(chǎn)大模型正式從以價(jià)換量流量階段邁向價(jià)值定價(jià) 商業(yè)化分層階段對(duì)于 Agent 應(yīng)用與企業(yè) IT 采購(gòu)方而言2026 年 Q4 之前的 TCO 模型需要重新校準(zhǔn)。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.來(lái)源 / Sources國(guó)產(chǎn)大模型從價(jià)格競(jìng)爭(zhēng)走向價(jià)值變現(xiàn) - 南方基金DeepSeek V4-Plus 漲價(jià)方案落地首日調(diào)用量仍增長(zhǎng) 18% - 財(cái)聯(lián)社Kimi K3.1 視頻原生多模態(tài)發(fā)布1M 上下文推理成本下降 22% - IT之家3、特斯拉弗里蒙特工廠首條 Optimus 專用產(chǎn)線進(jìn)場(chǎng)目標(biāo)年產(chǎn)能 100 萬(wàn)臺(tái)事件概要2026 年 8 月 12 日特斯拉 CEO 馬斯克在最新財(cái)報(bào)會(huì)議上確認(rèn)加州弗里蒙特工廠原 Model S/X 產(chǎn)線已停用第一代 Optimus 專用產(chǎn)線正在安裝預(yù)計(jì)很快啟動(dòng)生產(chǎn)目標(biāo)年產(chǎn)能 100 萬(wàn)臺(tái)得州超級(jí)工廠的 Optimus 機(jī)器人基地已進(jìn)入主體建設(shè)階段遠(yuǎn)期目標(biāo)年產(chǎn)能 1000 萬(wàn)臺(tái)預(yù)計(jì) 2027 年夏季投產(chǎn)。供應(yīng)鏈端特斯拉已于近期下發(fā)具體零件采購(gòu)指引8 月已現(xiàn)數(shù)百臺(tái)訂單EDI 訂單進(jìn)入自動(dòng)釋放狀態(tài)正式采購(gòu)訂單PO已下發(fā)供應(yīng)鏈預(yù)期 9-10 月周產(chǎn)水平可達(dá) 1000-2000 臺(tái)級(jí)別。7 月底至 8 月初弗里蒙特工廠已啟動(dòng)小批量 SOP 量產(chǎn)采用按月滾動(dòng)訂單模式。特斯拉 Optimus 量產(chǎn)從指引階段正式進(jìn)入訂單驗(yàn)證階段三季度有望成為量產(chǎn)數(shù)據(jù)集中驗(yàn)證窗口。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.來(lái)源 / SourcesOptimus 產(chǎn)線進(jìn)場(chǎng)高盛聚焦高技術(shù)壁壘核心部件 - 財(cái)聯(lián)社特斯拉弗里蒙特工廠首條 Optimus 專用產(chǎn)線正在安裝 - 路透社特斯拉 Optimus 量產(chǎn)進(jìn)入訂單驗(yàn)證階段 - 第一財(cái)經(jīng)4、微軟計(jì)劃 9 月發(fā)布 Maia 300 AI 芯片與臺(tái)積電洽談 30 萬(wàn)顆產(chǎn)能事件概要2026 年 8 月 10 日-11 日據(jù) The Information 援引兩位知情人士消息微軟計(jì)劃最早于 2026 年 9 月公開(kāi)發(fā)布其下一代 Maia 300 人工智能加速器并已與臺(tái)積電就超過(guò) 30 萬(wàn)顆的產(chǎn)能達(dá)成協(xié)議預(yù)計(jì) 2027 年開(kāi)始交付。微軟同時(shí)希望說(shuō)服 Anthropic 等大型云服務(wù)客戶采用 Maia 系列芯片作為對(duì)英偉達(dá)昂貴 GPU 替代方案的延展。微軟首批自研 AI 芯片 Maia 100 于 2023 年 11 月推出但在大規(guī)模部署上持續(xù)落后于競(jìng)爭(zhēng)對(duì)手新一代 Maia 300 被視為微軟自研 AI 芯片戰(zhàn)略的關(guān)鍵拐點(diǎn)。受該消息影響微軟股價(jià)在消息確認(rèn)日盤前交易中上漲約 0.7%。臺(tái)積電是此次供應(yīng)鏈的直接受益方。微軟 Azure Maia 總經(jīng)理 Andrew 的表態(tài)是從長(zhǎng)期看公司目標(biāo)是將 Maia 系列芯片的部署規(guī)模擴(kuò)大至吉瓦GW級(jí)別。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.來(lái)源 / SourcesMicrosoft plans to release next-gen Maia 300 AI chip in September - The Information微軟最快 9 月發(fā)布 Maia 300 芯片與臺(tái)積電洽談 30 萬(wàn)顆產(chǎn)能 - IT之家微軟 Maia 300 今秋亮相目標(biāo)產(chǎn)能 30 萬(wàn)片 - 財(cái)聯(lián)社5、愛(ài)芯元智下一代大算力 AI 芯片完成流片支持 2/4 芯級(jí)聯(lián)滿血大模型推理事件概要2026 年 8 月 12 日界面新聞從愛(ài)芯元智 2026 年中期業(yè)績(jī)電話會(huì)上獲悉公司下一代高性能大算力 AI 芯片已完成流片算力規(guī)格較上一代大幅提升同時(shí)配備高帶寬內(nèi)存HBM并支持兩芯或四芯級(jí)聯(lián)可實(shí)現(xiàn)滿血大模型在邊緣側(cè)的高性能推理。