SpaceX Acquires Cursor For Sixty Billion Dollars In Major Tech Deal On Record
English version above, Chinese version below.
Trend One: Open-source coding models close the gap without changing base models
On August 14, Zhipu AI released GLM-5.3, keeping the same base model as GLM-5.2 and relying entirely on post-training scaling. Coding ability rose 50% in its internal benchmark, Terminal-Bench 3.0 jumped from 4.6 to 28.3, and DeepSWE v1.1 went from 46.2 to 66.9. Zhipu said it will open the weights two weeks after release, following a security review.
The release came alongside a run of open-source coding models. DeepSeek shipped V4 Pro, a 1.6T-parameter MoE model with 49B active parameters and 1M-token context. On LiveCodeBench it scored 93.5, with a Codeforces rating of 3206. Alibaba open-sourced Qwen3.8-27B on the evening of August 14, a 270B-parameter native multimodal dense model that beats Qwen3.7-Plus in coding and office tasks. Alibaba reports Qwen has passed 3 billion downloads with over 300,000 derived models, and 460+ open-sourced models.
The pivot is uniform: none of the three changed base models. Gains came from post-training and architecture. GLM-5.3 uses only 50,000 output tokens per task on Z.ai Code Bench, against roughly 120,000 for Claude Opus 4.8, while reaching 31.4% accuracy versus Opus 4.8's 29.5%.
Trend Two: DeepSeek moves the race from models to the Agent runtime
On the night of August 13, DeepSeek open-sourced DeepSeek Harness under MIT license, with no new model. The framework splits the entire Agent skeleton — main loop, state management, session storage, task scheduling, sandbox, events, UI — into pluggable components, under a "Everything is a Plugin" design. It runs on a custom Cordis plugin base with hot load/unload and rollback, plus Preset scoping that runs multiple isolated Agents in a single process.
The same night as the Harness-V4 Pro pairing, Honor announced on August 15 that YOYO Claw (its "Shrimp Brain") now runs GLM-5.3, a direct consumer-side rollout of an open model. The shift is visible: model scores are converging, and the differentiation layer is moving to the runtime and the tooling that wraps models.
Trend Three: Compute power reshuffles ownership
On August 14, SpaceX closed its $60 billion acquisition of Cursor, one of the largest tech deals on record. Cursor's blog cited access to the largest GPU cluster to build lower-cost models. The two had already shipped the joint Grok 4.5 model in July and Grok 4.6 afterward, and Musk told employees AI revenue would exceed all other SpaceX business by September.
The same window saw Nvidia reconsidering its exposure. Nvidia holds roughly a $21 billion SpaceX stake converted from its xAI position after the SpaceX merger, with SpaceX data centers using Nvidia exclusively. Separately, reports point to Nvidia cutting its guarantee on OpenAI's Ohio 10GW project — from $250 billion to under $120 billion, covering only the first 5GW — and OpenAI's annualized revenue passing $40 billion, roughly double its prior run rate. Anthropic, which said its run rate topped $47 billion in May, is reported near a $6 billion acquisition of Decart, a company whose software runs training and inference more efficiently on chips.
Meanwhile, on August 14, Anthropic CEO Dario Amodei posted on AI regulation, arguing fair institutional process can itself disperse power, and that regulation should bind frontier companies while leaving room for open models and small players.
Trend Four: AI reaches unsolved mathematics
An August 14 report detailed how Jinshanmu, a neurosurgery resident physician at Peking Union Medical College Hospital, used OpenAI's GPT-5.6-Sol to prove the Crouzeix conjecture in about 16 hours — a problem open since 2004. The conjecturer Michel Crouzeix, and mathematicians Alex Townsend and Anne Greenbaum, reviewed the manuscript and confirmed the proof. The prior best constant was 2.414, set in 2017 after a week-long workshop; Crouzeix himself had proved 11.08 in 2007. Eight days later, mathematicians Emiel Lorist and Felix Schwenninger posted an independent five-page proof using a different method, also aided by ChatGPT 5.6.
Nvidia added hardware to the efficiency story on August 14, announcing full production of the Spectrum-X CPO switch — the first mass-produced 200G/lane co-packaged optics Ethernet switch. Power drops to one-fifth, laser count to one-quarter, and AI application uptime extends fivefold. The SN6810 fits 128 ports of 800 Gb/s in a 2U chassis (102.4 Tb/s), with TSMC, SPIL, Lumentum and Foxconn in the supply chain.