25位菲尔兹奖得主警告：AI正毁掉数学
9月12日，陶哲轩等 25 位菲尔兹奖得主联合签署公开信，警告AI公司与数学共同体之间正在出现“严重错位”。他们认为，AI公司正在把解决数学问题作为衡量模型能力的基准，而数学共同体真正追求的，并不是尽可能快地得到一道题的答案。AI以越来越快的速度批量生产“真/假”结论，可能不是在孕育新的思想，反而可能摧毁孕育思想的沃土。此前9月8日，OpenAI宣布其一款未公开内部模型的计算能力已解答千禧年大奖难题之一的纳维斯托克斯方程存在性与光滑性问题，并发布了166页的论文。
【点评】这份声明并没有主张拒绝AI。数学家们承认AI具有加强和加速真正数学研究与理解的潜力，数学这一职业也需要适应由此带来的变化。真正尚未确定的是，这种变化最终会推动数学发展，还是产生破坏性影响，而结果在很大程度上取决于掌握这项技术的人如何作出选择。
DeepSeek发布V4.1 Flash，最高降价60%
9月10日，经过两天内测，DeepSeek正式发布V4.1 Flash模型，“能力更强、速度更快、且成本更低”。据介绍，V4.1 Flash是DeepSeek全新模型结构系列中尺寸最小的一款，具备原生多模态视觉理解能力。这一系列架构的设计目标是：提高能力上限和推理速度，扩大吞吐量，同时支持向更大参数规模扩展。除了架构变化，V4.1 Flash还采用了新的预训练方式，并经过更大规模的强化学习后训练。由于技术上的创新，V4.1 Flash能以更低的成本服务用户，DeepSeek还公开下调了定价。9月10日12:00起，Flash系列模型的最新定价为：空闲时段输入缓存命中单价0.02元、输入缓存未命中单价1元，输出单价4元；高峰时段价格为空闲时段价格的2倍。相比之前，缓存命中输入的价格降幅达到60%。
【点评】对普通用户而言，用DeepSeek聊天最直观的变化是生成速度更快了；对开发者而言，则是调用成本进一步下降。凭借性能提升与价格下探，DeepSeek巩固了其在高性价比模型市场中的位置。
黄仁勋称网络安全可能成为AI下一个主要应用场景
在当地时间9月10日的旧金山高盛科技会议上，英伟达CEO黄仁勋表示，网络安全很可能成为AI的下一个主要应用场景。他认为，AI加速自动化编程的同时，也加速了代码被攻击和漏洞被利用的速度，企业对AI安全工具的需求将随之增长。“有什么比制造问题更能创造需求的方式呢？”他表示。
【点评】近一周还有多名海外模型头部公司科研人员密集发声，强调AI安全风险已从理论走向现实。例如，Anthropic对齐科学负责人Evan Hubinger表示，他个人认为未来十年内AI杀死所有人类的概率超过10%。不同业内人士对AI安全问题的关注层面不同，当前AI已经能被用于网络攻击，用AI守护网络安全是一种应对方式。随着AI能力演进，未来还有可能导致其他安全问题。
诺奖得主称AI有望使生产率增速每年提高1.08个百分点
“AI有望在未来十年使生产率增速每年额外提高约1.08个百分点，潜在影响或超过IT革命。”9月10日，在2026Inclusion·外滩大会上，诺贝尔经济学奖得主、法国经济学家菲利普·阿吉翁表示，仅考虑AI对商品和服务生产中任务的自动化，未来十年，AI就有望使生产率增速每年额外提高0.68个百分点。0.68个百分点只是一个下限，如果进一步考虑AI促进新想法产生的作用，生产率增速每年还可能额外提高至少约0.4个百分点，合计潜在提升约1.08个百分点。
【点评】阿吉翁的预计建立在一个前提下：竞争政策、教育体系和劳动力市场制度同步跟上，使技术潜力转化为持续增长。与此同时，要警惕AI领域集中度过高导致的垄断，垄断有可能压制该领域的创新。阿吉翁提到的另外一点也非常重要：需要良好的劳动力市场保障，帮助劳动者更顺利地进入新的工作。
燧原科技登陆科创板
9月11日，AI芯片厂商燧原科技登陆科创板，上市首日收涨179.22%。2023年~2025年，燧原科技年营收分别为3.01亿元、7.2亿元、9.9亿元，营收持续增长，净亏损则分别为16.65亿元、15.1亿元、11.64亿元。2026年上半年，该公司营收11.2亿元，同比增长279.08%，亏损6.32亿元。
【点评】国产AI芯片上市公司再增一员。燧原科技尚未实现扭亏，去年营收规模在国产AI芯片公司中偏低。今年上半年营收规模低于摩尔线程、壁仞科技和沐曦股份，高于天数智芯。对大客户的绑定是燧原科技业务结构的一大特点，客户集中度偏高带来了一定的风险。燧原科技第一大股东腾讯科技及关联方同时是燧原科技的大客户。2025年该公司来自腾讯的收入占比达83.79%。
软件股剧烈波动
当地时间9月8日，美股结束劳动节假期恢复交易后，Salesforce、Intuit均下跌约4%，ServiceNow跌约5%，标普500软件与服务指数下跌1.4%，连续两个交易日走低。触发市场新一轮担忧的，是OpenAI上周发布的GPT-6 Astra，更强的AI Agent可能进一步削弱甚至替代传统软件的部分价值。
【点评】不到两周前，市场还刚刚因为AI商业化进展重新买入软件股。Salesforce发布财报后盘后一度上涨14%，Snowflake最新财报后更上涨超过20%。软件股近期的剧烈波动，反映出市场还没有找到答案。AI已经开始进入部分软件公司的财务指标。而它最终成为传统软件新的增长点，还是重新改写整个SaaS行业的收入结构，现在才刚刚进入可以用数字验证的阶段。
美国科技巨头加速布局电力
9月9日，谷歌宣布未来两年在芬兰投资130亿欧元建设AI基础设施，并签下一份长达22年的核电协议；微软被曝计划到2032年将全球数据中心容量扩大至约38GW，超过目前水平的三倍；英伟达则与澳大利亚多家数据中心运营商等合作，计划到2027年推动最高2GW的AI基础设施容量，而澳大利亚目前全部数据中心容量约为1.6GW。
【点评】这些项目背后，AI公司不再被动等待公共电网扩容，并且越来越早地把电源、储能和数据中心放在一起规划。AI竞争也从模型、芯片和数据中心，继续延伸到电厂、电网和能源融资。大模型版本迭代以月计，GPU全新架构迭代周期约 1-2年，而新的输电和发电设施往往需要数年建设，电力正成为AI扩张中最难加速的一环。
