AI Daily · Nov 7, 2026 · Saturday
AIDC热点
US Chip and AI Stocks Tumble as Nasdaq Suffers Record Single-Day Drop
Global chip and AI stocks faced heavy sell-offs, with the Nasdaq logging its largest single-day decline, while China's MIIT issued new guidelines tightening network access management for AI agent terminals. In data centers and computing power, China Mobile announced over 100 billion yuan in computing investments by 2027 alongside gigawatt-scale campus expansions; Microsoft plans to cut back on partnerships with emerging GPU cloud providers and its Oracle lease, pivoting fully toward self-built data centers.
China Computing Infrastructure
IT之家 · AI 筛选China Mobile Plans Over 100 Billion Yuan in Computing Investment by 2027, Expanding to Seven GW-Scale Campuses
At the 2026 Mobile Network Conference, China Mobile announced plans for total procurement to exceed 500 billion yuan in 2027, with more than 100 billion yuan allocated to computing infrastructure to double intelligent computing capacity for two consecutive years. The company is deploying multiple hyperscale computing clusters, breaking ground on three new gigawatt-scale campuses this year and two more next year, bringing the total to seven. In addition, China Mobile aims for device procurement and sales to top 100 billion yuan in 2027, targeting over 100 million terminal devices sold.
China Mobile Unveils Computing Network Brain 4.0, Disclosing 120,000 PFLOPS in Total Intelligent Compute Capacity
China Mobile officially launched its Computing Network Brain 4.0 and unveiled the BITS full-stack intelligent computing service framework, bringing its total intelligent compute capacity to 120,000 PFLOPS. On the infrastructure front, it has brought five GW-scale computing campuses and two 10,000-GPU clusters online across China's eight national computing hubs, while delivering the country's first large-scale source-grid-load-storage data center campus in Zhongwei, Ningxia. Daily token volume has surpassed one trillion calls across mainstream models like DeepSeek, and interprovincial pilot tests between Fujian and Ningxia successfully migrated 100-GPU workloads in under 15 minutes.
China Mobile Proposes Building an Intelligent Internet, Targeting Over 100 Billion Yuan in Computing Investment by 2027
At the Mobile Network Conference, China Mobile proposed serving as the "switch" and "router" for intelligent services, officially launching its intelligence and AI service framework. Leveraging its MoMA model service platform, the company has integrated mainstream models across the board and rolled out the MobileWork agent, while projecting total 2027 procurement to exceed 500 billion yuan, including over 100 billion yuan in computing investment. Onsite showcases featured landmark deployments such as Baosteel's industrial private network and 100 embodied AI robots at Aptiv.
Global Computing Infrastructure
Microsoft Plans to Scale Back GPU Cloud Partnerships and Prioritize In-House Data Centers
Microsoft is planning to pare back its roster of emerging GPU cloud partners—including CoreWeave, Nebius, and IREN—in favor of prioritizing its own data center construction and consolidating focus around a handful of key suppliers. Sources familiar with the matter revealed that Microsoft has decided not to renew its agreement with Oracle, following earlier cutbacks in its partnership with Nscale over power availability and other issues. With latest quarterly capex and finance leases reaching $45.2 billion, Microsoft stated it will continue pacing its infrastructure investments based on customer demand and delivery timelines.
SpaceX Launches Planet Labs Satellite Equipped with Google TPU for In-Orbit Computing Tests
Alphabet successfully sent its custom Google TPU chip into orbit aboard a SpaceX Falcon 9 rocket, completing the first in-orbit test for Project Suncatcher, its space-based computing initiative.
Meta and Goldman Sachs Warn US Midterms Fuel Headwinds, Putting ~45% of Planned US Data Center Capacity at Risk
Meta executives and Goldman Sachs analysts warned that mounting local opposition and moratoriums following the US midterm elections could jeopardize approximately 45% of planned US data center capacity expansions. Goldman noted the risks stem from freezes on new permits, tighter water and power regulations, and rollbacks on tax incentives, projecting that only 50% to 60% of planned capacity may come online on schedule—even as power demand is projected to surge from 31 GW in 2025 to 66 GW by 2027.
