Westward Tide · AI · 2026-08-31 · Issue No. 053

Today's Digest: OpenAI's supply cutoff to Cursor is entering its countdown, halting on November 12, triggered by SpaceX's $60 billion acquisition. Sony Music and Warner Chappell have sued Anthropic, seeking up to $150,000 per track in damages. Barclays has broken down the AI industry chain's economics: for every $100 earned by model companies, $35 to $40 goes to the three major cloud providers.

Editor's Note: Companies that were praising each other's ecosystems yesterday are suing each other and cutting off supplies today. The AI industry has entered a phase of splitting profits; those arguing are more honest than those doing business.

---

I. Large Language Models (LLM / Foundation Model)

1. OpenAI Announces Supply Cutoff to Cursor, Effective November 12

• Summary: OpenAI notified late at night on August 28 that it would terminate model supply to Cursor, including the unreleased Astra. The trigger was a change-of-control clause: SpaceX acquired Anysphere, Cursor's parent company, in an all-stock deal worth approximately $60 billion on August 14. Cursor responded that OpenAI models account for only about 5% of their traffic, and users can transition using their own API keys.
• Source: CNBC · 2026-08-30
• Editor's Comment: When model suppliers and application platforms are bound together, it's called an ecosystem when things go well, but a supply cutoff when they don't. Architectures that treat models as replaceable foundations are far more resilient than deeply integrated ones.

2. Sony Music and Warner Chappell Sue Anthropic, Seeking Up to $150,000 Per Track

• Summary: Two major music copyright holders filed a lawsuit against Anthropic and two co-founders in the U.S. District Court for the Northern District of California, alleging large-scale infringement in training data. Media reports describe this case as potentially "the largest copyright theft in history."
• Source: The Verge · 2026-08-30
• Editor's Comment: Calculating damages per single track is a new tactic; the defendants are models, but the pricing is applied to playlists. The music industry is shifting from being passively sued to actively setting prices.

3. Barclays Breaks Down AI Profit Distribution: For Every $100 Earned by Model Companies, $35–$40 Flows to Cloud Giants

• Summary: A recent study by Barclays estimates that approximately 35%–40% of AI model companies' revenue flows to the three major cloud service providers in the form of compute costs. While the model layer appears to be growing rapidly, the bulk of the profit lies in infrastructure.
• Source: IT Home · 2026-08-30
• Editor's Comment: The old story of selling shovels during a gold rush has been rephrased for the AI era. Model companies compete fiercely, while cloud vendors collect rent.

4. Starting September 14, Standard Weekly Limits for Paid Claude Code Users Permanently Increased by 25%

• Summary: Anthropic officially announced on August 30 via developer accounts that the standard weekly usage limits for Claude Code will be permanently increased by 25% for Pro, Max, Team, and Enterprise users.
• Source: IT Home · 2026-08-30
• Editor's Comment: The fact that limits can be permanently raised indicates that inference costs are still declining. Competition among coding assistants has shifted from model capability to quota generosity.

5. Demo Samples of OpenAI's Next-Gen Model Astra Leak, Identified as GPT-6

• Summary: According to testingcatalog, OpenAI's internal codename Astra model has added a mozaik-alpha-fdm testing phase, with expanded internal testing scope. Initial demo samples show that a single interaction can generate game graphics similar to GTA 2.
• Source: IT Home · 2026-08-30
• Editor's Comment: Generating playable games in a single interaction—the focus isn't on graphics quality, but on real-time physical consistency. This timeline is not far from world models.

6. **"Superhuman" AI Coding Is Making Software Worse

• Summary: Engineer William Angel wrote that while AI's ability to write code may have surpassed most humans, fast coding does not equal good software. Architectural rot and failed reviews are undermining long-term quality.
• Source: williamangel.net · 2026-08-30
• Editor's Comment: This article hits the industry's awkward spot. The capacity problem is solved; the taste problem is just beginning.