該系列被命名為元曦系列大算力加速卡以單芯片或多芯組合形態(tài)提供算力能夠?qū)崿F(xiàn)滿血大模型高性能推理尤其擅長(zhǎng)超長(zhǎng)上下文處理、高清圖像視頻解析等任務(wù)可在邊緣側(cè)實(shí)時(shí)、就近數(shù)據(jù)源完成推理。愛(ài)芯元智同期公布的上半年業(yè)績(jī)顯示公司 2026 年上半年實(shí)現(xiàn)營(yíng)業(yè)收入 4.02 億元同比增長(zhǎng) 181.8%毛利率提升至 29%其中邊緣 AI 業(yè)務(wù)同比增長(zhǎng) 251.9%已成為業(yè)績(jī)?cè)鲩L(zhǎng)引擎新興業(yè)務(wù)智能汽車 邊緣 AI收入占比由 11.3% 升至 14.1%。愛(ài)芯元智也成為 8 月 9 日披露上市后首份半年報(bào)的港股 AI 芯片企業(yè)。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.來(lái)源 / Sources愛(ài)芯元智下一代大算力 AI 芯片完成流片 - 界面新聞愛(ài)芯元智 2026 上半年?duì)I收同比增長(zhǎng) 181.8% - 21 世紀(jì)經(jīng)濟(jì)報(bào)道邊緣 AI 業(yè)務(wù)增長(zhǎng) 251.9%愛(ài)芯元智構(gòu)建邊-車-端全棧算力生態(tài) - 財(cái)聯(lián)社6、高盛上調(diào)人形機(jī)器人出貨預(yù)期2027 年 7.6 萬(wàn)臺(tái)、2032 年 50.2 萬(wàn)臺(tái)事件概要2026 年 8 月 12 日高盛最新研究報(bào)告顯示隨著 AI 計(jì)算基礎(chǔ)設(shè)施投資持續(xù)擴(kuò)張AI 投資主題正從算力競(jìng)賽轉(zhuǎn)向終端應(yīng)用人形機(jī)器人有望成為下一階段增長(zhǎng)引擎。高盛預(yù)計(jì)全球人形機(jī)器人出貨量將在 2027 年達(dá)到 7.6 萬(wàn)臺(tái)并于 2032 年升至 50.2 萬(wàn)臺(tái)大規(guī)模部署將在 2027-2029 年逐步展開(kāi)。報(bào)告認(rèn)為勞動(dòng)力短缺和自動(dòng)化需求將為機(jī)器人產(chǎn)業(yè)提供長(zhǎng)期結(jié)構(gòu)性支撐。從主題驅(qū)動(dòng)看AI 投資主線的切換信號(hào)已經(jīng)出現(xiàn)算力板塊經(jīng)過(guò)近兩年的高速擴(kuò)張進(jìn)入訂單驗(yàn)證階段而人形機(jī)器人賽道從故事走向業(yè)績(jī)兌現(xiàn)。近期特斯拉 Optimus 量產(chǎn)進(jìn)入訂單驗(yàn)證階段、谷歌 DeepMind 發(fā)布新一代具身智能大模型、得州超級(jí)工廠 Optimus 基地進(jìn)入主體建設(shè)疊加國(guó)內(nèi)宇樹(shù)科技科創(chuàng)板 IPO 已于 8 月 10 日啟動(dòng)申購(gòu)構(gòu)成本輪人形機(jī)器人主題的核心催化組合。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.來(lái)源 / Sources高盛人形機(jī)器人出貨量 2027 年 7.6 萬(wàn)臺(tái)、2032 年 50.2 萬(wàn)臺(tái) - 財(cái)聯(lián)社Humanoid robot shipments to accelerate, Goldman Sachs projects - BloombergAI 主題切換高盛唱多 Optimus 量產(chǎn) - 第一財(cái)經(jīng)7、Stripe 聯(lián)合創(chuàng)始人加碼 AIRiver AI 完成 11 億美元融資英偉達(dá)、AMD 戰(zhàn)略投資事件概要2026 年 8 月 11 日由 xAI 聯(lián)合創(chuàng)始人 Igor Babuschkin此前亦為 Stripe 主要工程師創(chuàng)辦的全棧 AI 公司 River AI 宣布完成 11 億美元融資。本輪由 General Catalyst 和 AMP PBC 領(lǐng)投英偉達(dá)和 AMD Ventures 作為戰(zhàn)略投資方參與Y Combinator、淡馬錫等基金跟投。融資資金將用于建設(shè)開(kāi)放 AI 技術(shù)棧并為開(kāi)發(fā)者和企業(yè)提供訓(xùn)練、微調(diào)及部署自有 AI 模型的工具。River AI 稱其 API 可在無(wú)需專門基礎(chǔ)設(shè)施團(tuán)隊(duì)的情況下于 15-20 分鐘內(nèi)完成復(fù)雜的強(qiáng)化學(xué)習(xí)訓(xùn)練成本較閉源替代方案低 2-4 倍。從投資人結(jié)構(gòu)看本輪標(biāo)志著 GPU 龍頭英偉達(dá)、AMD首次以同時(shí)戰(zhàn)略投資的方式押注同一家全棧 AI 創(chuàng)業(yè)公司意味著 GPU 廠商對(duì)開(kāi)放生態(tài)的爭(zhēng)奪正在升溫。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.