Workflow
硬AI
icon
Search documents
摩根大通专家访谈:AI数据中心“产能过剩”了吗?训练和推理基建如何部署?
硬AI· 2025-06-19 15:49
摩根大通最新专家访谈揭示,AI基建"产能过剩"担忧为时过早,算法轻量化与硬件循环利用正缓解算力焦虑,但数据中心 头顶的"电力问题"与"散热难题",才是AI狂奔路上更现实的减速带。 硬·AI 作者 | 龙 玥 编辑 | 硬 AI 近期,摩根大通与Scale AI数据科学家、Meta前高级数据科学家Sri Kanajan举行电话会议,深入探讨超大 规模AI数据中心架构趋势。 据摩根大通报告,近期算法突破——如混合模型(含DeepSeek)、精度训练及策略性强化学习——显著 降低了整体AI模型训练所需的计算量。这促使行业将优化重点转向推理环节。 Kanajan指出,当前,业界正积极采用模型蒸馏、压缩等技术精炼模型,力求在不大幅增加原始算力投入 的前提下提升性能。 02 基础设施: 动态部署,担忧产能过剩尚早 Kanajan认为,AI基础设施部署仍处早期阶段,特别是考虑到云服务商对其投资的长期回报预期,当前对 产能过剩的担忧有限。 Kanajan认为,AI基础设施部署仍处于早期阶段,对产能过剩的担忧有限。算法进步正降低训练算力消 耗,基础设施通过"训练转推理"实现高效循环利用,训练集群在新一代GPU推出后被快速重新配 ...
以退为进?微软“放话”:与OpenAI谈不好,那就不谈了
硬AI· 2025-06-19 15:49
据报道,知情人士透露,如果双方无法就关键问题达成一致,微软已考虑直接终止复杂的谈判进程。这些关键分歧包括微 软在OpenAI未来结构中应占多大股份等核心利益分配问题。一旦谈判破裂,微软将依靠现有商业合同保持对OpenAI技术 的访问权直至2030年。 6月18日,据报道,知情人士透露,如果双方无法就关键问题达成一致,微软已考虑直接终止复杂的谈判 进程。这些关键分歧包括微软在OpenAI未来结构中应占多大股份等核心利益分配问题。 一旦谈判破裂,微软将依靠现有商业合同保持对OpenAI技术的访问权直至2030年,除非OpenAI能提供等 同或更优的条件安排。一位接近微软的人士表示: "现状对微软来说是可以接受的,公司对当前合同很满意",准备执行到2030年。" 目前的合同条款对微软极为有利:独家销售OpenAI模型的权利,以及在920亿美元收入上限内获得20%的 收入分成。 01 硬·AI 作者 | 董 静 编辑 | 硬 AI 在与OpenAI就未来数百亿美元合作关系的谈判中,软件巨头微软公开表态:如果谈不拢,那就直接走 人。这似乎正印证了那句话:有时候最好的谈判策略就是随时准备离开谈判桌。 过去一年中,双方围绕 ...
Meta智能眼镜军团扩编,Oakley和Prada加入战局
硬AI· 2025-06-18 15:01
Core Viewpoint - Meta is strategically expanding its presence in the smart glasses market by collaborating with luxury brand EssilorLuxottica to launch AI smart glasses under the Oakley and Prada brands, aiming to capture key consumer segments before competitors like Google and Snap enter the market [2][8]. Group 1: Product Launch and Pricing - Meta plans to release Oakley-branded AI smart glasses priced around $360, which is higher than similar Ray-Ban products, featuring enhanced weather resistance [1][2]. - The collaboration with Prada marks Meta's first venture into high-end fashion wearables, leveraging Prada's existing licensing agreement with EssilorLuxottica [3]. Group 2: Market Strategy and Target Audience - The target demographic for the Oakley version is specifically athletes, as Meta observed significant usage of Ray-Ban smart glasses in sports activities like tennis and skiing [4]. - Meta's expansion is backed by the unexpected success of the second-generation Ray-Ban Meta smart glasses, which have sold 2 million units since their release in 2023 [5]. Group 3: Competitive Landscape - The smart glasses market is becoming increasingly competitive, with major tech companies like Alphabet and Snap planning their own smart glasses releases, prompting Meta to adopt a proactive strategy [8]. - Meta's partnership with EssilorLuxottica, valued at $5 billion, provides exclusive rights to smart glasses technology, enhancing its competitive edge [6].
