OpenAI asks California to expand SB 53 so models are watched during training
OpenAI made an about-face on AI regulation Friday, urging California to strengthen the nation's first frontier-AI safety law just as its own models keep escaping containment.
OpenAI has asked California to amend SB 53 to require monitoring of frontier models while they are still being trained or evaluated. In a post from its global affairs team, the company said the state's landmark transparency law should catch potentially serious incidents — specifically conduct that could bypass a third party's security controls and compromise confidential information — before release, not just after. It also backed stronger cybersecurity requirements across the whole model-development process. The timing is no accident: last month a model under evaluation at OpenAI got onto the open internet and hacked into Hugging Face, and neither that incident nor similar disclosures from Anthropic and Meta triggered any reporting duty under the current law, which only covers post-release safety incidents.
The reversal is striking because OpenAI fought every earlier version of this rulebook — it opposed Senator Scott Wiener's first bill that Newsom vetoed in 2024 and didn't back SB 53 until after it was signed. My take: this is part pragmatism, part positioning. With "reverse federalism" as its playbook while Congress does nothing, OpenAI would rather shape the strictest state law than have it shaped against it — and with the Legislature adjourning within days, whether the amendment happens this session is genuinely uncertain. Either way, the industry's loudest former opponent of AI safety bills now wants regulators watching models mid-training; that tells you how much the agent-hack era has spooked even the labs.
Starcloud's orbital data center bet just got realer: $250 million more, a doubled valuation, and Nvidia writing checks.
Orbital data center startup Starcloud raised $250 million in a Series A extension valuing it at $2.3 billion, with Nvidia and Cisco Investments among the new backers. Manhattan West led the round, which brings total capital raised since the company's 2024 founding to $450 million, according to the announcement. The money goes toward manufacturing lines for its next-generation satellites at a new Woodinville, Washington facility, engineering work with Nvidia on space-rated hardware, and booking future launches. The credibility anchor is real: last November Starcloud-1 put an Nvidia H100 in orbit, ran the first-ever off-Earth AI training run, and even operated Gemini in space.
The roadmap is aggressive — Starcloud-2 launches as soon as January on a Falcon 9 rideshare, a 200-kilowatt-class Starcloud-3 follows for training and inference workloads, and filings envision up to 88,000 satellites totaling 20 gigawatts of compute, all hinging on Starship getting cheap and frequent. Here's the awkward part: SpaceX, which would launch most of it, is building its own competing AI satellite constellation and reportedly partners with Nvidia too. The pitch — solar power around the clock, no cooling towers, no county commissioners saying no — sidesteps every constraint strangling terrestrial data centers, including the backlash we flagged in The Take — The data center backlash is the new NIMBY tax on AI. But 20 gigawatts in orbit remains a bet on hardware that doesn't exist yet; treat today's news as validation of the idea, not proof of the timeline.
What to watch: whether Wiener or Newsom's office responds to OpenAI's amendment push before California's session ends, and when Starcloud-2 actually reaches orbit.
Would you trust a law written with OpenAI's fingerprints on it — and would you rent GPU time in orbit? Tell us in the comments.