來(lái)源 / SourcesRiver AI raises 1.1 billion USD with NVIDIA and AMD backing - BusinessWireStripe 聯(lián)合創(chuàng)始人加碼 AIRiver AI 完成 11 億美元融資 - 財(cái)聯(lián)社英偉達(dá)、AMD 罕見(jiàn)同時(shí)出手押注 xAI 聯(lián)合創(chuàng)始人新項(xiàng)目 - TechCrunch8、上海市經(jīng)信委印發(fā)《十五五軟件和信息服務(wù)業(yè)規(guī)劃》明確 GPU/NPU/QPU/HBM 協(xié)同方向事件概要2026 年 8 月 11 日上海市經(jīng)濟(jì)和信息化委員會(huì)印發(fā)《上海市軟件和信息服務(wù)業(yè)發(fā)展十五五規(guī)劃》提出攻堅(jiān)下一代模型架構(gòu)推動(dòng)基于狀態(tài)空間模型、循環(huán)神經(jīng)網(wǎng)絡(luò)變體、液態(tài)神經(jīng)網(wǎng)絡(luò)等非 Transformer 架構(gòu)的多技術(shù)路線探索加快布局物理智能、世界模型、量子智能、類腦智能等前沿基礎(chǔ)模型技術(shù)體系。規(guī)劃同時(shí)明確攻關(guān)超大規(guī)模智能算力集群組網(wǎng)技術(shù)圍繞高性能計(jì)算芯片GPU/NPU、量子芯片QPU、高速光互連CPO、高帶寬內(nèi)存HBM及異構(gòu)服務(wù)器等核心環(huán)節(jié)提升智算硬件的供給能力推動(dòng)自主芯片與主流大模型深度融合。從產(chǎn)業(yè)落地看該規(guī)劃延續(xù)了上海模塑申城工程方向上海數(shù)據(jù)港、網(wǎng)宿科技、云賽智聯(lián)、寶信軟件等本地算力鏈上市公司被券商視為規(guī)劃落地的直接受益方。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.來(lái)源 / Sources上海市經(jīng)信委印發(fā)《十五五軟件和信息服務(wù)業(yè)規(guī)劃》 - 財(cái)聯(lián)社上海十五五規(guī)劃攻關(guān) GPU/NPU/QPU/HBM 核心環(huán)節(jié) - 第一財(cái)經(jīng)上海算力建設(shè)相關(guān)受益方梳理 - 界面新聞行業(yè)數(shù)據(jù)速覽指標(biāo)數(shù)據(jù)備注DeepSeek V4-Plus 輸入價(jià)格3.5 元/百萬(wàn) tokens較前次漲價(jià)約 60%DeepSeek V4-Plus 輸出價(jià)格12 元/百萬(wàn) tokens首日調(diào)用量仍增長(zhǎng) 18%Kimi K3.1 1M 上下文推理成本下降 22%視頻原生多模態(tài)面壁智能半年累計(jì)融資超 50 億元估值突破 200 億元愛(ài)芯元智 H1 營(yíng)收4.02 億元同比增長(zhǎng) 181.8%愛(ài)芯元智邊緣 AI 業(yè)務(wù)增速251.9%新興業(yè)務(wù)占比 14.1%微軟 Maia 300 芯片產(chǎn)能30 萬(wàn)顆2027 年交付與臺(tái)積電合作River AI 融資規(guī)模11 億美元NVIDIA、AMD 戰(zhàn)略投資高盛 2027 年人形機(jī)器人出貨預(yù)期7.6 萬(wàn)臺(tái)全球高盛 2032 年人形機(jī)器人出貨預(yù)期50.2 萬(wàn)臺(tái)2032 年特斯拉 Optimus 弗里蒙特目標(biāo)年產(chǎn)能100 萬(wàn)臺(tái)產(chǎn)線正在安裝特斯拉 Optimus 得州遠(yuǎn)期目標(biāo)年產(chǎn)能1000 萬(wàn)臺(tái)2027 年夏季投產(chǎn)媒體人解讀2026 年 8 月 12 日的 AI 產(chǎn)業(yè)圖景可以用兩升一降來(lái)概括升的是廠商按價(jià)值定價(jià)的底氣與終端硬件的出貨定價(jià)降的是模型迭代敘事的邊際吸引力。兩條線最值得讀者關(guān)注第一DeepSeek V4-Plus 漲價(jià)首日調(diào)用量仍增長(zhǎng) 18%配合阿里 Qwen重度商用收入分成機(jī)制、Kimi K3.1 多模態(tài)升級(jí)、智源 FlagOS 2.0 同日落地國(guó)產(chǎn)大模型已經(jīng)從以價(jià)換量走向按價(jià)值定價(jià)。對(duì)開(kāi)發(fā)者而言這意味著 Agent 化應(yīng)用的 TCO 測(cè)算在 Q4 前必須重做對(duì)企業(yè) IT 而言自建 商用 API組合策略需要重新校準(zhǔn)權(quán)重。第二硬件端的中國(guó)制造開(kāi)始進(jìn)入兌現(xiàn)期。微軟 Maia 300 9 月發(fā)布、與臺(tái)積電 30 萬(wàn)顆產(chǎn)能意味著北美云廠的自研芯片替代英偉達(dá)從 PPT 走向量產(chǎn)愛(ài)芯元智下一代大算力 AI 芯片完成流片、2/4 芯級(jí)聯(lián)實(shí)現(xiàn)邊緣滿血大模型推理疊加上海十五五規(guī)劃明確 GPU/NPU/QPU/HBM 協(xié)同國(guó)產(chǎn)算力鏈從政策預(yù)期走向產(chǎn)品兌現(xiàn)2026 H2-2027 H1 是關(guān)鍵驗(yàn)證窗口。第三人形機(jī)器人賽道的訂單驗(yàn)證窗口已經(jīng)開(kāi)啟。特斯拉弗里蒙特工廠首條 Optimus 專用產(chǎn)線進(jìn)場(chǎng)、目標(biāo)年產(chǎn) 100 萬(wàn)臺(tái)高盛上調(diào) 2027 年 7.6 萬(wàn)臺(tái)、2032 年 50.2 萬(wàn)臺(tái)出貨預(yù)期特斯拉同步下發(fā)的零件采購(gòu)訂單已經(jīng)進(jìn)入自動(dòng)釋放狀態(tài)。