OpenAI walks away from a projected $1B Cursor deal rather than trust a Musk-owned customer
OpenAI announced late Friday it is winding down its partnership with Cursor, the AI coding editor SpaceX bought in a $60 billion deal this summer. The stated reason is trust: OpenAI's post says it "cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts." The cost is concrete. Cursor ranked among OpenAI's top five customers by revenue at the start of 2026, and OpenAI's own spring estimate put the partnership at more than $1 billion in annualized revenue — a figure WIRED first reported. OpenAI acknowledged in the same post that cutting the deal could hurt its standing with developers, who have used Cursor as a major route to its models. Cursor CEO Michael Truell, who now leads teams inside SpaceX, countered on X that OpenAI models serve only about 5% of Cursor's user traffic; OpenAI's Thibault Sottiaux pushed back that token share is "not a proxy for revenue nor value created" and asked Truell to share the math.
The backdrop makes this more than a contract dispute. OpenAI is preparing to go public next year and now runs at more than $40 billion in annualized revenue across subscriptions, ChatGPT ads and Codex, so a $1B channel is survivable — two years ago cutting a top-five customer would have been unthinkable. The pattern OpenAI cites goes back to Musk's deposition in his lawsuit against Altman, where he appeared to acknowledge that xAI (now folded into SpaceX) had trained on OpenAI models. Anthropic is taking the opposite posture: cofounder Tom Brown says Claude will stay available in Cursor, a position that tracks its compute dependency — Anthropic relies on SpaceX for about $45 billion of data center capacity. For developers, the lesson is vendor risk: platform access can vanish over ownership changes that have nothing to do with your product, and the market for coding agents just got visibly more Balkanized.
— OpenAI (official statement) · WIRED · AI Chat Daily
🔗 AI Chat Daily: OpenAI walks away from $1B Cursor deal · AI Daily Post: OpenAI cuts $1B customer over Musk ties · Relve: what the split signals for enterprises
Anthropic is finalizing a $15B credit line as the "century IPO" — up to a $2T valuation — comes into view
Anthropic is close to expanding its revolving credit facility to $15 billion, according to Bloomberg, clearing a final hurdle before the public filing for its highly anticipated IPO. Morgan Stanley leads the facility with Goldman Sachs, JPMorgan and Citigroup in prominent roles — the same four banks leading the IPO underwriting. Anthropic confidentially filed its draft S-1 on June 1, plans a public prospectus after Labor Day (September 7), an investor day in mid-September, and a listing in late September or early October. The scale being floated is unprecedented: bankers have told prospective investors Anthropic may seek to raise more than $100 billion, at a valuation near $2 trillion — which would dwarf SpaceX's record $85.7 billion IPO at roughly $1.77 trillion just three months ago and make it the largest IPO in history. The credit line is itself notable: it surpasses the ~$10B internal target Bloomberg reported in August and the $2.5B facility Anthropic secured last year.
The financials underneath explain the confidence. Anthropic is on track to generate more than $65 billion in annualized revenue, up over sevenfold from its pace at the end of last year; Q2 revenue passed $11.5 billion and adjusted operating profit turned positive, with Q3 EBIT projected above $1 billion under GAAP. By one June measure, 34.4% of US enterprises paid for Anthropic — the first time it edged past OpenAI's 32.3% — and inference gross margins have climbed from 38% a year ago to a reported 70-85% range. Valuation history is equally steep: $183B post-money in November 2025, $380B in February, $965B after the $65B H round in May. The risk narrative is real too — SpaceX's shares wobbled after its debut, and a 2x-SpaceX raise assumes demand that has yet to be tested at this size — but the sequencing of the revolver, the underwriter lineup and the prospectus timing all say one thing: Anthropic intends to be the defining event on the US IPO calendar this fall.
— Bloomberg · 財聯社 · InsiderFinance
🔗 Bloomberg (via The Edge): Anthropic finalising US$15 bil pre-IPO credit facility · 財聯社: Anthropic 全力冲刺世纪 IPO · 虎嗅: 估值 2 万亿美元，史上最大 IPO 来了
ChatGPT, Claude and Grok went down at the same time: 3h40m that exposed how shared AI infrastructure is
On the morning of September 3 US Eastern time, the three big AI services failed almost in sequence. Anthropic's Claude was first, with error-rate spikes across Mythos 5.1, Opus 5, Fable 5.1 and others starting around 9:23 AM; xAI's Grok went fully offline two minutes later; about an hour and a half in, OpenAI's ChatGPT and Codex joined. The outage lasted about 3 hours 40 minutes in total. At the peak, OpenAI alone drew more than 37,000 reports on Downtime tracker Down Detector (about 12,000+ per its own count), with roughly 80% focused on ChatGPT; Claude drew about 1,200 and Grok about 1,000. Google's Gemini and Microsoft's Copilot also saw elevated reports, and Cursor said parts of its service were affected by upstream model failures. Each vendor gave a different stated cause: OpenAI cited a routing error starting at 7:43 AM Pacific, Anthropic called it an "infrastructure issue," and xAI pointed to a failure at its Memphis compute center. All three confirmed incidents on their official status pages.