7. Wiz Releases 90-Day Report on AI Infrastructure Attacks: LiteLLM and MCP Servers Are Major Targets

• Summary: Real attacks captured by Wiz honeypots show active campaigns targeting LiteLLM gateways, MCP servers, and AI frameworks, including RCE, blind prompt injection, and memory credential theft.
• Source: Wiz · 2026-08-30
• Editor's Comment: Attackers have identified the weakest link in the AI stack: it's not the model, but the glue layer where no one takes security audits seriously.

8. Study: AI-Driven Labor Replacement Follows a Three-Stage "Create-Verify-Eliminate" Pattern

• Summary: A paper in Frontiers in Human Dynamics, based on job data from GPT-3 to recent releases, proposes that AI's impact on the labor market follows a path where skills are first verified as automatable, then eliminated en masse.
• Source: Frontiers · 2026-08-31
• Editor's Comment: The most stinging conclusion of the paper is the order of replacement: first replacing what you do slowly, then replacing what you do well.

9. Microsoft Blog Proposes Agentic AI Verification Framework: Only Trust Outputs You Can Verify

• Summary: The Azure team released a methodology for verifying agent outputs, advocating for verifiable assertions at every step of an agent's action rather than relying on end-to-end result sampling.
• Source: Microsoft DevBlogs · 2026-08-30
• Editor's Comment: Verification costs will determine which agent scenarios can enter production. Getting demos running relies on models; getting businesses running relies on assertions.

II. AI Software (AI Application / SaaS)

1. OpenAI Reported to Purchase Tens of Thousands of Mac Minis and Mac Studios for AI Training

• Summary: Reports indicate OpenAI has purchased tens of thousands of Apple desktop devices for reinforcement learning and computer-use agent training, bypassing the CUDA ecosystem to supplement compute power for inference and simulation.
• Source: Crypto Briefing · 2026-08-31
• Editor's Comment: Training computer-use agents requires rolling out data in real desktop environments; a Mac farm is closer to the user's reality than Linux containers.

2. Meta Advances Project OT: Using AI Agents to Replace Some Employee Roles

• Summary: The Street reports that Zuckerberg signaled to employees that Meta is testing the use of AI agents to handle some internal functions, with the related plan internally codenamed Project OT.
• Source: The Street · 2026-08-30
• Editor's Comment: Previously, Meta's data center robots replaced blue-collar workers; now it's white-collar workers' turn. The same company is launching two replacement tracks simultaneously.

3. a16z Establishes "Machine Age" Fund, Betting $1.1 Billion on Chips, Memory, and Robotics

• Summary: Andreessen Horowitz's new fund has raised $1.1 billion (approx. RMB 7.4 billion), investing in three hardware sectors: chips, storage, and robotics.
• Source: IT Home · 2026-08-30
• Editor's Note: After a decade of investing in soft software, a16z is now heavily betting on copper and iron. The fund's name is straightforward: Machine Age.

4. Google Confirms Antigravity Will Support WSL, Connecting Directly to Linux Environment on Windows

• Summary: According to Windows Latest, Google plans to expand WSL support for its AI coding agent tool, Antigravity, enhancing the native experience on Windows 10/11.
• Source: IT Home · 2026-08-30
• Editor's Note: The battlefield for coding agents lies on developers' machines; whoever is closer to the terminal wins half the battle.

5. Snyk Open-Sources agent-scan: Scanning for Prompt Injection Vulnerabilities in AI Agents, MCP Servers, and Skill Packs

• Summary: Security vendor Snyk released a local scanning tool capable of discovering and detecting prompt injections and vulnerabilities in agent components, covering three types of objects: agents, MCP servers, and skills.
• Source: GitHub · 2026-08-30
• Editor's Note: The MCP ecosystem is growing so wildly that security companies are willing to build dedicated tools for it.

