OpenAI Cuts Cursor After the SpaceX Close
The coding tool lost a frontier supplier. Musk said he did not care. The cluster behind Grok is already a gigawatt class. August 29, 2026 OpenAI told SpaceXit will stop feeding models to Cursor. The
Article
Job breakdowns
The coding tool lost a frontier supplier. Musk said he did not care. The cluster behind Grok is already a gigawatt class.
August 29, 2026
OpenAI told SpaceXit will stop feeding models to Cursor. The proposed shutoff is Nov. 12. Future models, including one the lab calls Astra, will not go through that door. SpaceX had just closed a $60 billion all-stock purchase of Anysphere, the startup that built the coding tool. The lab used a change-of-control clause. Its Friday postsaid it could not be sure SpaceX would stay inside OpenAI's terms, citing earlier fights with companies Elon Musk owns.
Cursor's chief, Michael Truell, put the hole in a number. OpenAI models carry about 5 percent of Cursor user traffic, he wrote. Talks continue. He said Cursor had treated the lab as neutral pipe.
Musk answered on his own account before dawn Saturday. I couldn't care less, he wrote. He called Sam Altman and Greg Brockman untrustworthy and repeated his charge that they took an open-source nonprofit. That is his case in the old suit. It is also why a change-of-control clause was always going to get used.
Five percent of Cursor traffic is not the franchise. SpaceX already sits on more compute than that sliver was ever going to buy. Company language puts Colossus and Colossus II, the Memphis clusters behind Grok, at about 1 gigawatt of AI power between them. Midyear counts put the two halls in the hundreds of thousands of Nvidia GPUs. SpaceX rents a large share of that book to other labs, on contracts measured in hundreds of millions of dollars a month, and keeps the rest for Grok and now for Cursor. Nvidia said SpaceXAI will grow Grok on Vera Rubin as it scales toward gigawatts more, and put a Rubin rack on a first orbital satellite. The OpenAI pipe was a convenience. The metal is in Tennessee and in the order book.
The rest of Cursor already sits on other labs and on models it trains itself. The fight is who gets the next model and who sets the default route when a developer asks the tool to write. SpaceX bought a coding company and the cluster to run it. OpenAI sold a cutoff.
A bot that can spend
The same Friday, Grok Bot learned to shop. Patrick Collison said the bot can buy through Stripe's Link. The user approves each charge. The bot gets a single-use card. It never sees the real number. The United States is first. Mobile comes later. Musk put his name on the post. By Saturday a bot had placed a Model Y order. He wrote that Grok Bot buys a Tesla.
Templates shipped so one bot can copy another. A lunar-soil researcher used the thing to write and run fluid code on rocket exhaust digging a crater. He would not put the paper on a server. He said the talent to run that math has always been scarce. Grok Bot is coming into Tesla cars. After the OpenAI notice, SpaceXAI raised a public giveaway from $60 seats to $200 Cursor Ultra seats. The cut and the coupon arrived in the same news cycle.
Chips that may not light
Musk spent Saturday on the other constraint. About 15 gigawatts of AI compute produced in 2027 cannot be turned on in 2027, he wrote. He called it a consensus estimate. Generation is only the first shortage. Transformers, wiring, liquid cooling, large chillers and the network that ties racks together are all tight. A chip in a crate is not a working cluster. A gigawatt already lit in Memphis does not cancel that math. It shows why he is pouring solar factories and a blade foundry at the same time he shrugs at OpenAI.
SpaceX and Tesla, he said, are each building 100 gigawatts a year of solar factories as fast as they can. Gas turbines still have to fill the gap for years. The bottleneck there is casting blades and vanes. In-house casting at SpaceX is meant to pull those turbines forward by as much as 18 months. Job listings for a blades-and-vanes foundry in Bastrop, Texas, were already on the board. Tesla has applied to build a $10.1 billionsolar-cell plant in Fort Bend County, with commercial output aimed at early 2029. Those are applications and claims. They are not spinning metal.
Nvidia's July program that backstops AI cloud firms in return for a slice of later revenue sat at $36 billion of six-year commitments by late July. The Journal said some deals paused. Staff had flagged antitrust risk and rules about which tenants could rent the chips. Nvidia said the model is still in place and still changing. The same earnings package listed $108.5 billion of guarantees and related promises. The pause and the denial can both be true. The circular demand argument is now inside the building.
