整理 | 华卫 即使你对生成式 AI 模型的内部运作了解不多，也大概率知道它们极其吃内存。正因如此，如今想买一根普通内存条都免不了被狠狠加价。 最近，谷歌研究院发布了 TurboQuant 压缩算法，能够在提升运行速度并保持准确性不变的前提下，降低大语言模型（LLM）的内存占用。如果 TurboQuant 成功落地，可将 AI 运行时的 "工作内存"，也就是键值缓存（KV cache）压缩至少 6 倍，并在 H100 显卡上实现最高 8 倍的速度提升，从 而大幅降低 AI 运行成本。 Cloudflare 的 CEO Matthew Prince 等人甚至称，这是谷歌的 "DeepSeek 时刻"。此前，中国 AI 模型 DeepSeek 实现这样的效率飞跃：该模型在性能 保持竞争力的情况下，训练成本仅为对手的零头，且使用的芯片性能较差。 而 TurboQuant 最关键的亮点是：精度零损失。无需微调，无需训练数据。直接接入任意 Transformer 模型，即可让键值缓存压缩至原体积的一小部 分，同时输出结果完全一致。如果这一效果能在实际生产环境中成立，将一夜之间改变长上下文推理的成本格局。 此外，Turb ...

While OpenAI has listed its reliance on Microsoft as a “risk,” dependence on the Sam Altman-led AI startup is a risk for Microsoft, too, which was on full display when the company released its most recent quarterly report.Meanwhile, as part of the funding from Amazon, OpenAI signed a $50 billion cloud deal with Amazon Web Services (AWS). Microsoft is reportedly considering a legal action over that deal, alleging that it violates its exclusive cloud deal with OpenAI.Earlier this year, OpenAI secured $110 bil ...

AI 搜索解析： 营销人员实用指南 AI 搜索解析：营销人员实用指南 | 目录 | | --- | | 01 | 前言 | 3 | | --- | --- | --- | | 02 | 本指南的目的 | 4 | | 03 | 解构大语言模型（LLMs） | 5 | | 04 | AI 搜索如何运作？ | 8 | | 05 | AI 搜索如何呈现品牌？ | 9 | | 0 6 | 从 SEO 走向 GEO：品牌如何被呈现？ | 12 | | 07 | 如何撰写清晰结构化的内容，以提升在 AI 搜索中的可信度 | 14 | | 0 8 | 内容策略的实用建议 | 16 | | 09 | 充分发挥 AI 价值的付费策略 | 18 | | 10 | 以人为本，回归人性 | 20 | | 11 | Microsoft 如何提供支持 | 21 | 2 AI 搜索解析：营销人员实用指南 前言 致营销人： 过去一年，我们正身处行业历史上发展最快的转型阶 段之一。 AI 概览、新一代 AI 浏览器，以及功能日益强大的多 模态智能助手，正在从根本上改变人们在线搜索信 息、发现内容以及做出决策的方式。因此，许多营销 人员向我们坦言 ...

This story was originally published on Marketing Dive. To receive daily news and insights, subscribe to our free daily Marketing Dive newsletter. Dive Brief: Apple plans to roll out a functionality that will allow marketers in the U.S. and Canada to post ads on Maps starting this summer, according to a press release. The ads will appear when consumers use the search function in the app, as well as a new Suggested Places functionality. The ad option is part of a new Apple Business platform, which is se ...

Google expands utility deals to curb datacenter power use during peak demand | Reuters Skip to main content Exclusive news, data and analytics for financial market professionalsLearn more aboutRefinitiv The Google logo is seen on the Google house at CES 2024, an annual consumer electronics trade show, in Las Vegas, Nevada, U.S. January 10, 2024. REUTERS/Steve Marcus/File Photo Purchase Licensing Rights, opens new tab Companies Google Inc Follow Alphabet Inc Power demand typically spikes on very hot or cold ...

Core Viewpoint - Alphabet, Google's parent company, plans to significantly increase its capital expenditures to between $175 billion and $185 billion in 2026, nearly doubling its 2025 spending, primarily to enhance its position in the competitive AI landscape [1]. Group 1: Capital Expenditures and AI Strategy - The substantial budget will focus on data centers, networking equipment, AI processors, and servers, aimed at maintaining a leading edge in the costly AI race [1]. - Executives are motivated by historical examples of companies that failed to adapt to innovation, emphasizing the need for timely adjustments to avoid becoming obsolete [2][3]. Group 2: Cloud Services and Financial Performance - Google Cloud is a key growth area, with Q4 2025 revenue rising 48% year-over-year to $17.6 billion, and operating income increasing 30.1% to $5.3 billion [4]. - The backlog for Google Cloud surged 55% sequentially and more than doubled year-over-year, reaching $240 billion, driven by strong demand for enterprise AI offerings [5].

