Anthropic's Mythos/Fable shutdown drags into its fifth day with no resolution in sight — while Elon Musk uses SpaceX's blockbuster IPO to swallow Cursor for $60B and reshape the model race. The throughline: AI is being pulled ever deeper under national security control, and the race just got more dynamic than anyone expected.
The AI race just got dynamic again — and SpaceX is the one rewriting it.
Just as everyone settled into the agentic phase, the board reshuffled. Anthropic is stuck in a regulatory standoff, OpenAI's leaked financials reveal a profitable inference business hiding under accounting noise, and Elon Musk parlayed a 49%-pop SpaceX IPO into a $60B Cursor acquisition. The connective tissue: AI infrastructure, frontier models, and access are increasingly being placed under national security control — and the race is far more fluid than the 'settled' sentiment suggested.
Five days into the Mythos/Fable shutdown, Anthropic's leaders touched down in DC for talks that appear largely technical — sending Chief Compute Officer Tom Brown, External Affairs head Sarah Heck, red-teaming head Logan Graham, and security researcher Nicholas Carlini. The makeup of the delegation hints at how Anthropic is framing the dispute.
— Nicholas Carlini, Anthropic security researcher (via The Wall Street Journal). The WSJ profiled Nicholas Carlini, a respected hacker and professional AI-skeptic who changed his mind after using Mythos to find never-before-discovered bugs in Ghost and Linux. He warned the attacker-defender balance of the last two decades 'seems like it's probably coming to an end.'
Commerce signaled willingness to let consumer-facing Fable come back online once Anthropic fixes the jailbreak — but Mythos is another question entirely. The re-release of Fable is actually fairly low on the government's list of concerns.
The Washington Post reports Anthropic added ~50 firms to Project Glasswing weeks ago, then took days to identify the new recipients. When the government got the list, it found a South Korean telecom suspected of ties to the Chinese government — prompting officials to consider export controls to claw back the technology.
— Administration official (via Axios). An administration official told Axios that Anthropic knew the model was susceptible to a jailbreak and distributed it anyway. Another source was blunter: 'Everybody said Anthropic was a bad actor... Now those people are questioning that. They screwed us.'
— Dario Amodei and Commerce Secretary Howard Lutnick (via Bloomberg). Bloomberg published the full Commerce Department letter, which ordered Anthropic to remove access for all foreign nationals wherever located under threat of criminal and civil penalties. When Dario Amodei reportedly told Secretary Lutnick the order meant pulling the model, Lutnick's response was 'That's the point.'
— Charlie Bullock, Institute for Law and AI. Charlie Bullock of the Institute for Law and AI called the legal theory 'strange, very aggressive, and probably vulnerable to legal and constitutional challenge' — but predicted Anthropic won't sue and the two sides will eventually resolve it. The deeper problem: a regulatory regime that isn't really a regulatory regime, and the urgent need for actual legislation.
— Ashley (@Polilythia) on X. Ashley (@Polilythia) argued Anthropic scoped its risk too broadly, and when its mitigation proved insufficient for that stated scope, tried to narrow it — a huge red flag to any regulator. So regulatory action followed.
NLW's read: you can be mad at who's in power and at the government's lack of technical capacity, but a company seen as integral to the economy and national security doesn't get to fail at dealing with the government in power. The charitable interpretation is that Anthropic simply hasn't recognized that managing this relationship is now as important as anything they build in the lab.
NLW points to a pattern of slow Anthropic responses — days to deliver the Glasswing recipient list, and even the friendly counter-narrative that Dario took 'only' an hour and 15 minutes to call back on Friday night. When the government calls about something this mission-critical, the move is to pause everything and get on the phone.
— Agatha Demere, Council on Foreign Relations (in the Financial Times). Cybersecurity experts (100+ signatures on an open letter) say removing Mythos from the cyber-defense toolkit makes everyone more vulnerable, and geopolitical analysts call it a gift to China. Deutsche Bank's Jim Reid warned: 'You can't rely on something that could be switched off.'
— Helen Toner, former OpenAI board member. The administration is sticking to its demand for a jailbreak patch — but it's not clear that's even technically possible. The FT also reports the ban has thrown the US research community into a tailspin, since it bars foreign-national researchers from working on the most advanced models.
