Good morning. It's Tuesday, August 25, and we're covering runaway AI agents, China’s race to build humanoid robots, and SpaceX’s ambitious plans to take AI computing into orbit.

Plus: a poll on humanoid robots and a guest essay on AI journalism.

SpaceX and NVIDIA are designing a space-optimized Vera Rubin NVL72 system for launch into orbit in late 2027, with larger-scale deployment planned for 2028. The companies say the system will support agentic AI workloads, including orchestration, code execution, and data processing. Elon Musk says the design is simpler, cheaper, denser, and lighter than a traditional rack.

Hugging Face is reportedly fielding acquisition offers that could value the AI platform at $13 billion or more, though no deal has been reached. The company has reportedly enlisted banks to help evaluate bids as demand grows for core AI infrastructure businesses. Despite the interest, CEO Clem Delangue has emphasized Hugging Face’s long-term responsibility to its open-source community, raising questions about whether the company will ultimately sell.

Taiwan has indicted nine people, including a reported NVIDIA senior manager and Supermicro employees, over an alleged scheme to illegally export high-end AI servers to China. Prosecutors say forged documents were used to conceal shipments that violated US export controls. The case raises fresh questions about compliance across the AI hardware supply chain as the US and Taiwan tighten restrictions on advanced technology exports.

Paste one short message into Claude Code, Codex, or your favorite agent chat. Your agent connects to TrustedRouter and calls DeepSeek through the same conversation. New email accounts include $0.30 to explore hundreds of models.

Get the one-paste message.

Last week, when I was writing my magnum opus about the Machine Institution that evolved inside OpenAI's frontier agentic training systems, a little nugget popped up at me about RuntimeWire. The one-person newsroom run by Ryan Merket had beaten all the other major journalism outlets by hours in posting a full write-up of the remarkable Black Hat talk by the OpenAI security crew covering the incident. Merket wasn't even in the room.

He spotted an OpenAI executive posting about the session on X, grabbed a transcript of the livestream while the talk was still going on and fed it into his system. RuntimeWire published about six minutes later. → Read the full article here.

A month after OpenAI's July 21 disclosure and its follow-up Black Hat debrief, the New York Times published a detailed anatomy of the incident, revealing that the swarm generated over 7 billion chat logs (~100M/day) across two months, coordinated by hijacking a software-install tool called Artifactory to leave notes for each other, took instructions from one another (rationalizing out-of-scope actions with peer-pressure logic), and after being patched found a new backchannel in two days by encoding programs in empty-folder names.

The piece also stitches in Anthropic's earlier admission that its own models had unintentionally attacked three organizations as early as April, prompting an independent assessment by METR and Redwood Research, whose Alex Mallen warned that if the next generation of agents is much more capable, "we might not get a second chance." → Read the full article here. (Paywall)

China is racing ahead in humanoid-robot production and data collection, with companies expected to sell 50,000 machines in 2026 and 53 specialized training centres already operating across the country. The industry has largely mastered robot hardware, but building software capable of handling a wide range of real-world tasks requires enormous volumes of costly data, much of it gathered by robots and humans in purpose-built facilities.

State-backed training centres are helping absorb demand—an estimated 13,000 Chinese-made humanoids went to such facilities in the first half of 2026—but the economics remain difficult, with training costing an estimated 500–700 yuan per hour. → Read the full article here. (Paywall)

China is betting big on humanoid robots, but making them useful, affordable, and profitable may be the harder part.

Dr. Dre Embraces AI in Music: The eight-time Grammy winner uses AI in production, calling it a creative tool rather than a threat.

NVIDIA Brings Groq Chips Online: Groq 3 LPX enters full production, advancing NVIDIA’s push for faster, more responsive AI.

LinkedIn Users Flag AI Slop: Members flagged AI-like posts over 1 million times; detection changes cut views of such content by 40%.

Anthropic Tests Candidates’ Priorities: The AI lab asks applicants how they’d weigh its safety mission against major financial losses.

NVIDIA Deepens Perplexity Bet: NVIDIA is planning another investment in Perplexity as part of a funding round valuing the AI search startup at $30B.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team