OpenAI just pulled the plug on Cursor
· The Aigentic · Morning Brief
OpenAI said Friday it will stop supplying models to Cursor, now owned by Elon Musk’s SpaceX, with a proposed shutoff on November 12, citing a change-of-control clause and past contract breaches by Musk-owned companies, Reuters reports. Anthropic said it will increase Claude compute for Cursor; Musk replied he “couldn’t care less.”
Amir Efrati and Stephanie Palazzolo at The Information write that Cognition, maker of the Devin coding agent, is running near $900 million in annualized revenue while projecting roughly $800 million in cash burn this year, mostly on leased Nvidia servers, and is raising again toward a valuation near $45 billion.
Nvidia-backed neocloud Lambda raised about $1 billion of short-dated private debt arranged by JPMorgan to buy Nvidia GPUs that Microsoft will lease, Bloomberg and TechCrunch report, on top of a separate $926 million term loan closed this week.
The Wall Street Journal says Nvidia paused some AI Compute Partnership revenue-share deals after employees flagged antitrust and control worries, with about $36 billion in related commitments disclosed this week, per Reuters.
Leo Schwartz and Qianer Liu at The Information report Commerce is drafting a slimmed replacement for the rescinded AI diffusion rule aimed at Chinese firms renting advanced U.S. chips remotely through third-country data centers, with industry feedback possibly as soon as September.
1. OpenAI pulls Cursor’s model supply
OpenAI said Friday it notified SpaceX that it intends to wind down the contract supplying OpenAI models to Cursor, the AI coding tool SpaceX finished buying earlier this month in an all-stock deal valued at about $60 billion. The proposed shutoff is November 12, 2026, the maximum notice OpenAI says its change-of-control clause allows, Reuters reports.
OpenAI did not accuse Cursor of a breach. It said it cannot be confident SpaceX will stay inside OpenAI’s terms of service, pointing to past violations by Twitter after Musk’s acquisition and to Musk’s sworn admission that xAI, now also under SpaceX, had violated similar terms. Cursor co-founder Michael Truell, now a SpaceX executive, said the teams are talking. Anthropic’s Tom Brown said Anthropic will increase compute so Claude stays available in Cursor. Musk posted that he “couldn’t care less.”
The one move is to treat November 12 as a hard migration date for any workflow that still routes through OpenAI inside Cursor, and to confirm Anthropic (or another provider) is already the default for the jobs that matter.
2. Devin prints revenue and still burns silicon
Amir Efrati and Stephanie Palazzolo at The Information report that Cognition, the San Francisco company behind the Devin coding agent, is generating about $900 million in annualized revenue, up more than threefold since the start of the year and about 83 percent above the $492 million ARR it disclosed with its May raise.
The same exclusive says Cognition is on track to burn roughly $800 million in cash this year, including about $200 million in the second quarter alone, largely on leased Nvidia servers that run Devin and train in-house models. Enterprise gross margins sit near 50 percent, and ex-training the business is nearly cash-flow neutral. Cognition is raising again in a round that could value it around $45 billion, after a $26 billion May post-money, with named customers including BNY, Santander, and Mercedes.
The one move is to separate the product story from the training bill: the coding agent layer can finally show real ARR, and the GPU lease still decides whether that ARR compounds or evaporates.
3. Lambda borrows a billion so Microsoft can lease the GPUs
Bloomberg reports that Nvidia-backed neocloud Lambda raised about $1 billion of private short-dated debt arranged by JPMorgan and sold to private-placement investors. The money buys Nvidia GPUs that Microsoft will lease under a multi-year infrastructure deal that already contemplated tens of thousands of chips.
The raise sits beside a separate $926 million senior secured term loan B Lambda closed this week for GB300 capacity tied to an investment-grade offtaker, and beside talks for as much as $3 billion of pre-IPO equity after a $1.5 billion round at a $5.43 billion post-money last November. Bloomberg’s tally puts global AI-related debt above $400 billion for 2026 so far.
The one move is to watch the offtaker, not the logo on the rack: Lambda is financing chips against Microsoft’s willingness to keep paying, which is how neoclouds are turning hyperscaler demand into bankable paper.
4. Nvidia hits pause on the revenue-share cloud deals
The Wall Street Journal reports that Nvidia stepped back last week from some deals under its July AI Compute Partnership, the financing program that offered credit support and capacity buybacks to AI cloud providers in exchange for a cut of their cloud revenue. A Reuters write-up of the Journal exclusive says some employees flagged antitrust risk and unease over Nvidia dictating approved renters and preferring capacity spread across smaller firms.