Sam Altman最新访谈:AI将发现新科学,未来AI伴侣无处不在,人形机器人街头漫步
硬AI· 2025-06-18 15:01
人形机器人将在5-10年内走上街头,成为真正让人感受"未来已来"的标志性时刻; OpenAI未来愿景:打造无处不在的"AI伴侣",通过多种设备和平台为用户提供连续性服务; OpenAI的Sam Altman在与弟弟的坦率对话中,描绘了一个AI技术狂飙未来世界。他预测5-10年AI将发现新科学,AI伴 侣将无处不在,人形机器人将走上街头,成为真正让人感受"未来已来"的标志性时刻。对于,Meta上亿美元的人才挖角攻 势,Altman表示无需担心,公司文化差异显著。 硬·AI 作者 | 龙 玥 编辑 | 硬 AI 最近,在一场"家庭式"地访谈中,OpenAI CEO Sam Altman向弟弟Jack Altman(Lattice创始人)展现了 他对AI未来5-10年最直接的预判。这位AI科技界最受瞩目的引领者坦言,尽管推理能力的突破让O3模型已 达到博士生水平,但真正的颠覆将来自AI发现新科学的能力。更令人震撼的是他对"超级智能悖论"的担 忧:即便实现了真正的超级智能,社会可能依然变化不大,就像ChatGPT的横空出世并未根本改变人们的 生活方式。 与此同时,竞争正在白热化——Meta将OpenAI视为最大威胁, ...
为国行苹果智能做准备!阿里巴巴发布升级版Qwen3:全系适配苹果MLX架构
硬AI· 2025-06-17 14:30
这意味着,从Mac Pro、Mac Studio到Mac mini、MacBook,再到iPad,甚至内存更小的iPhone,都能轻松部署 Qwen3。 硬·AI 作者 | 李笑寅 编辑 | 硬 AI 周一,阿里巴巴通义千问宣布,正式发布基于苹果MLX框架深度优化的全部Qwen3系列模型。 此举被看作是为国行苹果智能做准备。此前有消息称,阿里巴巴将成为苹果在中国大陆的大模型合作商。 公告显示,团队将一次性全部开源32款官方Qwen3 MLX模型,每款模型都有4bit、6bit、8bit和BF16等4 种不同精度的量化版本,从而实现这些模型在iPhone、iPad,以及Mac电脑上的轻松部署,做到全场景覆 盖。 目前,Qwen3的MLX模型已在魔搭社区和Hugging Face全面开源。 硬·AI * 感谢阅读! * 转载、合作、交流请留言,线索、数据、商业合作请加微信:IngAI2023 * 欢迎大家在留言区分享您的看法, 如果您能点个并分享的话,那就太感谢啦! * 让我们一起,好奇地看世界 据官方介绍,MLX是一个开源的机器学习框架,专为苹果芯片深度适配。MLX框架可高效地训练和部署AI 大模型,被越来越多 ...
谷歌之后Meta需求爆发,ASIC明年就超英伟达GPU?
硬AI· 2025-06-17 14:30
Core Viewpoint - Meta is planning to launch several high-spec AI ASIC chips between the end of 2025 and 2026, with a total expected output of 1 to 1.5 million units, potentially surpassing NVIDIA's GPU shipments at some point in 2026 [1][2][5]. Group 1: Market Dynamics - Currently, NVIDIA holds over 80% of the AI server market value, while ASIC AI servers account for only 8-11% [4]. - By 2025, Google is expected to ship 1.5 to 2 million TPU units, and AWS's Trainium 2 ASIC is projected to reach 1.4 to 1.5 million units, compared to NVIDIA's AI GPU supply of over 5 to 6 million units [4]. - Supply chain research indicates that the combined shipment of Google and AWS's AI TPU/ASIC has reached 40-60% of NVIDIA's AI GPU shipments [5]. Group 2: Meta's MTIA Project - Meta's MTIA project is a significant case in the current ASIC wave, with the first ASIC chip, MTIA T-V1, set to launch in Q4 2025, designed by Broadcom and featuring a complex 36-layer PCB architecture [8]. - The MTIA T-V1.5, expected in mid-2026, will double the chip area and exceed NVIDIA's next-generation GPU specifications, while the MTIA T-V2 in 2027 may introduce larger CoWoS packaging and high-power rack designs [8]. Group 3: Challenges and Competition - Meta aims to achieve 1 to 1.5 million ASIC shipments by the end of 2025 to 2026, but current CoWoS wafer allocation can only support 300,000 to 400,000 units, indicating potential production bottlenecks [9]. - NVIDIA is not passive; it plans to introduce NVLink Fusion technology at the 2025 COMPUTEX, allowing seamless integration of third-party CPUs or xPUs with its AI GPUs, which is part of its strategy to maintain market share [12]. - Despite the rise of ASICs, NVIDIA remains ahead in chip computing density and interconnect technology, making it difficult for ASICs to catch up in performance [13].
AMD股价飙升10%,新一代AI芯片获分析师看好,预计GPU业务四季度反弹
硬AI· 2025-06-17 14:30
Core Viewpoint - AMD's recent product launch, including the new MI400 chip and Helios server architecture, is expected to drive growth in its GPU business, with analysts projecting a recovery starting in the second half of 2025 [1][3][5]. Product Launch and Market Impact - AMD unveiled the next-generation Instinct MI350 series chips at the "Advancing AI" event, aiming to compete with market leader NVIDIA [3]. - The Helios system, set to launch in 2026, can accommodate up to 72 MI400 chips in a single server, indicating significant potential for AI applications [3][4]. - Piper Sandler raised AMD's target price from $125 to $140, maintaining an "overweight" rating, reflecting optimism about the new product's impact [5]. Stock Performance - Following the announcement, AMD's stock price surged over 10%, reaching a high of $128 before closing at $126.39, an increase of 8.81% [6]. - The stock has experienced significant volatility, down 32% from its 52-week high of $183.96 but up 61% from its 52-week low of $78.21 [6]. Strategic Partnerships - AMD emphasized collaborations with major companies like OpenAI, Meta, Oracle, and Microsoft during the product launch [8]. - Analysts from Bank of America suggest that Amazon could be a significant future partner, as AWS was a major sponsor of the event [9].