Group 1 - Big Tech companies are heavily investing in the AI sector, spending hundreds of billions or even trillions of dollars to compete against each other, and their investors currently support this strategy [1] - The competitive landscape in the AI arms race indicates a significant commitment from major technology firms to enhance their capabilities and market positions [1] Group 2 - The article reflects a perspective shaped by a background in data analytics and social media marketing, emphasizing the importance of datasets, trends, and pattern recognition in market analysis [2] - The author transitioned from a casual investor to an active trader, engaging in swing trading and options trading, which highlights the evolving nature of investment strategies in response to market dynamics [2]

Group 1: Microsoft - Microsoft is currently the worst-performing stock among the "Magnificent Seven" due to slower growth in its Azure cloud platform and high costs in the AI sector [1] - The stock is under pressure amid a relatively high valuation, but this does not indicate broader trouble for megacap stocks [1] Group 2: Alphabet - Alphabet has invested heavily in AI infrastructure, spending $91 billion on capital expenditures last year and pledging $175 billion to $185 billion for this year [3] - The Google Gemini AI engine has made significant competitive strides, with some users preferring it over ChatGPT [4] - Google Cloud is growing faster than its digital ad platform, indicating a stronger reliance on AI, and Waymo is expected to be a major revenue driver in the future [5] Group 3: Amazon - Amazon has committed to $200 billion in capital expenditures for 2026, following nearly $132 billion spent the previous year [8] - Despite concerns over high capex, Amazon Web Services has shown rapid growth recently, suggesting that investments may soon yield returns [10] - Amazon's e-commerce segments leverage AI for product selection and supply chain efficiency, with many segments reporting double-digit revenue growth [11] - The stock trades at a P/E ratio of 30, which is lower than its historical average, indicating a potential buying opportunity [12]

Core Viewpoint - Meta Platforms is experiencing a decline in stock value due to reports of postponing the release of its AI model, Avocado, which raises concerns about its overall AI strategy and spending [1][5][20] Company Performance - Meta's stock is down 3.5%, reflecting investor concerns about the delay in the AI model rollout, which was initially targeted for release this month but is now pushed to at least May [1][4] - The new AI model has reportedly not met performance expectations compared to leading models from competitors like Google and OpenAI, which could impact Meta's ability to attract talent [5][8][9] Industry Context - The broader technology sector is facing challenges, with the NASDAQ index showing a decline of approximately 4.75% this week, indicating a general downturn in tech stocks [10] - Major tech companies, including Meta, are committing significant resources to data center expansions, with Meta's future lease commitments reaching $104 billion, reflecting the high costs associated with AI development [12][15][16] Financial Implications - The total future lease commitments across cloud providers have exceeded $700 billion, with hidden costs that may not be immediately visible on balance sheets, raising concerns for investors [13][14][18] - Companies like Microsoft and Oracle are also making substantial commitments, with Microsoft at $155 billion and Oracle at $261 billion, indicating a competitive landscape in AI infrastructure spending [15][16][17]

Core Insights - Electricity costs for Americans have risen by 30% from 2021 to 2025, with further increases expected due to geopolitical tensions and rising energy demands from data centers [1][2] - Google and Tesla have formed a coalition named Utilize, along with other companies, to address high electricity costs and improve grid efficiency [2][3] Group 1: Electricity Costs and Demand - The war in Iran is threatening global oil supply, contributing to rising energy costs [1] - Data centers are significantly increasing energy demands, with costs rising as much as 267% for some consumers in neighboring towns from 2020 to 2025 [5] - In 2024, data centers accounted for approximately 4% of U.S. electricity use, a figure expected to more than double by 2030 due to increased AI usage [5] Group 2: Coalition and Solutions - The coalition, Utilize, aims to enhance grid utilization, arguing that the current underutilization is a key factor in high electricity bills [3][4] - A study from Duke University indicates that regional electric systems are operating at only 53% capacity on average, highlighting the potential for improvement [3] - Utilize's Executive Director emphasized the need to unlock excess capacity in the grid, likening it to an airplane that only flies full a few times a year [4] Group 3: Industry Implications - Google and Tesla's involvement in the coalition reflects their vested interests in data centers, which consume significant electricity [4] - As electricity demand grows, the coalition aims to ensure that this demand does not lead to increased costs for existing customers [6]