三季度財(cái)報(bào)與供應(yīng)鏈訂單數(shù)據(jù)將是該賽道從故事到業(yè)績(jī)切換的關(guān)鍵驗(yàn)證點(diǎn)。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.參考來(lái)源英文來(lái)源Microsoft plans to release next-gen Maia 300 AI chip in September - The InformationRiver AI raises 1.1 billion USD with NVIDIA and AMD backing - BusinessWireGoldman Sachs raises humanoid robot shipment forecast - BloombergRiver AI secures 1.1 billion USD with NVIDIA and AMD backing - TechCrunch中文來(lái)源面壁智能啟動(dòng) A 股上市輔導(dǎo)估值突破 200 億元 - 36氪面壁智能啟動(dòng) IPO - IT之家DeepSeek V4-Plus 漲價(jià)方案落地 - 財(cái)聯(lián)社Kimi K3.1 多模態(tài)升級(jí) - 鈦媒體特斯拉 Optimus 產(chǎn)線進(jìn)場(chǎng) - 財(cái)聯(lián)社微軟 Maia 300 9 月發(fā)布 - 財(cái)聯(lián)社愛(ài)芯元智下一代大算力 AI 芯片完成流片 - 界面新聞上海十五五軟件和信息服務(wù)業(yè)規(guī)劃 - 財(cái)聯(lián)社高盛人形機(jī)器人出貨預(yù)期 - 第一財(cái)經(jīng)本文內(nèi)容綜合整理自上述公開(kāi)信息來(lái)源媒體人解讀僅代表個(gè)人觀點(diǎn)僅供參考。作者署名劉一說(shuō) ?發(fā)布日期2026年8月12日 AI科技熱點(diǎn)日?qǐng)?bào) | 2026年8月12日
文章目錄AI科技熱點(diǎn)日?qǐng)?bào) | 2026年8月12日 今日摘要1、面壁智能啟動(dòng) A 股上市輔導(dǎo)端側(cè)大模型獨(dú)角獸估值突破 200 億元事件概要來(lái)源 / Sources2、DeepSeek V4-Plus 漲價(jià)方案正式落地國(guó)產(chǎn)大模型按價(jià)值定價(jià)進(jìn)入兌現(xiàn)期事件概要來(lái)源 / Sources3、特斯拉弗里蒙特工廠首條 Optimus 專用產(chǎn)線進(jìn)場(chǎng)目標(biāo)年產(chǎn)能 100 萬(wàn)臺(tái)事件概要來(lái)源 / Sources4、微軟計(jì)劃 9 月發(fā)布 Maia 300 AI 芯片與臺(tái)積電洽談 30 萬(wàn)顆產(chǎn)能事件概要來(lái)源 / Sources5、愛(ài)芯元智下一代大算力 AI 芯片完成流片支持 2/4 芯級(jí)聯(lián)滿血大模型推理事件概要來(lái)源 / Sources6、高盛上調(diào)人形機(jī)器人出貨預(yù)期2027 年 7.6 萬(wàn)臺(tái)、2032 年 50.2 萬(wàn)臺(tái)事件概要來(lái)源 / Sources7、Stripe 聯(lián)合創(chuàng)始人加碼 AIRiver AI 完成 11 億美元融資英偉達(dá)、AMD 戰(zhàn)略投資事件概要來(lái)源 / Sources8、上海市經(jīng)信委印發(fā)《十五五軟件和信息服務(wù)業(yè)規(guī)劃》明確 GPU/NPU/QPU/HBM 協(xié)同方向事件概要來(lái)源 / Sources行業(yè)數(shù)據(jù)速覽媒體人解讀參考來(lái)源英文來(lái)源中文來(lái)源AI科技熱點(diǎn)日?qǐng)?bào) | 2026年8月12日綜合整理自 TechCrunch、The Information、財(cái)聯(lián)社、界面新聞、IT之家、36氪、鈦媒體、第一財(cái)經(jīng)、新華財(cái)經(jīng)等科技財(cái)經(jīng)媒體當(dāng)日?qǐng)?bào)道聚焦國(guó)產(chǎn)大模型商業(yè)化拐點(diǎn)、AI芯片與人形機(jī)器人量產(chǎn)進(jìn)展、企業(yè)級(jí)融資動(dòng)態(tài)。 今日摘要2026 年 8 月 12 日AI 行業(yè)出現(xiàn)三條主線一是 DeepSeek V4-Plus 漲價(jià)方案正式落地輸入價(jià) 3.5 元/百萬(wàn) tokens、輸出價(jià) 12 元/百萬(wàn) tokens整體漲幅約 60%首日調(diào)用量仍逆勢(shì)增長(zhǎng) 18%被視為國(guó)產(chǎn)大模型按價(jià)值定價(jià)進(jìn)入實(shí)質(zhì)兌現(xiàn)階段二是端側(cè)大模型獨(dú)角獸面壁智能正式向北京證監(jiān)局提交 IPO 輔導(dǎo)備案半年內(nèi)累計(jì)融資超 50 億元、估值突破 200 億元成為端側(cè)智能賽道估值最大獨(dú)角獸三是特斯拉弗里蒙特工廠原 Model S/X 產(chǎn)線已停用、首代 Optimus 專用產(chǎn)線進(jìn)場(chǎng)安裝目標(biāo)年產(chǎn)能 100 萬(wàn)臺(tái)高盛同步上調(diào)人形機(jī)器人出貨預(yù)期2027 年全球預(yù)計(jì) 7.6 萬(wàn)臺(tái)、2032 年升至 50.2 萬(wàn)臺(tái)。