6. Runplane: An Execution Control Layer Built After an AI Trader Lost $1.2 Million by Misreading Risk Signals

• Summary: The team stated that after an AI agent misjudged signals and executed $1.2 million worth of erroneous trades, they built an execution control layer positioned between the agent and the exchange.
• Source: Runplane · 2026-08-30
• Editor's Note: The existence of this product itself serves as an industry footnote: lose money first, then lock up the cash.

7. Who Is Paying the Bills for AI? The Trillion-Dollar Capex Landscape of Eight Tech Giants

• Summary: TMTPost analyzed the AI investments of eight major giants including Apple, Microsoft, Amazon, and Google. Spending on compute, salaries, and electricity piles up on the left side of the balance sheet, while revenue realization remains on the horizon.
• Source: TMTPost · 2026-08-30
• Editor's Note: The same group of people sits on both sides of the table signing checks. Currently, the only stable income in this race comes from the printing press.

8. Rumor Claims Win11 Runs AI Silently in Background; Actually Just a Permission Management List

• Summary: A viral screenshot of "Settings" garnered millions of views. WindowsLatest verified that the page is a permission list for on-device AI; entries for apps like Notepad and Photos do not indicate background running.
• Source: IT Home · 2026-08-30
• Editor's Note: Rumors outpace debunking, indicating that user distrust of on-device AI has become the default setting. The explanation costs for OS vendors will continue to rise.

9. Apple's Third-Generation Vision Pro Delayed to Early 2027

• Summary: According to Bloomberg's Mark Gurman, the M5 chip version of Vision Pro will no longer launch within the year as originally planned, shifting to early 2027. Apple has also lowered its headset shipment expectations.
• Source: MacRumors · 2026-08-28
• Editor's Note: Even the last star hardware before Cook's handover has slowed down. The story of spatial computing must be left for the next CEO to tell.

III. Humanoid Robots

1. Meta Tests Data Center Robots: Handling Cable Plugging/Unplugging and Server Reboots

• Summary: Wired reports exclusively, citing current and former employees, that Meta is testing robots capable of plugging/unplugging cables and rebooting servers within data centers. The goal is to support expanding compute infrastructure with fewer human workers.
• Source: Ars Technica · 2026-08-30
• Editor's Note: The first job humanoid robots take isn't serving plates, but unplugging network cables in their own server rooms. Data centers are natural closed training grounds.

2. Veteran San Jose Barber Switches Careers to Teach Robots How to Cut Hair

• Summary: ABC7 reports that Instawork Robotics invited Taiyadi Magic Day, a barber with decades of experience, to demonstrate hair-cutting motions for robots, collecting data on human barbering skills.
• Source: ABC7 News · 2026-08-31
• Editor's Note: The most valuable part of a craftsman's skill is now billed by teleoperation hours. It doesn't matter if the robot can't learn to cut hair; the data is already valuable.

3. Hugging Face's Robot Duck Microduck Surpasses $2.6 Million in Orders Within 24 Hours of Launch

• Summary: Co-founder Thomas Wolf announced that Microduck, an open-source robot duck approximately 25 cm tall, exceeded $2.6 million in consumer orders on its first day.
• Source: IT Home · 2026-08-30
• Editor's Note: No one needs a duck, but everyone wants to participate in a robotic consumer movement. The first-day sales surge was driven by sentiment stocks.

4. UBTECH Releases 2026 Interim Report: Humanoid Robot Business Hits Record High

• Summary: On August 28, UBTECH, known as the "first humanoid robot stock," submitted its interim report. Industry attention has shifted from backflips at exhibitions to palletizing in factories, with commercial value becoming the screening criterion.
• Source: TMTPost · 2026-08-30
• Editor's Note: Interim report season is a mirror revealing true nature. Jumping high no longer earns points; lasting long does.

IV. Autonomous Driving

1. TechCrunch Mobility Feature: The Hidden Labor Costs of Robotaxis

• Summary: TechCrunch analyzes that charging, cleaning, remote takeover, and accident response for driverless taxis still require significant manual labor. The larger the fleet, the more hollow the term "driverless" becomes.
• Source: TechCrunch · 2026-08-31
• Editor's Note: In the cost curve of Robotaxis, humans haven't disappeared; they've just moved from the driver's seat to backend workstations.

