SpaceX Closes Its $60B Cursor Deal — AI Digest

SpaceX closed its record Cursor deal, Anthropic investors eye a $2T IPO, GLM-5.3 found 2,436 vulnerabilities in open source, and Qwen plus Meta gave local models their best week yet.
A day of big money and big weights: SpaceX closed its record $60B Cursor acquisition, Anthropic investors are eyeing a $2T IPO, Z.ai shipped GLM-5.3 with surprisingly strong cybersecurity skills, and Qwen dropped open weights for a multimodal 27B — local models just had their best week of the year.
Top stories
SpaceX closes the Cursor deal: $60 billion in stock. The acquisition announced in June (CNBC) was completed on August 14: per the Form 8-K filed with the SEC, Anysphere (Cursor) shareholders received roughly 389.3 million SpaceX Class A shares — a $60 billion all-stock valuation. Cursor becomes a wholly owned unit of the new SpaceXAI division (Investing.com). The most popular AI code editor now belongs to a rocket company — and somehow that no longer sounds odd: coding agents and harnesses have become infrastructure, priced accordingly.
Anthropic investors expect a $2 trillion IPO. According to Financial Times reporting relayed by Fortune and Quartz, six of the company's investors expect a valuation of $2 trillion or more at the October listing — the largest IPO in history, above SpaceX's current $1.77T mark. The case rests on revenue velocity: investors project $100–120 billion in annualized revenue by year-end, up from $47 billion in May (Silicon Republic). One important caveat: these numbers come from investors, not the company. More skepticism in Perspectives below.
GLM-5.3: cybersecurity as an RL side effect. Z.ai released GLM-5.3 — same base model as GLM-5.2, with every gain coming purely from scaled post-training RL on long agentic tasks (the open slime + SAO stack). The jumps are real: Terminal-Bench 3.0 at 28.3 versus 4.6 for its predecessor, DeepSWE v1.1 at 66.9 versus 46.2 (SiliconANGLE). The headline surprise is emergent cybersecurity: 84.5% on CyberGym (first place in Z.ai's comparison) and 2,436 vulnerabilities found across 269 open-source projects, 1,097 of them critical or high severity. Per VentureBeat, the model reportedly found a serious vulnerability in Cursor itself. Weights are promised roughly two weeks out, after a safety review.
Local models' best week. Qwen published open weights for Qwen 3.8 27B — a natively multimodal dense model (an FP8 variant is up too; the hosted version carries a 1M context window). Days earlier Meta opened Muse Glimmer 30B — 29.6B parameters under Apache 2.0, under 20 GB at ~4-bit, built for always-on local agentic workflows (NVIDIA's walkthrough). Add the promised GLM-5.3 weights, and local-inference enthusiasts are having a feast (see our curated Open Source section for tools and models to run yourself).
Numbers that matter
- $60B in stock, ~389.3M Class A shares — the closed SpaceX × Cursor deal (SEC 8-K).
- $2T — investors' valuation target for Anthropic's IPO; $100–120B annualized revenue projected by year-end versus $47B in May (Fortune).
- GLM-5.3: Terminal-Bench 3.0 28.3 (up from 4.6), DeepSWE v1.1 66.9 (up from 46.2), CyberGym 84.5%, 2,436 vulnerabilities across 269 open-source projects (Z.ai, SiliconANGLE).
- 27B — Qwen 3.8's newly open multimodal model (Hugging Face); 29.6B / <20 GB / 131K context — Muse Glimmer (NVIDIA).
- 110 countries — the reach of Apple's latest wave of mercenary-spyware threat notifications (9to5Mac).
Perspectives: is $2 trillion for Anthropic justified — or a bubble?
- Optimists (the investors themselves, per FT via Fortune): demand for frontier models is compounding faster than any precedent, and a $100–120B revenue run-rate by December supports a valuation above SpaceX.
- Neutral observers note the target comes from six investors, not the company — management has not endorsed the number (Quartz); going from $47B in May to $100B+ by December is an extrapolation, not reported results.
- Skeptics point to friction inside the tent: per a Reuters exclusive, some investors are unhappy with how the company is handling a defense-contract dispute — managing the government relationship is becoming as much a valuation factor as revenue.
Tools and techniques
- The AI Deputization Audit — a framework from the latest AI Daily Brief: decide what work to hand to AI entirely, what to do side by side, and what to keep human (episode); ready-made delegation pipelines live in our automation templates catalog.
- Matt Wolfe's week-in-models recap — what actually matters in the flood of releases, in 20 minutes (video).
- Pangram — the AI-text detector currently making rounds, including on the latest Hard Fork: $9M raised and a very public war on AI slop (TechCrunch).
In brief
- Prompt injection reached the courtroom: a Connecticut plaintiff hid AI instructions in 3pt white text inside e-filed documents — the court noticed and revoked his e-filing privileges (Volokh Conspiracy, 404 Media).
- Apple sent a fresh wave of mercenary-spyware threat notifications across 110 countries — delivered as lock-screen push alerts for the first time; advice: Lockdown Mode and prompt updates (9to5Mac, Apple).
- Mark Zuckerberg published “The Future Is for Everyone” — a manifesto on “personal superintelligence for everyone,” already being dissected on Hard Fork.
- The Friend 2.0 AI pendant now talks back — and is, per Matt Wolfe, the strangest AI launch of the year (video).