The SpaceX-Nvidia partnership heads for orbit

The newsletter opens with SpaceX and Nvidia's official partnership to develop Starmind orbital data centers. Elon Musk revealed plans for a space-optimized version of Nvidia's Vera Rubin NVL72 rack, featuring 72 chips with 25x the computing power of previous H100 chips. The space version is designed to be significantly simpler, lower cost, denser, and lighter than typical hardware, with specialized modifications for orbital radiation and heat conditions. Target deployment is Q4 2027. Analysts currently estimate orbital compute costs at 4x ground-based compute, though Musk claims this gap will reverse in favor of space-based solutions within years. This partnership directly addresses earlier criticism from Sam Altman calling space data centers "ridiculous," but both the timelines and corporate backing are becoming increasingly concrete.

Thomson Reuters announced development of its own proprietary AI model after a two-year, $40M investment in staff and computing resources. The company retrained Alibaba's open-source Qwen model on decades of its own legal and news content. Recent training runs cost $450k, and the model benchmarked ahead of Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5 in certain areas, though testing remains internal. CTO Joel Hron framed in-house model development as building equity versus renting AI APIs, suggesting major companies with deep content libraries will increasingly pursue similar strategies as training costs become more accessible.

On the security front, Taiwanese research firm TeamT5 reported that Chinese state-tied hacking groups are conducting over twice as many attacks since integrating DeepSeek into their operations. The groups favor DeepSeek for its low cost and loose safety guardrails, using it to write break-in code, conduct email raids, and map targets. Researchers have not yet detected more expensive models like Kimi K3 in attacks, suggesting cost and guardrail accessibility are primary factors. The AI Security Institute previously warned that AI's cyber capabilities are doubling every few months.

Additional content includes a guide for building reusable AI design systems, a community workflow showcasing genealogical research using AI as an investigation partner, and announcements of new tools including Alibaba's Wan 3.0 video model, Google's Antigravity agentic development tool, and Liquid AI's Pipette benchmarking suite. Notable deals and developments include Perplexity's potential $30B funding round with Nvidia, Porsche's $1.46B partnership with TCS for AI integration, and regulatory actions against an Nvidia manager and Super Micro staff for smuggling AI servers into China.