Reports
Report 2026-05-25 48 sources Input Tokens: 45,686 · Output Tokens: 5,518 · Total Tokens: 52,853

AI 戰情報告 (2026-05-25)

metagoogledeepseekalibabahuaweiasusmicrosoftanthropic gemini-embedding-2claude-mythosqwenllama agentraghbmsupply-chain-attackon-device-aimultimodalm2m-payment

AI 戰情報告 (2026-05-25)

昨日 AI 重點新聞

華為計畫 2031 年實現 1.4 奈米等效晶片,突破美國制裁限制 ref-6

華為在 IEEE 會議上提出全新晶片縮放定律,目標在 2031 年前設計出電晶體密度等效於 1.4nm 的晶片。這項長期戰略旨在透過架構優化,在缺乏先進光刻機的情況下持續提升其麒麟系列晶片的競爭力。

Google 發表 Gemini Embedding 2,支援全模態 Agentic RAG ref-0

Google 推出 Gemini Embedding 2,首度實現文字、影像、音訊與文件的單一語義空間映射。該模型支援 Matryoshka 維度縮減技術,能顯著降低 RAG (檢索增強生成) 的存儲成本,並優化 AI 代理處理複雜多模態資訊的效能。

DeepSeek 透過模型優化降低 HBM 依賴,挑戰硬體成本牆 ref-4, ref-19

DeepSeek 分享其模型優化經驗,旨在大幅減少對 HBM (高頻寬記憶體) 的需求。最新數據顯示,63% 的 AI 晶片成本來自記憶體,DeepSeek 此舉可能讓一般 CPU 或 ASIC 也能執行強大模型,降低企業部署門檻。

阿里巴巴發布專為「AI 代理」設計的新型 AI 晶片 ref-39

阿里巴巴 揭曉一款專為 AI Agents 優化的硬體,其架構針對代理任務中的頻繁邏輯跳轉與自主決策進行強化。這標誌著晶片研發已從單純追求運算速度轉向支援「代理邏輯執行」的新階段。

ASUS 推出混合代理式 AI 基礎設施,優化推論成本 ref-38

華碩 (ASUS) 發表 Hybrid Agentic AI 基礎設施,結合本地與雲端運算資源。該架構旨在幫助企業在保持極低延遲的同時,最大程度地降低 AI 代理在大規模營運下的推論成本。

Kore.ai 發表 Artemis 平台,啟動企業級代理治理機制 ref-15

Kore.ai 推出 Artemis 平台,專注於構建、治理與優化企業級 AI 代理。該平台提供透明的審計工具,旨在解決大型企業在自動化流程中對 AI 行為不可預測性的擔憂。

微軟因成本激增限制 Claude Code 存取,反映 AI 代碼代理的高昂支出 ref-18

微軟 內部已開始削減對 Anthropic Claude Code 的存取權限。儘管 AI 輔助編程提升了效率,但其產生的巨大 Token 成本已對大型企業的營運預算造成壓力,驅動市場轉向更具成本效益的開箱即用方案。

AI 代理自主撰寫 SpaceX IPO 備忘錄,微型支付經濟成型 ref-20

Circle 執行長展示了一名 AI 代理以約 1 美元成本撰寫的高質量投資備忘錄。此案例再次強調了以 USDC (加密貨幣) 作為「機器支付 (M2M)」載體的重要性,能有效處理低額高頻的代理服務費用。

TrapDoor 供應鏈攻擊橫跨多個軟體生態系,竊取開發者憑證 ref-11

資安界偵測到名為 TrapDoor 的協同攻擊,在 npm、PyPI 等平台發布惡意軟體包以竊取開發者帳密。這對日益依賴 AI 自動化生成的代碼供應鏈構成了嚴峻的資安威脅。

Meta 技術長 Andrew Bosworth 領軍,全面推動「AI 優先」工作流 ref-3

Meta 正處於轉型關鍵期,技術長 Andrew Bosworth 致力於利用 AI 徹底重塑內部萬人團隊的生產力。此舉旨在將 Meta 轉化為一家完全由 AI 驅動的技術公司。