同期微軟計(jì)劃 9 月發(fā)布下一代 Maia 300 AI 芯片并與臺(tái)積電洽談 30 萬(wàn)顆產(chǎn)能Stripe 聯(lián)合創(chuàng)始人創(chuàng)立的 River AI 完成 11 億美元融資。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、面壁智能啟動(dòng) A 股上市輔導(dǎo)端側(cè)大模型獨(dú)角獸估值突破 200 億元事件概要2026 年 8 月 11 日端側(cè)大模型企業(yè)北京面壁智能科技股份有限公司向北京證監(jiān)局提交 IPO 輔導(dǎo)備案輔導(dǎo)券商為中信證券8 月 12 日正式出現(xiàn)在證監(jiān)會(huì)公開(kāi)發(fā)行輔導(dǎo)公示平臺(tái)。備案報(bào)告顯示面壁智能無(wú)控股股東第一大股東為北京清語(yǔ)啟航科技中心有限合伙持股 16.45%公司 2022 年孵化自清華大學(xué) NLP 實(shí)驗(yàn)室專注高效大模型全棧研發(fā)與端側(cè)落地旗下 MiniCPM 系列端側(cè)模型全球下載量超 4300 萬(wàn)次已在汽車、手機(jī)、具身智能等場(chǎng)景實(shí)現(xiàn)產(chǎn)業(yè)落地。融資方面2026 年上半年公司累計(jì)融資金額超 50 億元估值突破 200 億元正式成為端側(cè)智能賽道公開(kāi)估值最大的獨(dú)角獸本輪涉及京國(guó)瑞、茅臺(tái)基金、龍芯創(chuàng)投、中金保時(shí)捷等機(jī)構(gòu)。本輪輔導(dǎo)備案的提交意味著這家成立僅四年的清華系 AI 公司正式踏上 A 股資本市場(chǎng)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.來(lái)源 / Sources面壁智能啟動(dòng)上市輔導(dǎo)中國(guó)最大端側(cè) AI 獨(dú)角獸估值突破 200 億 - 36氪AI 公司面壁智能啟動(dòng) IPO - IT之家面壁智能啟動(dòng) A 股上市輔導(dǎo) - 財(cái)聯(lián)社2、DeepSeek V4-Plus 漲價(jià)方案正式落地國(guó)產(chǎn)大模型按價(jià)值定價(jià)進(jìn)入兌現(xiàn)期事件概要2026 年 8 月 12 日DeepSeek V4-Plus 新一輪價(jià)格調(diào)整正式落地輸入價(jià)格調(diào)整至 3.5 元/百萬(wàn) tokens、輸出價(jià)格調(diào)整至 12 元/百萬(wàn) tokens整體漲幅約 60%。值得注意的是漲價(jià)首日 DeepSeek 平臺(tái)調(diào)用量仍較前一日增長(zhǎng) 18%顯示市場(chǎng)對(duì)頭部國(guó)產(chǎn)模型的提價(jià)承受力明顯抬升。同日阿里 Qwen 正式實(shí)施重度商用收入分成機(jī)制年 API 調(diào)用金額超過(guò) 500 萬(wàn)美元的大客戶需進(jìn)入新簽約框架月之暗面 Kimi K3.1 發(fā)布視頻原生多模態(tài)能力1M 上下文推理成本進(jìn)一步下降 22%智源社區(qū) FlagOS 2.0 同步上線。本次同日集中亮相的漲價(jià) 收費(fèi) 能力升級(jí)組合意味著國(guó)產(chǎn)大模型正式從以價(jià)換量流量階段邁向價(jià)值定價(jià) 商業(yè)化分層階段對(duì)于 Agent 應(yīng)用與企業(yè) IT 采購(gòu)方而言2026 年 Q4 之前的 TCO 模型需要重新校準(zhǔn)。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.來(lái)源 / Sources國(guó)產(chǎn)大模型從價(jià)格競(jìng)爭(zhēng)走向價(jià)值變現(xiàn) - 南方基金DeepSeek V4-Plus 漲價(jià)方案落地首日調(diào)用量仍增長(zhǎng) 18% - 財(cái)聯(lián)社Kimi K3.1 視頻原生多模態(tài)發(fā)布1M 上下文推理成本下降 22% - IT之家3、特斯拉弗里蒙特工廠首條 Optimus 專用產(chǎn)線進(jìn)場(chǎng)目標(biāo)年產(chǎn)能 100 萬(wàn)臺(tái)事件概要2026 年 8 月 12 日特斯拉 CEO 馬斯克在最新財(cái)報(bào)會(huì)議上確認(rèn)加州弗里蒙特工廠原 Model S/X 產(chǎn)線已停用第一代 Optimus 專用產(chǎn)線正在安裝預(yù)計(jì)很快啟動(dòng)生產(chǎn)目標(biāo)年產(chǎn)能 100 萬(wàn)臺(tái)得州超級(jí)工廠的 Optimus 機(jī)器人基地已進(jìn)入主體建設(shè)階段遠(yuǎn)期目標(biāo)年產(chǎn)能 1000 萬(wàn)臺(tái)預(yù)計(jì) 2027 年夏季投產(chǎn)。供應(yīng)鏈端特斯拉已于近期下發(fā)具體零件采購(gòu)指引8 月已現(xiàn)數(shù)百臺(tái)訂單EDI 訂單進(jìn)入自動(dòng)釋放狀態(tài)正式采購(gòu)訂單PO已下發(fā)供應(yīng)鏈預(yù)期 9-10 月周產(chǎn)水平可達(dá) 1000-2000 臺(tái)級(jí)別。