Under proposed terms, Nvidia would set a base hourly rate covering provider costs and take 50 percent of revenue above that line. Nvidia disclosed about $36 billion in related commitments this week, and CFO Colette Kress told investors the model could still generate billions over the medium and long term. An Nvidia spokesperson said the July business model “is still in place and continues to evolve due to high demand,” and the company may revamp the initiative or fold it into another program.
The one move is to treat the pause as a control and antitrust problem, not soft demand, and to ask any neocloud pitch whether its financing still assumes Nvidia as both supplier and revenue partner.
5. Commerce wants to close the remote-chip loophole
Leo Schwartz and Qianer Liu at The Information report that a small group inside the Commerce Department is drafting a slimmed-down replacement for the rescinded Biden-era AI diffusion rule. The new draft is aimed at Chinese AI firms that rent advanced U.S. chips remotely through data centers in third countries instead of importing the hardware.
The Bureau of Industry and Security could circulate a draft to AI companies and other stakeholders as early as September. That would shift export control from who physically holds the GPU to who can log into it from abroad, and it lands on the same remote-access fight that earlier Remote Access Security Act drafts tried to close in Congress.
The one move is to inventory any overseas GPU rental that Chinese teams can reach from a browser, and to assume September industry comments are the start of a compliance project, not a rumor.
Watch
Matt Wolfe — OpenAI’s Jalapeño chip vs Nvidia
Matt Wolfe walks through OpenAI’s first public Jalapeño inference-chip results, the custom silicon OpenAI is building to cut reliance on Nvidia for serving prompts, not for training new frontier models.
OpenAI published numbers on open-weight models so outsiders can re-run the tests, and Wolfe notes the company says the gap widens further on its own closed models. He pairs that with the still-unconfirmed Information report that Nvidia agreed to buy Hugging Face, the default host for open-weight files, and with Cerebras-style ultrafast inference as the other escape hatch from a single GPU vendor.
The one move is to separate inference independence from training independence: Jalapeño is a serving story, and Nvidia still owns the training narrative until a lab ships something else.
Nate B Jones — friction-maxxing against AI brain rot
Nate B Jones walks through friction maxxing, his deliberate practice of making AI harder to use so the work sharpens judgment instead of replacing it.
Most AI use removes friction: ask, accept, move on. Jones runs the opposite loop every day across Codex, Grok, Claude, and about ten trusted humans, hunting for the disagreement that breaks the first easy answer. He treats the MIT “Your Brain on ChatGPT” paper as a preliminary setup the researchers themselves refuse to brand as brain rot, then shows how surviving four to ten rounds of argument changes the output.
The one move is to ask, after every AI reply, whether you feel more capable or less, and to force at least one second model or human to attack the draft before you ship it.
Alex Finn — Omarchy, DHH’s AI-native Linux
Alex Finn walks through Omarchy, the free open-source AI-first Linux desktop from David Heinemeier Hansson, the creator of Ruby on Rails and Basecamp, installed on Finn’s seven-year-old Dell Alienware.
Omarchy ships with ChatGPT desktop, Claude, Grok bots, and Hermes agents built in, and it is keyboard-first with tiled windows instead of floating ones. Because the OS is open source, Finn asks a Hermes agent to add close buttons on critical notifications and asks a “Machi Dev” Grok bot for a Super-E keybind that equalizes every tile, edits that take about thirty seconds. He flashed the installer from a USB drive with an AI walking the steps.
The one move is to try Omarchy on a spare machine first: the point is not Linux cosplay, it is an OS your agents can rewrite while you work.
The AI Advantage — Claude’s browser vs ChatGPT Work vs Grok Bot
Igor at The AI Advantage walks through the new wave of cloud browsers and cloud computers bundled into ChatGPT Work, Claude Cowork, and Grok Bot after a round of price cuts.
He separates old web search and desktop-extension remote control from a real browser that lives in a rented machine and keeps working while you are away. Demos include auditing other subscriptions from a prompt and letting the agent open YouTube Studio analytics and build an artifact without a click-by-click script.
The one move is to pick one recurring browser chore this week, hand it to the cloud browser with the least-privileged login you can spare, and see whether the artifact comes back without you babysitting the tab.