Good morning,

This week we cover the public release (then ban by US Gov) of Anthropic Mythos model as Fable. We unpack that, then return to a story we covered previously and which has now actually happened: SpaceX's record IPO. Plus the usual bites: Nvidia's move into laptops, hedge funds shorting the call-centre industry, Whatsapp for Business new AI features and more.

On 4 June, Anthropic's research institute published When AI Builds Itself, claiming: AI is now meaningfully accelerating the building of AI. More than 80% of the code Anthropic merges is written by its own model; the typical engineer ships 8x as much code per day as in 2024; and on one internal test, the model went from a roughly 3x speed-up over human-written code a year ago to 52x today. Extrapolate, and you reach recursive self-improvement: a system capable of designing its own successor. The authors conclude by recommending it would be good for the world to have the option to "slow or temporarily pause frontier AI development," warning that the gap between AI's pace and society's ability to adapt is widening.

Then the irony. Two days before it, on 2 June, Anthropic widened access to its most powerful and most restricted model, Mythos, from roughly 50 to about 150 organisations across 15 countries, including NATO, the EU's cyber agency, Samsung, and, reportedly, the NSA (to help it spies). It is an uncomfortable pairing: a lab that wants the option to halt AI handing its most dangerous model to the agency Edward Snowden exposed for mass surveillance. Then, five days after the essay, on 9 June, it released Fable 5 to everyone, a contained sibling of Mythos, made safe for general use by routing high-risk requests away from the model, but still, on most benchmarks, the most capable thing on the market. Arguing for the brakes and flooring the accelerator in the same weeks is quite the straddle.

However the party was short. On 12 June, the US Commerce Department ordered Anthropic to disable both Fable 5 and Mythos for any foreign national, anywhere — including its own foreign-born staff. Unable to filter in real time, Anthropic switched the models off for everyone. As far as we can tell, it is the first time a government has forced a publicly released frontier model offline. Anthropic is complying while disputing the decision, calling it a "misunderstanding".

What we glimpsed, and what it cost. While it was live, Fable was, by most accounts, remarkable: Mythos-class reasoning, multi-hour autonomous work, and a cyber capability that has already surfaced more than 10,000 critical software vulnerabilities. However, it was also extraordinarily expensive, listed at $10 / $50 per million input/output tokens, roughly double Opus (which was already judged expensive). And here is the strategically interesting part: Anthropic had said Fable 5 will only be included in its flat-rate subscriptions only until 22 June, after which it moves to metered usage credits. To illustrate this change, I used Fable 5 to build a simple game from 1995. The work took 3h and maxed out twice my 5h compute limit with Claude Max subscription. Had I been paying per token, this work would have costed me $150. Some users even reported running work at $10 per minute.

This is another highlight that price is becoming everyone's problem. First, in early June GitHub Copilot abandoned flat-fee billing for token-metered "AI credits" — a change developers have nicknamed the "tokenpocalypse", reporting cost jumping from $20 per user to thousands of dollars. Then, SemiAnalysis ran the numbers on the labs' own subscriptions and found them heavily subsidised: a $200 Claude Max plan can deliver roughly $8,000 of compute at API rates; the equivalent ChatGPT tier, up to $14,000. The top tiers turn loss-making once a user's utilisation passes single digits. The maths does not hold. Either prices rise or availability falls.

Finally, on 12 June, SpaceX completed the largest IPO in history: a $75bn raise at a ~$1.75tn valuation, popping nearly 30% on debut to clear $2tn (a whopping 2.5% of the $70tn of the US stock market cap), making Elon Musk the world's first trillionaire. Yet, as Richard Waters argue, the listing says less about SpaceX than about Musk's command of narrative — data centres in space, xAI as the AI backbone of business — "transmuted into Wall Street gold," with a financial establishment (some $500mn in fees) happy to oblige.

More interesting for the rest of us is the new direction xAI seems to take to make money: renting its Colossus Data Center. Colossus was built last year at Musk demand to provide a compute facility to train Grok (xAI model). However it seems they do not need it so far, but instead of sitting idle, they transformed the data center into a gold mine. Anthropic is paying $1.25bn a month for Colossus 1 through 2029; and on 5 June, Google agreed to pay $920m a month from October for roughly 110,000 chips. Together those two deals give SpaceX about $2.17bn a month — roughly $26bn a year — in compute rent. Let’s note that: Google, plausibly the largest single owner of AI compute on earth and itself spending more than $180bn this year, is renting from a rival because it cannot build fast enough.

The macro-story behind this IPO and the ones to come. In a sharp Big Read, the FT's Robert Armstrong argues we are watching "regime change not just in technology but in finance." For two decades, Big Tech were cash machines (Alphabet, Microsoft, Meta and Amazon threw off over $200bn of free cash flow in 2024) and the market's biggest buyers of their own shares (around $800bn of buybacks in a decade). AI capex has flipped that: free cash flow is falling, and to keep buying back stock these firms will increasingly have to borrow. The age of "de-equitisation" is becoming one of re-equitisation. Alphabet is raising $85bn in new shares, Meta may follow, and with OpenAI and Anthropic also filing, five companies could issue some $400bn this year. Contrary to some belief, buying capacity isn't the worry ($8tn sits in US money-market funds). The cycle is. On Armstrong's reckoning the AI and data-centre boom is now around 7% of GDP — about the size of the 2007 housing boom at its peak, and bigger than the late-1990s dotcom build-out. History is unkind here: large IPO waves tend to be followed by more muted, more volatile, and more often negative returns.