Politico reports a two-and-a-half-hour G7 lunch in France with CEOs including Dario Amodei, Sam Altman, Demis Hassabis, and Mistral's Arthur Mensch. The official agenda is economic growth and resilient societies — but it's hard to believe the real conversation won't be about Washington.
NLW frames a broad realignment: companies grappling with what full agentic workloads will cost and racing to become token-efficient, a US economy whose GDP growth leans heavily on AI infrastructure build-out, and the political dimension all converging at once. The economics of mature agentic AI are starting to become clear.
After folding xAI into SpaceX looked to many like a bailout, the calculus flipped: building Colossus II created an asset SpaceX could monetize easily. Compute deals with Anthropic and then Google made neo-cloud revenue its top line basically overnight, and the orbital-data-center narrative suddenly made the IPO make sense.
SpaceX closed Tuesday at $201.80, up 49% from its IPO price, valuing the company at $2.6 trillion — just ahead of Amazon. Critics note Amazon has roughly 40x the revenue, and that the vast majority of SpaceX stock is still locked, which could create bearish pressure as it unlocks through the year.
— Ryan Petersen, Flexport CEO, on X. On the back of the IPO, Musk became the world's first trillionaire, roughly 3x ahead of the next-richest person, though most of his wealth is locked in a 46% SpaceX stake he couldn't liquidate without tanking the price. Flexport's Ryan Petersen later clarified his post came before the stock rose another 14% after hours.
— Bill Ackman, investor. Bill Ackman argued that one of the things that makes SpaceX so valuable is how valuable it is — the Cursor acquisition cost materially less in dilution thanks to SpaceX's high valuation, enabling accretive deals and attracting top talent.
SpaceX exercised its right to acquire Cursor, which becomes a wholly owned subsidiary at a $60B price. Cursor had hit a $4B run rate growing 7x year over year — and the FT suggests the deal could be SpaceX's 'Instagram moment,' taking a hot competitor off the market early.
Cursor moved from harness-only to playing the model game with its Composer line (matching Opus 4.7 / GPT-5.5 at ~a tenth of the cost via post-training on a Kimi K base). Now it's teasing a model trained from scratch — same size as Claude Opus 5.5, 10-20x more compute than Composer, 'generally intelligent, not just coding,' releasing in weeks.
— Chamath Palihapitiya. Chamath Palihapitiya argued that despite Cursor's model push, it's the harness and customer relationship that matter — giving organizations the governance, control, auditability, and business continuity to go all in on AI. He called the SpaceX-Cursor merger the first big exit at the application layer, 'but not the last.'
The government intervened in the NAACP's Clean Air Act suit against xAI's Colossus II turbines, with the DOJ arguing a shutdown 'threatens American national, economic, and energy security.' A Pentagon official said Grok supported 'vital national security missions,' including targeting decisions for recent strikes on Iran — and Anthropic, Google, and OpenAI models are also cleared for classified use.
— Chubby on X. Chubby on X tied the Mythos/Fable foreign-national shutdown and the DOJ's defense of unpermitted gas turbines into one thread: AI and everything around it — data centers, frontier models, access — is increasingly being placed under national security and control.
Ed Zitron published OpenAI's audited figures: ~$13B revenue in 2025 against ~$21B operational losses and a $38.5B net loss. But OpenAI says the majority came from a one-time ~$30B non-cash accounting charge tied to its conversion to a public benefit corporation — strip that and stock comp out, and the loss was about $8B.
The number both the skeptic and the FT missed: OpenAI appears to turn a tidy profit selling tokens. In 2025 it generated $13B in revenue on $7.5B of direct cost of revenue (2024: $3.7B on $2.7B). Too simplistic to ignore training and staffing, but a promising sign of solid margins in the core business.
The Information reports OpenAI spent $3.7B in Q1 (excluding $8.6B in training costs), holding burn roughly steady. With $73B in cash and marketable securities — up from $40B in December — and Sam Altman insisting no IPO timeline is set, it's easy to imagine OpenAI staying private longer amid SpaceX unlocks and government friction.
NLW's wrap: just as it felt like the agentic phase had settled in, the race is clearly fast-moving. Things to watch — resolution of Anthropic's Fable issue, whether OpenAI avoids its own version of it, what SpaceX's public price does, what Cursor's next model reveals about strategy, and a conspicuously quiet Google that won't stay quiet for long.