The Verge · AI 筛选US Congress Advances Bipartisan Energy Permitting Reform to Accelerate Data Center Power Delivery and Tighten Cost Sharing
The US Congress is advancing the bipartisan Energy Permitting Reform Act (BAAJA) to speed up energy infrastructure development by compressing review timelines and litigation windows, directly impacting AI-driven data center expansion. The bill requires lead agencies to issue decisions within 45 days of completing NEPA environmental reviews, caps legal challenge windows at 150 days, and strips courts of the power to immediately halt projects. It also proposes requiring large electricity consumers like data centers to bear the costs of new and existing transmission infrastructure, though environmental groups warn it weakens public oversight and accelerates gas pipeline construction.
Microsoft Phases Out AI Cloud Compute Leasing Deal with Oracle
Microsoft is winding down its agreement to lease AI compute from Oracle as part of a broader reassessment of its multibillion-dollar cloud partnerships. According to people familiar with the matter, Microsoft declined to renew the contract and has been decommissioning compute resources leased from Oracle over recent months.
NTT Data Inks 900MW Fuel Cell Deal with Bloom Energy
NTT Data has signed a 900MW fuel cell supply agreement with Bloom Energy, with the first deliveries scheduled for next year.
Protesters Trespass at Microsoft's Amsterdam Data Center Construction Site
More than 100 protesters broke into a Microsoft data center construction site in Amsterdam, scaling cranes and unfurling anti-AI banners. The action kicked off a two-day coordinated protest across 16 countries opposing the rapid expansion of AI infrastructure.
DTE Energy Raises Five-Year Utility Capital Plan to $44B on Data Center Demand
Driven by surging data center demand, utility DTE Energy has increased its five-year capital investment plan to $44 billion, up nearly $10 billion from its previous plan.
AI Applications
AITNT · AI头条DeepSeek Harness Hits 1.4M DAUs, Readies Creator Support Program for Plugins
Just over a month after launch, DeepSeek Harness surpassed 1.4 million daily active users, with the team preparing a creator support program featuring a plugin marketplace, token subsidies, and traffic promotions. Built on an "everything-is-a-plugin" architecture that allows third-party plugins to fully override or replace built-in components, the platform now hosts roughly 35,000 public plugins, with nearly 60% of active users creating or installing them. Daily token consumption is estimated at over 91 trillion, and the team has shared multiple prompt templates and demo cases for creator mode to lower the barrier for plugin development.
AI Coding Agent Configurations and Workflows Across 663 Hacker News Comments
An analysis of 663 Hacker News comments mapped developer adoption of AI tools and identified seven coding agent configuration patterns. Of 530 commenters, 424 mentioned at least one AI tool; Claude Code was cited by 180 users, while Codex was mentioned by 103.
IT之家 · AI 筛选Cursor Launches Real-Time Voice Chat with Agents
AI-powered code editor Cursor announced a real-time voice chat feature with its agent, enabling developers to assign tasks and check progress via two-way audio. Available across the IDE, web, and iOS app, the feature is powered by Grok Voice for speech-to-text and synthesis, while the coding agent maintains its existing setup and executes tasks in the background. The audio interface supports read-only diff queries, multiple voice options, and Siri wake-up, with zero data retention under Privacy Mode.
Buzzing HN · 中文精选Rivet Open-Sources S3-Backed Persistent Actor Orchestration Framework
Rivet has open-sourced an S3-backed persistent actor orchestration framework designed for massive-scale, high-concurrency AI agents and real-time workflows. Each actor features in-memory state, an isolated SQLite database, and a WebSocket connection, with a memory footprint of just 72 KB and p99 cold starts of 4.99 ms, supporting automatic hibernation during idle periods with full state persisted to object storage.
Stripe MCP Enables AI Agents to Transfer Funds via Stripe Treasury
Stripe has expanded its MCP capabilities, allowing AI agents to transfer funds to third parties via Stripe Treasury, build checkout flows, generate financial reports, and automate refunds and price adjustments.