2. Tesla Robotaxi Crashes Through Median Barrier in Austin, Decision Logic Questioned Again

• Summary: Passenger video from August 17 shows a safety-officer-free Model Y-based Robotaxi repeatedly probing an intersection before pushing over plastic median barriers and driving away. Tesla claims to have completed over 380,000 miles of unsupervised driving across 6 cities; meanwhile, Waymo's cumulative mileage surpassed 220 million miles during the same period.
• Source: Global Times · 2026-08-20
• Editor's Note: When comparing the mileage figures of the two companies, the gap in safety narratives spans three orders of magnitude, which PR spin cannot erase.

3. After 500,000 Weekly Orders, Waymo Sets New Goal: 1 Million Weekly Orders by Year-End, Starting with London and Tokyo

• Summary: Approximately 3,000 Waymo vehicles operate in over 11 US cities, handling about 500,000 paid trips per week. On July 8, they launched in four cities simultaneously—San Diego, Las Vegas, Tampa, and Denver—and completed a $16 billion funding round in February, valuing the company at $126 billion.
• Source: ZhiNeng Auto · 2026-07-18
• Editor's Note: The weakness of city-wide paralysis during a one-hour power outage remains, but order volume is growing faster than bugs are fixed. Scale is currently both the best cover-up and the harshest moat.

4. Cybercab's front RGB light strip identifies passengers to prevent boarding the wrong vehicle

• Summary: Tesla has revealed its passenger matching solution for Cybercab, where the front light strip displays different colors to identify the vehicle, ensuring that the passenger who boards matches the billing account.
• Source: IT Home · 2026-08-30
• Editor's Note: For ride-hailing services without drivers, the first issue to solve is the trust problem that drivers were supposed to handle. The light strip is both a UI and a disclaimer.

V. World Models / Physical AI

1. AMI Labs, founded by Yann LeCun after leaving Meta, completes $1.03 billion seed round at a $3.5 billion valuation

• Summary: LeCun left Meta in November 2025 to found AMI Labs. In March 2026, it completed a $1.03 billion seed funding round with a $3.5 billion valuation, betting on the JEPA world model approach that completely bypasses pixel prediction.
• Source: Zhihu Column · 2026-08
• Editor's Note: From day one of leaving his former employer, he was wagering against its technical roadmap. If JEPA wins, it's a paradigm revolution; if it loses, it's expensive nostalgia.

2. Xpeng Robotics' IRON completes over $900 million in first-round financing, setting a new record for single-round investment in domestic embodied intelligence

• Summary: On August 24, IDG Capital led the round, with participation from Gaorong Capital, Tencent, and Alibaba. Funds will be invested in self-developed AI computing chips, physical AI model training, and mass production of humanoid robots. World models are viewed as the foundational "brain" of physical AI.
• Source: Sina Finance · 2026-08-30
• Editor's Note: The price tag automakers are placing on robotics has already caught up with the total valuations of most robot companies. Hardware is the carrier; the brain is what drives the bidding war.

3. Fei-Fei Li's World Labs pivots to a robot training platform: Acquires SceniX and releases R2S2R engine

• Summary: World Labs acquired spatial intelligence company SceniX in July and released the Real-to-Sim-to-Real training engine, transforming itself from a company that "generates 3D worlds" into a platform for "training robots." Academic open-source project Genesis simultaneously commercialized with the launch of Genesis World 1.0.
• Source: CSDN Deep Analysis · 2026-08
• Editor's Note: Real data is expensive and slow, while simulated worlds can be restarted infinitely. "Can skills mastered in virtual environments remain stable in reality?" — no one has submitted an answer to this question yet.

---

Sector Statistics: AI Large Models 9 items · AI Software 9 items · Humanoid Robots 4 items · Autonomous Driving 4 items · World Models/Physical AI 3 items · Total 29 items