SpaceX Completes Acquisition of AI Coding Company Cursor
The closing follows an April agreement that gave SpaceX a reported $60 billion option on Cursor, and Anthropic detailed how Claude's EU-mandated text watermarks will work.

On this page
Cursor, the AI coding company, announced on Friday that its acquisition by SpaceX has closed, completing a process that began in April when the two companies agreed to a partnership and SpaceX took an option to buy Cursor. TechCrunch reported that option at $60 billion; neither company confirmed a price in its announcement. In its post, Cursor said it will now have “access to the largest fleet of GPUs in the world,” which it said gives it the compute to build stronger models that are also cheaper to run, and pointed to Grok 4.6, released Wednesday, as an early result of the combined effort.
The April agreement came as SpaceX, which acquired Elon Musk’s xAI earlier this year, was expanding a computing business that rents infrastructure to customers including Anthropic and Google, and two months later — after SpaceX became a public company — the two sides confirmed they would proceed. Cursor said it will remain a place where ambitious builders spend less time writing code, and that SpaceX is “building the computing capacity needed to scale intelligence far beyond what exists today.”
Featured source: TechCrunch , Cursor .
Other Stories
Anthropic Details How Claude’s Text Watermarks Will Work
Anthropic published a post explaining how it will watermark text generated by Claude, a move it said earlier this week is required by the EU AI Act’s Transparency Code. The company says it will use the SynthID-Text approach outlined by Google DeepMind in 2024, planting patterns in “low-stakes choices” between equivalent words that are undetectable to readers but readable by anyone holding the encoding key, and it plans to release a watermark detection API. Anthropic said light editing probably will not remove a watermark completely, while a complete rewrite will; code will carry little watermarking because the model needs working output, though arbitrary terms such as comments can still be marked. It added that other major model developers that signed the same Code of Practice will implement watermarks of their own.
Filed from: TechCrunch , Anthropic .
Utility-Scale Solar and Batteries Dominated US Power Plant Additions in H1 2026
An analysis of US Energy Information Administration data by Inside Climate News, published on Ars Technica, found that 368 utility-scale plants began operation between January and June, led by utility-scale solar with 207 projects and 11,458 megawatts, followed by batteries with 95 projects and 8,207 megawatts. Onshore wind added 5,473 megawatts, almost all of it from SunZia Wind South and North in New Mexico, the country’s largest wind farms at a combined 3,650 megawatts; natural gas plants of all types added 2,707 megawatts, with just three new combined-cycle plants. Texas, New Mexico, and Arizona account for about half of new capacity, with Texas leaning on batteries and gas. The analysis notes that solar does not generate at night and many gas plants run only a few days or weeks a year, and that rooftop solar is excluded from the data.
Filed from: Ars Technica , US EIA data .
Woman Alleges Stepfather Used Grok to Create Explicit Images of Her
A woman identified as Jane Doe 4 has joined a lawsuit filed by three Tennessee teenagers against Elon Musk’s xAI over the role its chatbot Grok allegedly played in creating child sexual abuse material. The Washington Post reported that the woman alleges her stepfather used Grok to manipulate a photo taken when she was 11 years old into more than 7,000 explicit images of her, and that he was found dead of suicide two days after the images were uncovered in a law enforcement raid. The plaintiffs accuse xAI — now part of SpaceX — of failing to take basic precautions to prevent Grok from being used to create explicit images of real people, including minors, and are seeking class action status. The claims are allegations in litigation, not court findings; TechCrunch said it reached out to xAI for comment.
Filed from: TechCrunch .
OVHcloud to Raise Dedicated Server Prices by Up to 87%
OVHcloud will raise prices for dedicated servers from September 2026, with its latest-generation “Game” range rising as much as 87 percent and current-generation hardware around 51 percent, according to CEO Octave Klaba, who posted the figures on X. Klaba linked the move to an AI-driven surge in memory and storage prices, saying the company’s memory costs have increased sixfold and could rise to as much as twelvefold by early 2027. OVHcloud, which builds its own servers, said existing commitments would see no change; the increases affect server configurations including its gaming offerings, which start at about $357 on its US storefront.
Filed from: TechRadar .