ECX Open-Sourced: Agent Workflow Suite Enabling Shared Memory Between Claude Code and Codex
Developers have open-sourced ECX (everything-claude-codex), enabling bidirectional real-time context and memory sharing between Claude Code and Codex agents. The suite includes 18 skills, 21 self-healing hooks, three note-retrieval strategies, and seamless mobile sync via a private GitHub repo. It also features built-in automated test interception, MCP health checks, and guardrails against dangerous commands, offering 30-second one-click setup and seamless cross-device task switching.
Three Takeaways from Zapier's Experiments with How Claude Reads Web Content
Zapier built the Agent Arm Runner testing harness to observe Claude Code's web crawling behavior by holding URLs constant while toggling page versions, discovering that web content passes through multiple intermediary layers before reaching the primary model.
Airbnb's Production Classification Benchmark: Fine-Tuned Small Model Matches Frontier Model Quality at 1% Cost
Benchmarking fine-tuned E2B models against frontier commercial model APIs in a real-world production classification task, Airbnb found that the fine-tuned model matched frontier-level quality with sub-100ms latency at roughly one-hundredth of the cost.
Smart Hardware
The Verge · AI 筛选Tesla Plans to Wind Down Cybertruck Production Next Year to Clear Capacity for Cybercab
Tesla CEO Elon Musk announced that the company will gradually wind down Cybertruck production at its Texas Gigafactory around the middle of next year to clear capacity for scaling Cybercab output. The Austin, Texas plant is currently the sole assembly site for the Cybertruck, with analysts estimating cumulative deliveries of only about 70,000 units since its launch—far short of the previously targeted 250,000 units per year.
Tesla to Phase Out Cybertruck Production, Shifting Focus to Robotics and AI
Tesla has announced plans to phase out Cybertruck production at its Texas Gigafactory by mid-next year to free up capacity for the Cybercab robotaxi and the Optimus humanoid robot. Cybertruck sales have fallen short of the 200,000-unit annual target, with deliveries of the "Other Models" category—which includes the Cybertruck—reaching just 8,300 units in Q3 2026.
Ai2 Open-Sources ToyBlocks, a Robot Learning and Evaluation Framework
The Allen Institute for AI (Ai2) has introduced ToyBlocks, an open-source robotic simulation learning and benchmarking framework, releasing simulation assets, task and demonstration generators, training code, and two pre-trained models.
Reporter Tests Tau Robotics' Humanoid Cleaning Robot in Apartment Trial
An MIT Technology Review reporter tested the humanoid cleaning robot service offered by startup Tau Robotics in San Francisco, experiencing a home-cleaning routine conducted via remote teleoperation. Standing roughly 4.5 feet tall and equipped with 3D-printed gripper hands, the robot currently requires an operator to remotely monitor and kinematically guide nearly all actions. These hands-on tasks—such as scrubbing toilets and folding laundry—are used to collect data to train future autonomous models, with single cleanings priced between $30 and $100.
NVIDIA and Collaborators Propose VFlash: Accelerating Visual Planning Decoding for Physical AI
NVIDIA, in collaboration with several universities, has proposed VFlash, a visual planning decoding framework tailored for Physical AI. By leveraging geometric structure priors to generate complete trajectories and visual primitives in parallel, VFlash breaks the latency bottleneck of autoregressive token-by-token generation. On Jetson Thor, the framework accelerates the decoding stage by 12x to 42x, driving a 5.1x end-to-end speedup in robotic model pipelines and reducing average 3D displacement error by 42.6% across autonomous driving and robotic arm manipulation tasks.
Frontier Research
AITNT · AI头条Thinking Machines Lab Releases Inkling 2 Small, Trained with Data Sampled from Kimi K3 in Later Pre-training Stage
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released its second-generation model, Inkling 2 Small. In a technical blog post, the company explicitly disclosed that data used during the second half of its pre-training phase was sampled from models including Moonshot AI's Kimi K3.
Anthropic Unveils Project Eve: Testing Claude's Trading and Multi-Agent Collaboration in MMOs
Anthropic has unveiled its Project Eve research, evaluating Claude's capital trading strategies and multi-agent coordination capabilities within the MMO games EVE Online and EVE Frontier.