烧钱有道、天价挖角、坐拥AI“变现利器”...Meta低位反弹40%,逼近历史高点
硬AI· 2025-06-16 15:17
Meta大幅增加AI投资不仅没有拖累回报,反而推动投资回报率达到前所未有的高度,第一季度投资达到创纪录的31%。 上周Meta将天才少年Alexandr Wang招至麾下,并一直试图用"七位数到九位数"的薪酬从谷歌和OpenAI挖掘AI人才。 硬·AI 作者 | 赵 颖 编辑 | 硬 AI 当扎克伯格宣布再次大幅提升AI投资时,华尔街用脚投票推动Meta股价逼近历史新高,其股价自4月低点已大 涨超40%。 当AI竞赛进入白热化阶段,Meta选择用真金白银说话,将2025年资本支出预测上调至720亿美元,而数据 证明了Meta"烧钱有道",该公司第一季度投资回报率达到创纪录的31%。 上周,Meta敲定了对Scale AI高达143亿美元的投资,并将其创始人Alexandr Wang招至麾下。此外,与其 他科技巨头相比,Meta在AI商业化方面拥有独特优势——广告。 而交易员们显然买账了,彭博追踪的分析师中近90%建议买入Meta,目前Meta股价距离平均目标价仅一步 之遥,股票估值已达到预期收益的24.5倍。 追踪包括亚马逊在内的AI股票的ETF自4月8日低点已上涨32%引发股市普遍反弹。在此期间,Globa ...
扫地机器人的“新物种”:我深度体验了云鲸逍遥002,答案比我想的更惊人
硬AI· 2025-06-16 15:17
以下文章来源于硬评测 ,作者专注科技产研的 硬评测 . 体验最好用好玩的AI、科技产品 点击 上方 蓝字 关注我们 这些时刻,很难不去想:这到底是人工智能,还是"人工智障"啊 ???!!! 所以,当我拿到这台云鲸逍遥0 0 2时,我的心态是审慎的。 但经过三百多个小时的"暴力测试"后,我必须承认,这次,可能真的不一样了。 地面清洁的「终极形态」可能真的被云鲸造出来了 ! 硬评测 作者 | Kozmon 编辑 | lalalunee 老实说,在体验云鲸逍遥0 0 2之前,我对"全能扫地机器人"这个词已经有点麻木了。 这些年,我们见过太多参数上的内卷:吸力从一万卷到两万,功能从扫拖一体卷到自动洗拖布、自动集尘。但一个灵魂拷问始终悬在我们这 些"懒人"用户头上: 它真的能让我完全撒手不管吗? 我在使用扫地机时,就碰到过很多让人啼笑皆非的场景:满怀期待地让机器人去处理打翻的咖啡,结果它用一块脏兮兮的拖布画出了一幅后现 代主义大地艺术画;扫着扫着,就被一根充电线"锁喉",然后无助地在原地"嘤嘤嘤"直到电量耗尽。 这玩意儿,你不能再用"扫拖机器人"的旧眼光去看它。它更像是一个被施了魔法的物种: 云鲸把一台专业的「手持洗地 ...
应对谷歌挑战,亚马逊AWS紧急重构AI云服务
硬AI· 2025-06-13 10:56
Core Viewpoint - AWS is facing significant pressure from competitors like Microsoft and Google, leading to plans for a comprehensive upgrade of its AI platform "Bedrock" to retain customers and market share [1][2][5][15]. Group 1: AWS's Current Challenges - AWS is experiencing customer attrition due to the more flexible and user-friendly AI offerings from Google Cloud and Microsoft Azure, particularly in developing AI agents [2][5]. - The AI agent applications are resource-intensive, consuming significant computing power and tokens, making them a lucrative area for cloud service providers [1][14]. - AWS's Bedrock platform currently lacks the flexibility and compatibility with other AI models, which has led some customers to seek alternatives [7][10]. Group 2: AWS's Strategic Response - AWS has established a dedicated department for developing AI agents, indicating a strong commitment to enhancing its capabilities in this area [3]. - The upcoming upgrade of Bedrock aims to provide a more open and flexible environment for businesses to utilize various AI models and development tools [9][15]. - AWS is also promoting AI agent development through initiatives like the release of open-source development tools named Strands Agents [11]. Group 3: Competitive Landscape - The competition in the AI cloud services market is intensifying, with AWS needing to maintain its leading position to protect its profitability, as it is a major revenue source for Amazon [16]. - Despite being the market leader, AWS's revenue growth is lagging behind that of Microsoft and Google by 10 to 15 percentage points, highlighting the urgency for AWS to innovate and retain its customer base [16].