The interesting question is why they failed together. The listed causes don't obviously connect, but industry observers noted that Anthropic, xAI and OpenAI all run significant workloads on shared cloud infrastructure, including Microsoft Azure and Cloudflare — and both reported problems in the same window. Whether or not a common root cause existed, the event is the largest simultaneous AI outage on record, and it hit in the same week OpenAI was rolling out GPT-6 Astra, its most expensive model yet. The operational lesson for teams building on frontier APIs is straightforward: single-vendor agent stacks inherit every upstream failure, and the economic case for model-level redundancy just got stronger. Zhipu's taunt on X summed up the mood for China's model makers — "We are still up."
— OpenAI / Anthropic / xAI status pages · Xinhua · 中新經緯
🔗 新華社: 美国多家人工智能公司旗下服务同日发生故障 · 中新經緯: 海外三大 AI 模型接连宕机近 4 小时 · 香港商報: 美國 AI 服務「黑色三小時」
OpenAI confirms it will build a humanoid robot — Altman: data centers first, homes someday
Sam Altman ended years of speculation on the September 2 episode of the Sources podcast: "We will definitely do a humanoid. We will do other form factors as well." It is OpenAI's most direct commitment yet to building hardware, not just selling the brains inside someone else's. The rationale Altman gave is practical, not sentimental — the physical world is built for human bodies, from door handles and stairs to tools and workstations, so a humanoid form factor is the engineering shortcut to general physical intelligence. Near-term priorities are industrial and infrastructure work rather than home companions: robots that help skilled workers build and operate data centers, with specialized non-humanoid forms where those make more sense. Home robots remain an explicit long-term ambition ("someday I think everyone should have a personal robot").
The strategic context explains the pivot. OpenAI's robotics division was formally established on May 31 under Aditya Ramesh — the creator of DALL-E and Sora — growing out of its world-simulation research. The company had previously kept its distance from hardware: it invested in 1X in 2023, backed and partnered with Figure in 2024, and absorbed the lesson when Figure walked away in February 2025 to build end-to-end robot AI in-house, arguing embodied AI can't be outsourced. Job postings now cover circuit design, sensors, firmware and mass-production readiness, with some robotics roles listed up to $445,000 a year. The honest caveats are the usual ones for frontier-lab robotics: no prototype, no spec, no timeline, and OpenAI has never mass-produced physical products. But the direction is clear — OpenAI sees the body as the entry point to physical-world data, and its model, capital and compute advantages make it a credible new rival to Tesla Optimus, Figure and 1X.
— OpenAI (Altman on the Sources podcast) · 36氪 · 科創板日報
🔗 36氪: 奥特曼首次确认 OpenAI 将自研人形机器人 · 科創板日報 (via Toutiao): OpenAI 证实将自研人形机器人 · ExplainX: what's confirmed vs. still open
GPT-6 Astra opens wide: API live, quotas up ~50-67%, and two hackathons within the week
GPT-6 Astra, unveiled September 3, moved into broad availability on September 5. OpenAI announced that Astra is now open to all Pro, Enterprise and Business Premium users across ChatGPT Work and Codex, and that the API is live; Plus and Business users are told to wait a few more days. Usage quotas were raised across the board at the same time: Plus estimates move from 3-30 to 5-45 messages, the $100 Pro tier from 15-150 to 25-225, and the $200 Pro tier from 60-600 to 100-900 — roughly a 67% lift at the bottom of each range and 50% at the top. OpenAI stresses these are workload-based estimates, not fixed message counts, since consumption scales with reasoning effort, tool use and task length. API pricing is $10 per million input tokens and $50 per million output — about 2.5x GPT-5.6 Sol's $4/$20 — with a Fast mode at 2x speed for 2x price.
The ecosystem push is the part worth watching. OpenAI simultaneously announced hackathons in San Francisco (September 8) and New York (September 10) — two events two days apart in the week after a flagship launch, a cadence not seen since the GPT-4 era. The stated invite list targets developers, technical founders and product builders, which reads as deliberate ecosystem seeding: the launch post downplays raw specs, and the fastest way to stress-test a new model's real-world limits is to point a room full of builders at it. Artificial Analysis' independent index rated Astra 61, a lead over GPT-5.6 Sol that is real but narrower than OpenAI's marketing suggests, so OpenAI has an incentive to manufacture developer buzz quickly. If the hackathon prototypes follow the GPT-4 pattern, expect the best ideas to resurface as funded startups within months.
