SpaceX completes its Cursor acquisition — a $60B all-stock deal: The process that began in April is done. Cursor gains access to the world’s largest GPU fleet, in what a16z calls the largest venture-backed startup acquisition ever. Cursor · a16z
China’s Zhipu ships GLM-5.3, its “strongest open-weights coding model”: Same base as GLM-5.2, with all gains from extended post-training; weights go open source in two weeks after a security review. The Decoder
Meta gives up control of AI startup Manus after China blocked the deal: The acquisition was reversed after Chinese regulators cited export controls; Manus returns to independence and must delete some data gathered under Meta. Computerworld
French startup Kog squeezes more inference out of existing GPUs: Software optimization on AMD MI300X and Nvidia H200 hardware; the solo founder reports 200 business leads, with software engineering as the first use case. TechCrunch
Google open-sources HEIR, a compiler for private AI via homomorphic encryption: A new tool in the Private Computing Toolkit that unlocks cryptographically-secure inference. Google
Compute owners are buying the application layer: SpaceX acquiring Cursor shows frontier-compute holders moving to own the coding-tool “killer app.” a16z · Cursor
Open-weight models gain “cyber” capability: GLM-5.3 was trained to find vulnerabilities and now reasons across “multiple stages of exploitation” — a new variable in open-model security. The Decoder
Inference on encrypted data becomes practical: Homomorphic encryption is moving from slow research to a usable open-source compiler, opening the door to privacy-preserving AI products. Google
The gap for neutral, self-hosted coding tools: With Cursor inside SpaceX, teams wanting an ecosystem-neutral option can lean on open-weight models like GLM-5.3. a16z · The Decoder
An app layer that infers over encrypted data: Health, finance, and legal products are natural experiments for HEIR-style privacy-preserving inference. Google
Fix latency bottlenecks in software, not silicon: Kog’s approach shows teams can unlock speed and cost gains on GPUs they already own for latency-sensitive agents. TechCrunch
Cursor’s independence and neutrality: Whether Cursor stays model-neutral or reorients around SpaceXAI models is an open question. a16z
GLM-5.3’s cyber capability, once open-sourced: A model that plans “complete exploitation chains” could aid offense as well as defense when the weights drop in two weeks. The Decoder
US-China AI regulation is chilling cross-border deals: The Meta-Manus reversal shows both sides tightening; cross-border AI transactions are getting harder. Computerworld