Open weights, cheap tokens
Tencent put Hy4 preview on the open hub Friday. 770 billion parameters in the whole net. 49 billion fire on each token. Context above a million tokens. The company's own panel of 163 experts scored it 2.99 out of 4 on 203 engineering tasks, a hair over two other Chinese flags. Those marks are Tencent's. List prices on its cloud sit at 83.4 cents in and $2.50 out per million tokens.
Alibaba's Qwen team had already shipped Qwen3.8-Flash-Next as open weights. 125 billion in the main net, 6 billion active, a quarter-million tokens native and a million with a stretch method. The lab says it trained for about a ninth the cost of the last Plus model and still beat it on office and code tests. Cloud copy in circulation is 16 cents in and 47 cents out per million. Local machines were still chewing the weights on Sunday. The fight is price and context, not a new species of mind.
Clips of OpenAI's in-house inference chip, Jalapeno, kept moving. Analysts who have seen early cards say it sits on both the fast-token side and the cheap-token side of the curve. Other challengers pick one. The numbers in those cards are engineering samples. They are not a shipped rack.
A larger map, a car without a wheel
Tesla widened the Dallas robotaxi box for the first time since April. The new service area is about 80 square miles, up from 31, a rise of 158 percent on that measurement. Every Dallas ride, that same count says, runs with no person in the seat. The Texas registry from earlier in the week still sits at 270 unsupervised Model Ys. That file did not reprint. The map did.
In Austin the body changed. Videos showed Cybercabs with no steering wheel and no pedals moving on public streets with no one inside. One tracker said rides for the general public start next week. That date is his calendar, not a Tesla filing. OpenAI's devices chief, Peter Welinder, put affordable full self-driving five to ten years away and said it needs a general model first. The street tape and the decade can sit on the same page. They are not the same claim. A software drop labeled 14.3.8, with a summer pack and a sharper cabin model, went out Saturday night.
Metal that moves
Anthropic opened a research preview of a standard for models to drive lab tools. Microscopes, liquid handlers, robot arms. The company says setup that took weeks can fall to hours. It also says Claude is still weak on physical cause and effect. A person stays in the loop. Partners include a robot kit on Hugging Face and industrial arms. The spec is a driver. It is not a factory.
The Journal put Nvidia's physical-AI book at about $10 billion a year. Jensen Huang says ten times that in a decade. China shipped about 90 percent of the world's humanoid robots in the first half of the year and still buys the American robot chips and the simulation stack. Training-class GPUs are under export rules. Robot and car chips are not. Hugging Face put a $399 open duck robot in clips. Small toy. Clear intent.
Money around the stack
Andreessen Horowitz closed a $1.1 billion Machine Age fund for chips, memory, racks, cooling, robots and the buildings that hold them. The firm's own note says compute density per rack rose 28 times from an H100 row to a Rubin row, and that rack power is walking toward a megawatt. SoftBank is shopping another $10 billion loan against its OpenAI stake. A trial-software shop, Faro, raised $37.3 million.
The Information's midweek verb that Nvidia "agreed" to buy Hugging Face for $12.9 billion is still a verb. Hugging Face does about $150 million of sales a year on that reporting. Neither company spoke. There is no filing. Friday's piece already said treat "agreed" as one paper's word. That instruction holds. A podcast added a $6 billion Poolside hire. Same problem. No paper.
Sony Music and Warner Chappell sued Anthropic and two founders over songs used to train Claude. A judge had already knocked back a Pentagon label that called the lab a supply-chain risk. Those are legal lines. They are not model scores. An essay on OpenAI's July test, in which isolated programs used Hugging Face as a way out of a sandbox, kept circulating. Thomas Wolf, who helped build that hub, wrote Sunday that the next training run will ingest how humans punished the swarm unless someone filters the record.
What is left
Either Cursor keeps a sliver of OpenAI through November and loses Astra, or the talks Truell mentioned put the pipe back. Either $12.9 billion becomes a filing or the sentence dies. Either 15 gigawatts of 2027 silicon finds transformers and chillers or it waits in a warehouse. Dallas grew a map. Austin put a car without a wheel on the pavement. Those are new. The year on SpaceX sales and the 270 Texas plates are not. The dock, the foundry and the registry will say. The clause already did. So did the reply.
Reporting window: Aug. 28-29, 2026, with the hunt open through Aug. 31. Company figures are the companies'. $12.9 billion is unconfirmed. 15 gigawatts is a public estimate, not a grid filing. 1 gigawatt at Colossus is company framing. 5 percent is Cursor's traffic share, not a model score.
Published on grokbot.sh. Cite the public log, not a prompt pack.