机构预测AI主机将与PC、服务器并列三大算力品类
9月9日，联想与IDC发布《AI主机白皮书》预测，国内AI主机产业将经历三个阶段：2026年为品类导入元年，全年出货量约40万台、对应市场规模约50亿元；2027-2028 年为早期渗透期，AI OPC群体与中小企业逐步采购，市场仍处品类教育期；2029-2030 年市场进入加速爬坡期，出货量从约350万台攀升至约700万台、对应市场规模从约460亿元跃升至约770亿元。
【点评】企业推进AI应用普遍面临“不敢用、不会用、用不起、不好用”四大痛点，担心核心数据外流且缺乏专业团队。AI主机走向产业普及的关键，在于是否能在商业逻辑上算清投入产出比与安全合规这两笔账。
谷歌用AI破解罕见病
当地时间9月8日，谷歌DeepMind发布了人类基因组预测平台AlphaGenome Atlas，可以理解为，这是一张预测人类基因组中所有可能的DNA单字母变化的图谱，能帮助研究人员理解尚未破解的罕见病等问题，并加速生物学研究。根据谷歌DeepMind发布的博客，AlphaGenome Atlas预先计算了人类基因组中约90亿种单核苷酸变异可能产生的分子影响，并将结果整理成一个规模达1PB的数据集。谷歌称，这是目前关于基因突变如何影响分子生物学最全面的预测目录，其数据体量是AlphaFold数据库的30多倍。
【点评】凭借AlphaFold拿到诺贝尔奖之后，谷歌的AI for Science版图开始向更多环节延伸。研究人员认为，人类基因组是一个极其庞大的搜索空间，要找到一个对特定疾病或性状真正重要的遗传变异，就像大海捞针，而谷歌的AlphaGenome Atlas的价值在于“缩小这片大海”。
摩尔线程限售股解禁当天股价大跌
摩尔线程9月7日跌20%，当天该公司2577.45万股首次公开发行网下配售限售股迎来解禁。摩尔线程回应称，公司总流通股扩大85%，公司生产经营持续向好。另一家国产GPU厂商沐曦股份也将于9月17日迎来1396.6万股首发机构配售股解禁，总流通股将扩大75%。
【点评】这批限售股解禁前，摩尔线程流通股数仅为3022.55万股，解禁后这一数字升至约5600万股。如果按9月4日收盘价计算，这批解禁的限售股市值达133.9亿元。摩尔线程还将于12月7日迎来1.86亿股首发股东限售股份解禁，规模更大。除了流通盘扩容压力，国产GPU厂商的长期走势，还取决于基本盘成色。
多家厂商在光博会展出NPO方案
9月9日至11日的中国国际光电博览会（下称“光博会”）上，多家光互连供应链厂商展出NPO（近封装光学）相关产品。华工科技子公司华工正源用展板介绍了公司3.2T NPO的设计。东山精密旗下索尔思光电展出了6.4T NPO模组。中际旭创子公司智禾光通展出了3.2T NPO产品。Coherent则展出了NPO、CPO（共封装光学）和XPO（超高密度可插拔光模块）可用的光子集成电路。
【点评】NPO方案在此次光博会上尤其惹眼，体现了数据中心互连方案从铜向光演进的趋势。要提升光互连性能，业界正在尝试各种方案，期间产业链存在持续的分歧和博弈。因CPO系统集成度高，有能力做CPO的厂商不多，一些产业链环节对CPO方案有所“忌惮”。近期NPO方案的火热，正是产业链博弈在现阶段的产物。
本文来源：https://www.163.com/dy/article/L6MULS6D0519DDQ2.html
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
25 Fields Medalists Warn: AI is Destroying Mathematics
On September 12, 25 Fields Medalists, including Terence Tao, signed an open letter warning of a "dangerous disconnect" between AI companies and the mathematical community. They argue that while AI firms use solving math problems as a benchmark for model power, mathematicians seek deep understanding rather than speed. The rapid production of "true/false" outcomes by AI risks stifling innovative thought in the field. This follows OpenAI's September 8 announcement of a model that solved the Navier-Stokes equation existence/smoothness problem with a 166-page proof.
The statement does not reject AI but acknowledges its potential to enhance and accelerate genuine mathematical research if used correctly. The outcome depends on how those wielding this technology choose to apply it.
DeepSeek Releases V4.1 Flash with Up to 60% Price Cut