富士通發表自我演化多代理技術,提升業務適應力 ref-13

富士通 (Fujitsu) 研發出一種可根據業務環境動態調整策略的 Self-evolving AI Agents 技術。該技術能讓多個代理在複雜工作流中協同演化,適用於變動快速的商業場景。

Google Cloud Agents CLI 簡化代理從開發到生產的流程 ref-39

Google 推出專屬指令列工具 Agents CLI,讓開發者能快速將 AI 代理原型轉化為生產環境服務。該工具透過優化基礎設施部署,能減少 30% 以上的開發週期。


市場洞察

1. 記憶體牆 (Memory Wall) 成為軟硬體協同優化的核心戰場 ( ) ref-44, ref-4, ref-39

延續前幾日關於 HBM 成本飆升 485% 的報導 (2026-05-22),今日數據證實記憶體成本已佔晶片總價的 63% (ref-25)。這驅動了如 DeepSeek 這種從軟體端降低記憶體需求的技術路徑,以及 阿里巴巴 針對「代理邏輯」而非「純算力」開發專用晶片的趨勢 (ref-29)。這預示著未來 AI 的競爭力將不再僅僅是算力,而是「記憶體使用效率」。

2. 「代理經濟」進入微型支付實踐階段 ( ) ref-39, ref-46

前幾日報告提到的 USDC 機器支付佔比 98.6% (2026-05-24) 今日獲得具體案例驗證。AI 代理以極低成本 ($1.02) 交付專業文獻 (ref-38),顯示了「任務顆粒化」後的經濟潛力。對於軟體開發商而言,如何讓產品具備接受 AI 代理自主支付的能力,將成為進入新經濟體系的門票。

3. 多模態 RAG 轉向「語義空間對齊」 ( ) ref-0

Google 的 Gemini Embedding 2 (ref-47) 標誌著 RAG 技術的重大演進:從單純的文字比對轉向文字、音訊、影像的全模態融合。這解決了過去 AI 代理在處理掃描文件、多媒體簡報時的語義斷裂問題。

4. 開發者生態系的安全信任危機 ( ) ref-11, ref-15

TrapDoor 供應鏈攻擊 (ref-23) 與 Node.js 對 AI 生成代碼的爭論 (ref-21),揭示了 AI 輔助開發的陰暗面。當 AI 被用來大規模生成代碼或軟體包時,審核速度已跟不上惡意代碼的擴散速度,這將迫使企業導入更嚴苛的「代理資安治理 (Agent Security Governance)」。

