A government letter prompted Anthropic to disable its most advanced AI models within days, with no congressional vote, court order, or public deliberation. Separately, the SpaceX IPO transferred no meaningful control, as Elon Musk retained his 42% stake and supervoting shares despite the listing.
The U.S. government took a $2 billion equity stake in quantum startups, including one with political ties, raising questions about how early-stage winners are selected. AI agents are now trading cryptocurrency autonomously through Coinbase, while the gap between frontier AI capability and what reaches consumers continues to widen.
The thing that matters most from this week got the least attention. While everyone tracked the SpaceX IPO and counted Musk’s billions, the US government sent a letter that effectively disabled Anthropic’s most advanced models, and Anthropic complied. A government letter, compliance within days, and the most capable AI system in public deployment went dark. No congressional vote, no court order, no public deliberation about what got weighed against what. Just a letter.
That is the actual story of the week. Everything else is context for it.
SpaceX goes public, Musk keeps the keys
The SpaceX IPO was the largest in history, and the ownership structure barely changed. Musk’s 42% stake and supervoting shares mean that public investors bought exposure, not influence. The company runs the American launch infrastructure and is the only privately operated satellite internet network with global coverage, and none of that changed when it listed. The governance conversation that should have accompanied the offering did not happen. Washington got a valuation number; Musk got a liquidity event. The keys stayed where they were.
Quantum picks and political selection
The US took a $2 billion equity stake in nine quantum startups, one of which is backed by a firm with ties to the Trump family. Picking quantum winners with political ties is not a process that invites trust, and the early stage of a technology category is exactly when the picks matter most. Alice and Bob’s fault-tolerant quantum hardware announcement this week, the first commercially deployable system targeting on-premises deployment, landed at a moment when the government is already deciding who gets to lead this race. The companies that get backing early will have a structural advantage that is nearly impossible to overcome later, as I wrote about in my piece on Illinois’s $20 billion quantum park. The infrastructure race is not just about who builds the hardware. It is about who the government decides deserves to build it.
AI agents trade crypto autonomously
Coinbase launched Coinbase for Agents, a standalone product that lets AI assistants from Claude and ChatGPT trade cryptocurrency and move money without human sign-off per transaction. OpenAI acquired Ona, a startup focused on managing long-running agents, the same week. Andrew Ng has been pushing for “unbiggen” AI, meaning better orchestration of existing models rather than more powerful ones. The shift from chatbots to agents I wrote about after CES 2025 is now operational, and it is starting where the incentive is highest: money that moves continuously, at machine speed, without waiting for someone to press approve.
FAANG is dead. The new power list is MANGOS: Meta, Anthropic, Nvidia, Google, OpenAI, SpaceX. Half are going public this year. NVIDIA’s RTX Spark targets creators rather than gamers; Qualcomm’s Snapdragon C promises capable Windows laptops at half the price of a MacBook. Carl Pei made the most clarifying comment of the week: RAM is now the single most expensive component in a phone. Not the display, not the camera sensor. The memory. If the component economics of building at the consumer tier are getting worse while the governance of building at the frontier tier is tightening, the gap between what capable AI looks like and what reaches a normal person is not closing on its own.
Which brings it back to the Anthropic shutdown. It is not a cautionary tale about capability. It is a data point about what oversight of the most powerful AI systems actually looks like in practice: informal, fast-moving, and operating without the kind of public process that would let anyone outside the room understand the tradeoff. A letter. Compliance. Silence. Whether that is appropriate depends entirely on how much you trust the people sending the letters. Right now, that trust is doing a lot of work it was not designed to carry.