7 月底至 8 月初弗里蒙特工廠已啟動(dòng)小批量 SOP 量產(chǎn)采用按月滾動(dòng)訂單模式。特斯拉 Optimus 量產(chǎn)從指引階段正式進(jìn)入訂單驗(yàn)證階段三季度有望成為量產(chǎn)數(shù)據(jù)集中驗(yàn)證窗口。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.來(lái)源 / SourcesOptimus 產(chǎn)線進(jìn)場(chǎng)高盛聚焦高技術(shù)壁壘核心部件 - 財(cái)聯(lián)社特斯拉弗里蒙特工廠首條 Optimus 專用產(chǎn)線正在安裝 - 路透社特斯拉 Optimus 量產(chǎn)進(jìn)入訂單驗(yàn)證階段 - 第一財(cái)經(jīng)4、微軟計(jì)劃 9 月發(fā)布 Maia 300 AI 芯片與臺(tái)積電洽談 30 萬(wàn)顆產(chǎn)能事件概要2026 年 8 月 10 日-11 日據(jù) The Information 援引兩位知情人士消息微軟計(jì)劃最早于 2026 年 9 月公開(kāi)發(fā)布其下一代 Maia 300 人工智能加速器并已與臺(tái)積電就超過(guò) 30 萬(wàn)顆的產(chǎn)能達(dá)成協(xié)議預(yù)計(jì) 2027 年開(kāi)始交付。微軟同時(shí)希望說(shuō)服 Anthropic 等大型云服務(wù)客戶采用 Maia 系列芯片作為對(duì)英偉達(dá)昂貴 GPU 替代方案的延展。微軟首批自研 AI 芯片 Maia 100 于 2023 年 11 月推出但在大規(guī)模部署上持續(xù)落后于競(jìng)爭(zhēng)對(duì)手新一代 Maia 300 被視為微軟自研 AI 芯片戰(zhàn)略的關(guān)鍵拐點(diǎn)。受該消息影響微軟股價(jià)在消息確認(rèn)日盤前交易中上漲約 0.7%。臺(tái)積電是此次供應(yīng)鏈的直接受益方。微軟 Azure Maia 總經(jīng)理 Andrew 的表態(tài)是從長(zhǎng)期看公司目標(biāo)是將 Maia 系列芯片的部署規(guī)模擴(kuò)大至吉瓦GW級(jí)別。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.來(lái)源 / SourcesMicrosoft plans to release next-gen Maia 300 AI chip in September - The Information微軟最快 9 月發(fā)布 Maia 300 芯片與臺(tái)積電洽談 30 萬(wàn)顆產(chǎn)能 - IT之家微軟 Maia 300 今秋亮相目標(biāo)產(chǎn)能 30 萬(wàn)片 - 財(cái)聯(lián)社5、愛(ài)芯元智下一代大算力 AI 芯片完成流片支持 2/4 芯級(jí)聯(lián)滿血大模型推理事件概要2026 年 8 月 12 日界面新聞從愛(ài)芯元智 2026 年中期業(yè)績(jī)電話會(huì)上獲悉公司下一代高性能大算力 AI 芯片已完成流片算力規(guī)格較上一代大幅提升同時(shí)配備高帶寬內(nèi)存HBM并支持兩芯或四芯級(jí)聯(lián)可實(shí)現(xiàn)滿血大模型在邊緣側(cè)的高性能推理。該系列被命名為元曦系列大算力加速卡以單芯片或多芯組合形態(tài)提供算力能夠?qū)崿F(xiàn)滿血大模型高性能推理尤其擅長(zhǎng)超長(zhǎng)上下文處理、高清圖像視頻解析等任務(wù)可在邊緣側(cè)實(shí)時(shí)、就近數(shù)據(jù)源完成推理。愛(ài)芯元智同期公布的上半年業(yè)績(jī)顯示公司 2026 年上半年實(shí)現(xiàn)營(yíng)業(yè)收入 4.02 億元同比增長(zhǎng) 181.8%毛利率提升至 29%其中邊緣 AI 業(yè)務(wù)同比增長(zhǎng) 251.9%已成為業(yè)績(jī)?cè)鲩L(zhǎng)引擎新興業(yè)務(wù)智能汽車 邊緣 AI收入占比由 11.3% 升至 14.1%。愛(ài)芯元智也成為 8 月 9 日披露上市后首份半年報(bào)的港股 AI 芯片企業(yè)。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.來(lái)源 / Sources愛(ài)芯元智下一代大算力 AI 芯片完成流片 - 界面新聞愛(ài)芯元智 2026 上半年?duì)I收同比增長(zhǎng) 181.8% - 21 世紀(jì)經(jīng)濟(jì)報(bào)道邊緣 AI 業(yè)務(wù)增長(zhǎng) 251.9%愛(ài)芯元智構(gòu)建邊-車-端全棧算力生態(tài) - 財(cái)聯(lián)社6、高盛上調(diào)人形機(jī)器人出貨預(yù)期2027 年 7.6 萬(wàn)臺(tái)、2032 年 50.2 萬(wàn)臺(tái)事件概要2026 年 8 月 12 日高盛最新研究報(bào)告顯示隨著 AI 計(jì)算基礎(chǔ)設(shè)施投資持續(xù)擴(kuò)張AI 投資主題正從算力競(jìng)賽轉(zhuǎn)向終端應(yīng)用人形機(jī)器人有望成為下一階段增長(zhǎng)引擎。高盛預(yù)計(jì)全球人形機(jī)器人出貨量將在 2027 年達(dá)到 7.6 萬(wàn)臺(tái)并于 2032 年升至 50.2 萬(wàn)臺(tái)大規(guī)模部署將在 2027-2029 年逐步展開(kāi)。