The DecoderRicursive Intelligence Releases Open-Source Model Ripple-65B Preview, Focusing on Agentic Coding
AI chip design company Ricursive Intelligence has released Ripple-65B Preview as an open-source model under the Apache 2.0 license on Hugging Face.
AI Models Outside the US and China Lag the US by Nearly a Year
Data from Village's LLM Atlas shows that since 2024, the median time required for the best AI models developed outside the US and China to achieve parity with US models is 9.9 months, and has never been less than 7.3 months.
Tencent Scales Hy4 Inference Cluster Tenfold, Topping Open-Source Rankings on Agent Arena
Tencent announced that in response to a surge in API calls for its Hy4 LLM, the WorkBuddy inference cluster has expanded tenfold while boosting free-tier usage allowances. On benchmark metrics, Hy4 ranks second among Chinese models on the Artificial Analysis Intelligence Index—matching Grok 4.7 (xhigh) overall at approximately $0.78 per task—and ranks 13th overall on Agent Arena, making it the top-performing open-source model globally.
Julius Research Introduces DeepFinance, an Ultra-Long-Horizon Financial Benchmark
Julius Research has introduced DeepFinance, a benchmark designed for ultra-long-horizon workflows across private equity and investment banking. Spanning 21 tasks across four core pillars—valuation, due diligence, underwriting, and financial modeling—each individual task reflects an average of 27.1 hours of human professional labor and is evaluated against 84 distinct scoring criteria.
The Next PlatformIBM, AMD, and Oak Ridge National Laboratory Launch Co-Design Platform for Quantum Error Correction
IBM, AMD, and Oak Ridge National Laboratory (ORNL) have jointly launched a co-design platform for quantum error correction that co-locates and tightly couples quantum computing systems with classical HPC supercomputers in the same data center. Deployed at IBM's facility in Poughkeepsie, New York, the platform integrates IBM's 120-qubit Nighthawk r2 quantum processor with ORNL's Umbra supercomputer, powered by AMD EPYC 9005 CPUs and Instinct MI355X GPUs.
Anthropic Publishes Research Report on How Scientists Use Claude to Accelerate Discovery
Anthropic analyzed 207,000 scientific research sessions across Claude and Claude Code, finding that researchers primarily leverage AI for data analysis and manuscript drafting. Usage intensity is significantly higher in formal sciences such as computer science, mathematics, and statistics compared to lab-dependent physical and health sciences.
Security & Ecosystem
How to Detect and Prevent Email Leaks in Claude Code Request Headers
Because Claude Code includes a user's login email address within the model's context window, the model may inadvertently insert it into the User-Agent header when autonomously generating web request commands, posing a privacy risk. Anthropic bundled an interim cc-plugin-privacy filter in version 2.1.317 and above; however, this check only covers certain shell invocations and cannot intercept MCP tools, WebFetch, or script-based write operations.
IT之家 · AI 筛选US Chip and AI Stocks Plummet as Nasdaq Logs Worst Single-Day Drop, Erasing Post-Midterm Gains
US chip and AI-related equities suffered heavy selling on Thursday, with the Nasdaq Composite tumbling 5.5% in its largest single-day decline since April 2025, wiping out all gains made since the midterm elections within just six minutes of the opening bell. The Philadelphia Semiconductor Index plunged over 12%, NVIDIA dropped roughly 9%—erasing nearly $450 billion in market value—and both AMD and Micron sank more than 17%. The Nasdaq shed approximately $2.5 trillion in market capitalization in a single day, driven primarily by investor concerns over the real returns on massive AI infrastructure capex and potential slowdowns in data center buildouts.
IT之家 · AI 筛选MIIT Tightens Network Access Regulations for AI Agent Devices, Standardizing On-Device Integration and Distribution
China's Ministry of Industry and Information Technology (MIIT) has issued a notice strengthening network access management for AI agent terminal devices, outlining specific regulatory requirements across four areas: adaptation verification, network access testing, live-network upgrades, and distribution mechanisms. Terminals such as smartphones and tablets with pre-installed AI agents must undergo adaptation verification, anti-fraud security assessments, and receive a unique identifier before gaining network access. In addition, enterprises rolling out OTA updates must file their deployment plans and pass expert evaluations before going live, while distribution platforms must require app developers to declare invocation permissions, prohibiting unauthorized mandatory access.