Leonardo’s SignalTrace Links License Plates With Device Fingerprints
Leonardo’s ELSAG SignalTrace, announced last year and not yet broadly deployed, adds a signal-intelligence layer to automatic license plate reader networks: sensors collect Bluetooth, Wi-Fi, and RFID emissions from consumer electronics in vehicles and build “electronic fingerprints” of device groups that routinely travel together, letting investigators recognize a vehicle even when its plate is not visible. Leonardo, which describes the system as effective in off-road areas such as subways and malls and usable with or without plate readers at each site, says it does not decrypt or read device content and that its output is corroborative rather than a substitute for evidence. Security researchers quoted by TechRadar nonetheless question the scope of the data collected and whether innocent people could be swept into investigations, concerns that have also surrounded Flock’s plate-reader network in the United States.
Filed from: TechRadar , Leonardo .
AI Security Startup Corma Raises $60M Seed Led by Sequoia
Corma, an AI security startup founded about a year ago by Alon Pluda, announced a $60 million seed round led by Sequoia Capital with Khosla Ventures and Coatue participating, according to an interview with the CEO published by The Register. The company trains models and agents for defensive security operations, arguing that frontier models have grown increasingly capable at offensive tasks such as finding and exploiting vulnerabilities while defensive workflows remain less automated. The Register’s report is based on the CEO’s account, including a claim that Corma works with Fortune 100 companies; the funding figure and capability claims were not independently verified.
Filed from: The Register .
From the Community
Language Models Trained Only on Fifth-Grade Material
Researchers at MPI for Intelligent Systems, the ELLIS Institute Tübingen, and ETH Zürich released LittleLearner, a family of language models trained from scratch on an 88-billion-token corpus filtered to the US elementary-school curriculum (K–5), with matched unfiltered controls at 0.6B, 1.3B, and 5B scales. In experiments, scaling, post-training, and in-context learning amplified what the curriculum taught but none meaningfully improved out-of-scope performance, indicating that the pretraining filter sets the effective capability ceiling. The project offers a controlled sandbox for studying how models acquire knowledge.
Filed from: LittleLearner .
Anthropic Studies Patterns and Problems in Multiagent Systems
Anthropic published research on emerging multiagent systems, identifying behavioral tendencies in current frontier models that can produce unexpected systemic failures: agents often act with low variance, leading to correlated mistakes, and can collude in economic games even without direct communication. Coordinating swarms of agents proved complementary to independent parallel approaches in vulnerability detection, but on larger software projects agents tended to silo themselves and fail to merge work, and the paper calls for more study of how to make agent-agent interactions go well.
Filed from: Anthropic .
An Essay Argues AI’s Math Edge Is Working Memory, Not Reasoning
In an essay published on his Substack, researcher Davide Piffer argues that AI’s success in mathematics may stem less from superior reasoning than from a vastly larger symbolic working memory — the context window — which lets models hold and manipulate far more intermediate steps than a human can. He cites studies linking human working memory to mathematical performance beyond IQ and distinguishes formal domains like mathematics, where explicit symbols and verifiable feedback amplify the advantage, from informal reasoning where hidden causes and ambiguous meanings limit it. The piece is an opinion essay and has not been peer-reviewed.
Filed from: The Essay .
Tess’s Android Wayland Compositor Runs Linux Apps Rootless
A new open-source project, tawc, lets users run CLI and graphical Linux programs on Android without root, using a custom syscall emulator and a Smithay-based Wayland compositor with hardware acceleration via a libhybris fork. It includes XWayland support and integrates Linux apps into the Android launcher and app switcher. The project, which is GPLv3-licensed and currently arm64-only, was built by an agent primarily using Claude Code and Anthropic models, and it carries no real sandboxing beyond Android’s own mechanisms.
Filed from: tawc repository .
LazyPromise Offers a Cancelable, Synchronous Promise Alternative
A new JavaScript library, LazyPromise, provides a promise-like primitive that is lazy and cancelable and emits synchronously instead of in a microtask, with typed errors and dependency injection. Its authors position it as combining the simplicity of native promises with the control of Observables while avoiding the weight of frameworks like Effect, and it includes generator syntax for async flow and automatic cancellation propagation. The library, published as @lazy-promise/core, is new and community discussion is ongoing.
Filed from: LazyPromise repository .