報(bào)告認(rèn)為勞動(dòng)力短缺和自動(dòng)化需求將為機(jī)器人產(chǎn)業(yè)提供長(zhǎng)期結(jié)構(gòu)性支撐。從主題驅(qū)動(dòng)看AI 投資主線的切換信號(hào)已經(jīng)出現(xiàn)算力板塊經(jīng)過(guò)近兩年的高速擴(kuò)張進(jìn)入訂單驗(yàn)證階段而人形機(jī)器人賽道從故事走向業(yè)績(jī)兌現(xiàn)。近期特斯拉 Optimus 量產(chǎn)進(jìn)入訂單驗(yàn)證階段、谷歌 DeepMind 發(fā)布新一代具身智能大模型、得州超級(jí)工廠 Optimus 基地進(jìn)入主體建設(shè)疊加國(guó)內(nèi)宇樹(shù)科技科創(chuàng)板 IPO 已于 8 月 10 日啟動(dòng)申購(gòu)構(gòu)成本輪人形機(jī)器人主題的核心催化組合。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.來(lái)源 / Sources高盛人形機(jī)器人出貨量 2027 年 7.6 萬(wàn)臺(tái)、2032 年 50.2 萬(wàn)臺(tái) - 財(cái)聯(lián)社Humanoid robot shipments to accelerate, Goldman Sachs projects - BloombergAI 主題切換高盛唱多 Optimus 量產(chǎn) - 第一財(cái)經(jīng)7、Stripe 聯(lián)合創(chuàng)始人加碼 AIRiver AI 完成 11 億美元融資英偉達(dá)、AMD 戰(zhàn)略投資事件概要2026 年 8 月 11 日由 xAI 聯(lián)合創(chuàng)始人 Igor Babuschkin此前亦為 Stripe 主要工程師創(chuàng)辦的全棧 AI 公司 River AI 宣布完成 11 億美元融資。本輪由 General Catalyst 和 AMP PBC 領(lǐng)投英偉達(dá)和 AMD Ventures 作為戰(zhàn)略投資方參與Y Combinator、淡馬錫等基金跟投。融資資金將用于建設(shè)開(kāi)放 AI 技術(shù)棧并為開(kāi)發(fā)者和企業(yè)提供訓(xùn)練、微調(diào)及部署自有 AI 模型的工具。River AI 稱其 API 可在無(wú)需專門基礎(chǔ)設(shè)施團(tuán)隊(duì)的情況下于 15-20 分鐘內(nèi)完成復(fù)雜的強(qiáng)化學(xué)習(xí)訓(xùn)練成本較閉源替代方案低 2-4 倍。從投資人結(jié)構(gòu)看本輪標(biāo)志著 GPU 龍頭英偉達(dá)、AMD首次以同時(shí)戰(zhàn)略投資的方式押注同一家全棧 AI 創(chuàng)業(yè)公司意味著 GPU 廠商對(duì)開(kāi)放生態(tài)的爭(zhēng)奪正在升溫。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.來(lái)源 / SourcesRiver AI raises 1.1 billion USD with NVIDIA and AMD backing - BusinessWireStripe 聯(lián)合創(chuàng)始人加碼 AIRiver AI 完成 11 億美元融資 - 財(cái)聯(lián)社英偉達(dá)、AMD 罕見(jiàn)同時(shí)出手押注 xAI 聯(lián)合創(chuàng)始人新項(xiàng)目 - TechCrunch8、上海市經(jīng)信委印發(fā)《十五五軟件和信息服務(wù)業(yè)規(guī)劃》明確 GPU/NPU/QPU/HBM 協(xié)同方向事件概要2026 年 8 月 11 日上海市經(jīng)濟(jì)和信息化委員會(huì)印發(fā)《上海市軟件和信息服務(wù)業(yè)發(fā)展十五五規(guī)劃》提出攻堅(jiān)下一代模型架構(gòu)推動(dòng)基于狀態(tài)空間模型、循環(huán)神經(jīng)網(wǎng)絡(luò)變體、液態(tài)神經(jīng)網(wǎng)絡(luò)等非 Transformer 架構(gòu)的多技術(shù)路線探索加快布局物理智能、世界模型、量子智能、類腦智能等前沿基礎(chǔ)模型技術(shù)體系。規(guī)劃同時(shí)明確攻關(guān)超大規(guī)模智能算力集群組網(wǎng)技術(shù)圍繞高性能計(jì)算芯片GPU/NPU、量子芯片QPU、高速光互連CPO、高帶寬內(nèi)存HBM及異構(gòu)服務(wù)器等核心環(huán)節(jié)提升智算硬件的供給能力推動(dòng)自主芯片與主流大模型深度融合。從產(chǎn)業(yè)落地看該規(guī)劃延續(xù)了上海模塑申城工程方向上海數(shù)據(jù)港、網(wǎng)宿科技、云賽智聯(lián)、寶信軟件等本地算力鏈上市公司被券商視為規(guī)劃落地的直接受益方。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.