Sources 48 items
ref-0 Google Developers Blog Building with Gemini Embedding 2: Agentic multimodal RAG and beyond ref-1 Google Developers Blog Agents CLI in Agent Platform: create to production in one CLI ref-2 TechMeme A look at Xiaomi's AI push to future-proof its hardware and EV ecosystem, as it recently committed ~$8.8B in AI investments over the next three years (Ben Jiang/South China Morning Post) ref-3 TechMeme A profile of Meta CTO Andrew "Boz" Bosworth, a top lieutenant of Mark Zuckerberg who is leading the gargantuan effort to transform Meta into an AI-first company (Meghan Bobrowsky/Wall Street Journal) ref-4 TechMeme A look at DeepSeek's model optimization to reduce HBM use, potentially enabling domestic memory, ASIC, and CPU makers to create a Chinese AI hardware ecosystem (@bookwormengr) ref-5 TechMeme The UK, outpaced by the US and China in AI, is turning to experimental technologies like neuromorphic computing in search of computing sovereignty (Charles Clover/Financial Times) ref-6 TechMeme At an IEEE conference, Huawei proposed a new chip scaling law and said it aims to design chips with transistor density equivalent to 1.4nm by 2031 (Nikkei Asia) ref-7 TechMeme Plenary Americas LP, La Caisse's infrastructure arm, agrees to acquire and take private Canadian database company ISC for $872M, expected to close in Q3 (Melissa Shin/Bloomberg) ref-8 TechMeme Quartermaster, which is building an analytics platform and a SmartMast, with sensors on a ship's mast to relay real-time maritime data, raised a $43M Series A (Sean O'Kane/TechCrunch) ref-9 InfoQ AI/ML NodeJS Proposes Built-In Virtual File System, Sparking Debate Over AI-Generated Contributions ref-10 InfoQ AI/ML OpenJDK News Roundup: Vector API, Compact Object Headers and G1GC as Default in JDK 27 ref-11 The Hacker News (Security) TrapDoor Supply Chain Attack Spreads Credential-Stealing Malware via npm, PyPI, and CratesIO ref-12 EU Startups Finland’s Grundium acquires Denmark’s Visiopharm to build an end-to-end AI precision pathology platform ref-13 Google News Fujitsu develops self-evolving multi-AI agent technology that learns and adapts to business operations ref-14 Google News New AI model detects hidden antibiotic resistance genes beyond standard databases ref-15 Google News Kore.ai launches Artemis, the new generation of the Kore.ai agent platform for building, governing, and optimising enterprise AI ref-16 Google News AI agent payments surge as crypto rails dominate ref-17 Google News Outsiders breached Anthropic’s Claude Mythos same day ‘potentially dangerous’ AI model was revealed: report ref-18 Google News Microsoft cuts Claude Code access as AI coding costs surge ref-19 Google News DeepSeek To Make Permanent 75% Discount on Flagship AI Model ref-20 Google News AI agent drafts SpaceX IPO memo for just over a dollar, Circle CEO wants you to notice the currency ref-21 Google News DAPPOS launches xBubble: An AI agent that learns and uses AI for you ref-22 Google News AI model finally learns to say ‘I don’t know’ in breakthrough to curb chatbot overconfidence ref-23 Google News What AI coding benchmarks still miss about software quality ref-24 Google News Perplexity reveals new AI agent tool: What to know ref-25 Google News Google Cloud opens SEA AI startup accelerator ref-26 Google News AI startup Sereact raises $110 million for robots that predict consequences ref-27 Google News Qualcomm (QCOM) Stock; Surges 11% After Stellantis Expands AI Partnership Deal ref-28 Google News US-China AI funding race hits record highs amid tech rivalry ref-29 Google News Meta Halts Work With Mercor After Major AI Data Breach ref-30 Google News Pope Leo brings Silicon Valley into the AI ethics fight ref-31 Google News United Nations Opens Registration for the 2026 Global Dialogue on AI Governance in Geneva ref-32 Google News Australia, UK sign AI safety pact as frontier model fears grow ref-33 Google News OpenAI offers up to $445,000 for AI safety role as focus shifts to self-improving systems ref-34 Google News A professor who helped write UC Berkeley's strict new AI policy says it's about preserving the value add of a lawyer ref-35 Google News Govt. to soon bring in an AI policy to transform Bihar: CM ref-36 Google News Trump scrapped a major AI safety plan — here’s why that matters for ChatGPT users ref-37 Google News At APEC Xi Moots Global Body For AI Governance ref-38 Google News ASUS Takes the Lead in Hybrid Agentic AI Infrastructure- Maximizing Performance While Reducing Inference Costs ref-39 Google News Alibaba reveals new AI chip designed for 'agents' ref-40 Google News SoftBank to launch AI GPU cloud in October 2026 ref-41 Google News Why are NVIDIA H100 GPU rental prices surging by 40% ref-42 Google News Core42 raises US$ 550 million to scale AI infrastructure ref-43 Google News After $2.5 billion Supermicro smuggling bust, Nvidia CEO urges company to fix export control compliance — Taiwan also begins to crack down on AI GPU chip smuggling to China ref-44 Google News 63% of the cost of AI chips goes to memory; we're entering an era where memory costs more than the GPU. ref-45 Google News Huawei Introduces New AI Chip Strategy for Future Computing ref-46 Google News Oracle stock’s breakout is real and the long-term AI infrastructure case is only getting stronger ref-47 Google News AMD unveils $10 billion AI infrastructure push in Taiwan