On September 10, after two days of internal testing, DeepSeek launched V4.1 Flash, promising better performance, faster speed, and lower cost. This model, based on a new architecture series, features native multimodal vision understanding and was trained with a novel approach followed by extensive reinforcement learning. The lower computational cost has allowed DeepSeek to reduce pricing significantly — from September 10, cache-hit input rates drop by up to 60%, with peak and off-peak pricing structures updated accordingly.
For users, this means faster response times and lower operational costs for developers — consolidating DeepSeek's position in the affordable model market.
Jensen Huang Says Cybersecurity Could Be AI's Next Major Application
At the Goldman Sachs Technology Conference on September 10, NVIDIA CEO Jensen Huang suggested cybersecurity will become a major AI application area as AI accelerates both code generation and attack capabilities. Enterprises will increasingly need AI-based security tools to defend against automated threats.
This aligns with recent warnings from top AI researchers about real-world security risks — including Evan Hubinger's estimate that AI could kill all humans within a decade. While AI can be used for attacks, it also offers defense mechanisms against them.
Nobel Laureate Claims AI Could Boost Productivity Growth by 1.08% Annually
Philippe Aghion, Nobel-winning economist, stated on September 10 that AI could add ~1.08% to annual productivity growth over the next decade — surpassing the impact of the IT revolution. He breaks this into two components: a direct boost from automation (~0.68%) and an additional ~0.4% from AI-driven innovation. This potential is contingent on supportive competition policies, education, and labor market reforms.
If these conditions are met, AI could significantly enhance global productivity; without them, benefits may be limited by monopolistic concentration or workforce displacement.
Moore Threads Faces Stock Plunge After Lockup Expiry
On September 7, Moore Threads' shares dropped 20% following the expiry of 25.77 million initial public offering shares under lockup. The company responded that operations remain healthy, but the sharp increase in float (85% growth) likely triggered selling pressure. Similar dynamics may affect MuXi on September 17 when 13.97 million restricted shares unlock — raising questions about how these GPU makers will manage dilution and market expectations.
Despite strong revenue growth — particularly in 2024H1 where sales rose 279% — losses persist at ~6.32 billion CNY, indicating profitability remains elusive for now.