來(lái)源 / Sources上海市經(jīng)信委印發(fā)《十五五軟件和信息服務(wù)業(yè)規(guī)劃》 - 財(cái)聯(lián)社上海十五五規(guī)劃攻關(guān) GPU/NPU/QPU/HBM 核心環(huán)節(jié) - 第一財(cái)經(jīng)上海算力建設(shè)相關(guān)受益方梳理 - 界面新聞行業(yè)數(shù)據(jù)速覽指標(biāo)數(shù)據(jù)備注DeepSeek V4-Plus 輸入價(jià)格3.5 元/百萬(wàn) tokens較前次漲價(jià)約 60%DeepSeek V4-Plus 輸出價(jià)格12 元/百萬(wàn) tokens首日調(diào)用量仍增長(zhǎng) 18%Kimi K3.1 1M 上下文推理成本下降 22%視頻原生多模態(tài)面壁智能半年累計(jì)融資超 50 億元估值突破 200 億元愛(ài)芯元智 H1 營(yíng)收4.02 億元同比增長(zhǎng) 181.8%愛(ài)芯元智邊緣 AI 業(yè)務(wù)增速251.9%新興業(yè)務(wù)占比 14.1%微軟 Maia 300 芯片產(chǎn)能30 萬(wàn)顆2027 年交付與臺(tái)積電合作River AI 融資規(guī)模11 億美元NVIDIA、AMD 戰(zhàn)略投資高盛 2027 年人形機(jī)器人出貨預(yù)期7.6 萬(wàn)臺(tái)全球高盛 2032 年人形機(jī)器人出貨預(yù)期50.2 萬(wàn)臺(tái)2032 年特斯拉 Optimus 弗里蒙特目標(biāo)年產(chǎn)能100 萬(wàn)臺(tái)產(chǎn)線正在安裝特斯拉 Optimus 得州遠(yuǎn)期目標(biāo)年產(chǎn)能1000 萬(wàn)臺(tái)2027 年夏季投產(chǎn)媒體人解讀2026 年 8 月 12 日的 AI 產(chǎn)業(yè)圖景可以用兩升一降來(lái)概括升的是廠商按價(jià)值定價(jià)的底氣與終端硬件的出貨定價(jià)降的是模型迭代敘事的邊際吸引力。兩條線最值得讀者關(guān)注第一DeepSeek V4-Plus 漲價(jià)首日調(diào)用量仍增長(zhǎng) 18%配合阿里 Qwen重度商用收入分成機(jī)制、Kimi K3.1 多模態(tài)升級(jí)、智源 FlagOS 2.0 同日落地國(guó)產(chǎn)大模型已經(jīng)從以價(jià)換量走向按價(jià)值定價(jià)。對(duì)開(kāi)發(fā)者而言這意味著 Agent 化應(yīng)用的 TCO 測(cè)算在 Q4 前必須重做對(duì)企業(yè) IT 而言自建 商用 API組合策略需要重新校準(zhǔn)權(quán)重。第二硬件端的中國(guó)制造開(kāi)始進(jìn)入兌現(xiàn)期。微軟 Maia 300 9 月發(fā)布、與臺(tái)積電 30 萬(wàn)顆產(chǎn)能意味著北美云廠的自研芯片替代英偉達(dá)從 PPT 走向量產(chǎn)愛(ài)芯元智下一代大算力 AI 芯片完成流片、2/4 芯級(jí)聯(lián)實(shí)現(xiàn)邊緣滿血大模型推理疊加上海十五五規(guī)劃明確 GPU/NPU/QPU/HBM 協(xié)同國(guó)產(chǎn)算力鏈從政策預(yù)期走向產(chǎn)品兌現(xiàn)2026 H2-2027 H1 是關(guān)鍵驗(yàn)證窗口。第三人形機(jī)器人賽道的訂單驗(yàn)證窗口已經(jīng)開(kāi)啟。特斯拉弗里蒙特工廠首條 Optimus 專用產(chǎn)線進(jìn)場(chǎng)、目標(biāo)年產(chǎn) 100 萬(wàn)臺(tái)高盛上調(diào) 2027 年 7.6 萬(wàn)臺(tái)、2032 年 50.2 萬(wàn)臺(tái)出貨預(yù)期特斯拉同步下發(fā)的零件采購(gòu)訂單已經(jīng)進(jìn)入自動(dòng)釋放狀態(tài)。三季度財(cái)報(bào)與供應(yīng)鏈訂單數(shù)據(jù)將是該賽道從故事到業(yè)績(jī)切換的關(guān)鍵驗(yàn)證點(diǎn)。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.參考來(lái)源英文來(lái)源Microsoft plans to release next-gen Maia 300 AI chip in September - The InformationRiver AI raises 1.1 billion USD with NVIDIA and AMD backing - BusinessWireGoldman Sachs raises humanoid robot shipment forecast - BloombergRiver AI secures 1.1 billion USD with NVIDIA and AMD backing - TechCrunch中文來(lái)源面壁智能啟動(dòng) A 股上市輔導(dǎo)估值突破 200 億元 - 36氪面壁智能啟動(dòng) IPO - IT之家DeepSeek V4-Plus 漲價(jià)方案落地 - 財(cái)聯(lián)社Kimi K3.1 多模態(tài)升級(jí) - 鈦媒體特斯拉 Optimus 產(chǎn)線進(jìn)場(chǎng) - 財(cái)聯(lián)社微軟 Maia 300 9 月發(fā)布 - 財(cái)聯(lián)社愛(ài)芯元智下一代大算力 AI 芯片完成流片 - 界面新聞上海十五五軟件和信息服務(wù)業(yè)規(guī)劃 - 財(cái)聯(lián)社高盛人形機(jī)器人出貨預(yù)期 - 第一財(cái)經(jīng)本文內(nèi)容綜合整理自上述公開(kāi)信息來(lái)源媒體人解讀僅代表個(gè)人觀點(diǎn)僅供參考。作者署名劉一說(shuō) ?發(fā)布日期2026年8月12日