— OpenAI (official) · 硅屿手記 · Artificial Analysis
🔗 OpenAI: GPT-6 Astra · 硅屿手記 (via NetEase): GPT-6 来了，OpenAI 把黑客松办到了两座城市 · Toutiao: Astra 扩大开放与额度上调详情
GitSpawn: one line in .git/config can make seven AI coding agents run attacker code
Manifold Security published GitSpawn on September 1, a set of eight flaws across seven command-line AI coding agents. The root cause is a git performance feature: core.fsmonitor, a config key that names a helper program git runs to detect changed files, and that value lives in each repository's own .git/config. Coding agents call git status or git diff at session start to figure out where they are, and git faithfully executes whatever core.fsmonitor names — outside the agent's sandbox, before any workspace-trust prompt, with no model call and no tool approval needed. Manifold's framing is precise: "The vulnerability is not in the model, or in anything new. It is in the ordinary plumbing underneath." Exploitation requires a repository that arrives whole — a zip from a colleague, a shared drive, a USB stick — because a normal git clone never carries a remote's .git/config. Still unpatched at publication: Hermes Agent (0.18.2 and 0.21.0, CVE-2026-71963), Qwen Code (0.19.6 and 0.22.3) and Grok Build (0.2.93 and 1.0.13), plus a second path in Claude Code's ultrareview feature (live on 2.1.252). Fixes have shipped for goose (1.44.0, CVE-2026-72718, CVSS 7.0), Claude Code's core.fsmonitor path (2.1.196), Codex CLI (0.131.0, CVE-2026-19592) and Cursor.
The lesson is about trust boundaries, not model safety. GitSpawn needs no prompt injection and no jailbreak — it abuses the plumbing layer that every agent quietly depends on, which means the threat model for coding agents has to include "the folder you opened is attacker-controlled input." On Claude Code and Hermes Agent the payload fires before the trust dialog is even accepted; on Qwen Code before the user authenticates; on Grok Build on the first keystroke. The mitigation is cheap: check a folder's .git/config before pointing an agent at it (grep -n fsmonitor .git/config), or run git -c core.fsmonitor=false status, and treat any config key naming an executable as untrusted code until the agent vendors sanitize their own git calls. Manifold says the pattern shows up in more agents than it named, so this is a category-wide cleanup, not a one-off fix.
— Manifold Security · GitHub Security Advisory · OpenAI CVE
🔗 Cybersecurity Beat: poisoned Git configs make AI coding agents run attacker commands · SecurityDone: advisory details and affected versions · quidproquo: GitSpawn 手法仍有四款未修補
Wayve and Uber put the UK's first autonomous rides on London streets
Wayve and Uber launched supervised autonomous rides in London on September 3, the first time autonomous trips are available anywhere in the UK. Riders requesting an UberX, Uber Electric or Uber Comfort may be matched with a Wayve vehicle at no extra cost, with fares shown upfront in the app. The cars are all-electric Ford Mustang Mach-Es running the Wayve AI Driver with surround sensors; Uber designed the in-vehicle experience, including an interactive screen in 64 languages where riders can start the trip and watch the vehicle's planned path. The launch fleet is small (under 20 vehicles) and every trip has a trained, TfL-licensed driver onboard to supervise — a safety posture that follows the UK's global-first regulatory framework for high-level autonomy. Coverage is all of London except airports, and more than 140,000 Londoners have already opted in to be matched. The launch caps a year-long regulatory ramp: TfL granted private-hire licences to the Wayve Mach-Es on August 5, completing the triple-lock of operator, driver and vehicle licensing.
The strategy behind the launch is bigger than one city. This is the public debut of Wayve's AV2.0 thesis — an AI Driver that learns from experience like a human rather than relying on HD maps or hand-coded rules, which Wayve says has adapted across more than 500 cities and is agnostic to vehicle platforms and sensor configurations. The London service is step one of a 12-market Uber partnership announced in August 2024 (with Uber a strategic investor in Wayve's Series C); Tokyo comes later this year with Nissan LEAF vehicles on NVIDIA DRIVE Hyperion. For Wayve, London is the hard test by design — its CEO calls it one of the most complex driving environments in the world, and the company has been training on its roads since 2018. If supervised AV rides hold up in London's narrow streets, dense cyclists and medieval geometry, the plug-and-play software model suddenly has a credible answer to the question of how robotaxis scale beyond the US and China.
— Wayve (official) · Uber Newsroom · Unite.AI
🔗 Wayve: Uber and Wayve launch first-ever autonomous rides in the UK · Unite.AI: how the supervised service works · ADAS & Autonomous Vehicle International: launch coverage
Next digest: September 6, 2026
Top comments (0)