Software Stocks Waver Amid AI Agent Threat
On September 8, Salesforce, Intuit, and ServiceNow all fell by ~4–5% as investors re-evaluate software companies' value in light of emerging AI agents. OpenAI's release of GPT-4 Astra suggests large language models are evolving into autonomous decision-makers, potentially reducing demand for traditional SaaS platforms. While AI integration has driven stock rallies recently, its long-term effect on industry structure is uncertain — some software companies may thrive as AI enablers, while others risk obsolescence.
The volatility reflects this ambiguity: does AI amplify software's role, or replace it? Early data shows mixed results, with some firms embedding AI successfully while others struggle to monetize.
Tech Giants Accelerate Power Plant Acquisitions for AI
Google, Microsoft, and NVIDIA are aggressively securing energy supply to support expanding AI operations. Google announced $13.2 billion in Finnish investments plus a 22-year nuclear power deal; Microsoft plans to triple its global data center capacity to 38 GW by 2032; NVIDIA partners with Australian energy firms for 2 GW of AI infrastructure. These moves highlight that electricity availability is becoming a bottleneck for rapid AI deployment — companies are vertically integrating energy procurement to ensure reliability and cost control.
This trend underscores that scaling AI isn't just about computing power — it's equally about having enough electricity to run those computers efficiently.
Agencies Predict AI PCs Will Rival Servers and Laptops
IDC and Lenovo report that "AI workstations" will become a third major computing category alongside servers and laptops, driven by demand for local inferencing and data privacy. In 2024, such devices — including NVIDIA Ising-ANNU and Intel Xeon/NVIDIA combos — will account for ~400k units/¥50B, rising to ~7M units/¥770B by 2030. This shift is fueled by growing awareness of cloud-based AI risks and the need for enterprise-level computing power in edge devices.
The success of this market depends on convincing businesses that on-prem AI solutions are both secure and cost-effective compared to cloud alternatives — a narrative being pushed aggressively by manufacturers.
Google Uses AI to Crack Rare Disease Codes
On September 8, DeepMind launched the AlphaGenome Atlas, a massive dataset predicting molecular effects of genetic variations — effectively "reading" DNA faster than ever. While not a diagnostic tool, it helps researchers identify candidate genes for rare diseases or traits, speeding up experimental workflows. With 90+ billion predictions covering 90% of the human genome, it's a key step toward personalized medicine and automated genetic discovery.
This follows DeepMind's Nobel Prize for AlphaFold and expands its impact from protein folding to genome-wide analysis, highlighting AI's role in accelerating biological research.
NPO Solutions Shine at Optical Expo
At the 2024 China International Optoelectronics Expo, multiple firms showcased Near-Package Optics (NPO) modules as an alternative to Co-Packaged Optics (CPO). Companies like Wuhan Fiber Optics and Solis Optical displayed 3.2T and 6.4T NPO solutions, indicating rising interest in this hybrid approach that offers easier maintenance and lower costs than fully integrated CPO systems.
This reflects broader industry debate over optimal optical interconnects for high-performance computing — NPO may bridge the gap between traditional pluggable modules and future co-packaging, depending on cost, scalability, and thermal management improvements.
Multiple vendors showcase NPO solutions at Photonics Expo
At the China International Optoelectronic Expo (referred to as the "Photonics Expo") held from September 9th to 11th, several optical interconnection supply chain vendors displayed NPO (Near-Package Optics) related products. Hualan Group's subsidiary, Hualan Zhengyuan, used a display board to introduce the company's 3.2T NPO design. Dongshan Precision's subsidiary, Solace Photonics, exhibited a 6.4T NPO module. Zhongji Xuchuang's subsidiary, Zhihe Guotong, showcased a 3.2T NPO product. Coherent displayed photonic integrated circuits suitable for NPO, CPO (Compact Package Optics), and XPO (Ultra-High Density Pluggable Transceiver) applications.
[Comment] The NPO solution was particularly eye-catching at this Photonics Expo, reflecting the trend of data center interconnections evolving from copper to optical. To improve optical interconnection performance, the industry is experimenting with various solutions, during which the supply chain is experiencing continuous disagreement and negotiation. Because CPO has high system integration, there are not many vendors capable of producing it, and some parts of the supply chain are "wary" of the CPO solution. The current popularity of NPO solutions is the result of this supply chain negotiation at this stage.