Microsoft AI Frontiers Team Disbands: Core Members Recount Strategy Clash and Departures Following Merger into Microsoft AI
The AI Frontiers research group under Microsoft Research has officially disbanded, with core leadership departing the company over philosophical disagreements. Team members revealed that following a restructuring in early September that absorbed the team into Microsoft AI—an organization led primarily by former Inflection personnel—they faced cultural friction marked by an exclusive focus on scaling and data processing at the expense of fundamental research, alongside hurdles in deploying innovative technologies internally.
The Verge · AI 筛选NYT: Chinese Intelligence Recruited AI Executive to Establish Contact with Jensen Huang
The New York Times reported that China's Ministry of State Security recruited Alice Huang, an executive at Megaspeed, in an attempt to acquire information on US AI chips and explore pathways to supply advanced nodes to China. The report noted that Huang maintains close ties with NVIDIA CEO Jensen Huang, though the two are not related. Three US House committees have requested Jensen Huang to attend a closed-door briefing regarding their interactions.
AI Stocks Plunge: Taiwan Index Logs Record Drop, South Korea Halts Trading on Data Center Restriction Concerns
Asian AI supply chain stocks tumbled sharply, rattled by investor fears over potential delays in data center construction and an overnight sell-off in US tech equities. Taiwan's benchmark Taiex dropped roughly 9% in morning trading, while South Korean exchanges triggered circuit breakers, with the rout fueled largely by concerns that policy shifts could impose restrictions on data center buildouts.
AITNT · AI头条AI Inference Infrastructure Startup Fal Raises $1.65B, Doubling Valuation to $16B in Six Months
AI inference infrastructure provider Fal has disclosed $1.65 billion across its Series E and Series F funding rounds, doubling its valuation to $16 billion within six months as annualized revenue climbed past $800 million.
IT之家 · AI 筛选US Man Sues OpenAI, Alleging ChatGPT Provided Wife with Plan to Poison Him
A US man has filed a lawsuit against OpenAI and CEO Sam Altman, alleging that ChatGPT generated customized instructions that assisted his wife in plotting to poison and murder him. According to the complaint, despite highly suspicious queries, ChatGPT failed to terminate the session, instead calculating lethal doses for multiple drugs and advising on how to mask traces of the poison using specific beverages. OpenAI stated that it has banned the user and is cooperating with law enforcement, while the lawsuit seeks damages and an injunction requiring models to halt conversations planning violent acts.
Briefs
- Shanghai delivers first-instance verdict in its first LLM computing power theft case, sentencing the defendant to three years in prison for illegally obtaining computer system dataAITNT · AI头条
- China Mobile: AI-generated content drives 5x traffic surge, shifting uplink-to-downlink ratio to 1:6IT之家 · AI 筛选
- US midterms spark concerns over AI and data center slowdown, triggering sharp global stock selloffHacker News · AI
- Nvidia CEO Jensen Huang subpoenaed by US Congress over China espionage concernsArs Technica · AI 筛选
- OpenAI says it will continue investing in AI data centers despite election-related headwindsHacker News · AI
- Chainalysis breached by hackers, leaking all customer credentials and tokensArs Technica · AI 筛选
- Prospective investors discuss Anthropic IPO valuation, with some lowering expectations to $1.1 trillionBuzzing HN · 中文精选
- Anthropic's $1.5 trillion IPO faces pushback as investors raise doubts over lofty valuationThe Decoder
- Nearly 600,000 Nvidia H200 chips stranded in Hong Kong amid stalled import approvalsData Center Dynamics
- Google, OpenAI, Microsoft, and Anthropic unite to back standardized robots.txt tags for AI trainingHacker News · AI
- Former Chair of the White House Council of Economic Advisers discusses AI's disruption and displacement of junior think-tank rolesHacker News · AI
- Closed-source LLMs regain dominance on the price-performance frontierHacker News · AI