curated · 2026-06-13

The only reason to click is that $965 billion number. If it holds, this would be the most expensive public debut in history, topping even the OpenAI valuation rumors from earlier this year. But I read the full article body — there's nothing beyond the headline. No revenue, no losses, no underwriter names, no roadshow timeline. A confidential filing is just the first step; final pricing is still several rounds of negotiation away. This number is more a marker of how feverish AI funding has gotten than a real price tag. I'd discount this until the S-1 drops. For now, treat it as a sentiment indicator, not a valuation anchor.

The timeline here is what makes this worth reading: Google pledged in 2018 not to use AI for weapons, quietly removed that restriction in February 2025, and signed a classified military AI deal with the Pentagon in April. Mayrhofer says none of this went through internal discussion — leadership decided unilaterally. He also points out the company dropped its carbon-neutral goals, citing AI energy consumption. His line is blunt: Google AI products could "be used directly against me and my family." He's upfront about having an EU tenured professorship as a safety net, so quitting wasn't a financial gamble for him the way it would be for others. DeepMind researcher Andreas Kirsch also publicly said he was "deeply ashamed" back in April. Google's spokesperson response is that they're "proud" to serve national security and that AI shouldn't be used for domestic mass surveillance or unsupervised autonomous weapons. But Mayrhofer's concern is exactly about where the "any lawful purpose" clause ends. No further internal records have surfaced yet — I'll keep an eye on this one.

The useful bit here is Zuckerberg saying the quiet part out loud. After cutting 10% of staff in May and shoving roughly 7,000 people into AI roles, the new Applied AI Engineering org ran a 50:1 IC-to-manager ratio — and now they're walking it back. He says the company will 'almost certainly make more mistakes' but ruled out another company-wide layoff this year. I'd read this as internal pressure relief: admit the pace was off, then throw a July hackathon and more team-building budget at morale. The memo doesn't mention Llama 4 or any model progress, so this is an organizational brake, not an AI capability problem.

The price gap here is wild: $200 a month gets you $8K–$14K worth of tokens at API rates, a 40–70x spread. SemiAnalysis bought every plan from both companies and ran heavy coding tasks until hitting weekly caps to get these numbers. I'd discount this on two fronts. First, the post doesn't say which model versions or token prices were used for the conversion—Claude Opus 4.8 and Sonnet 4.5 have very different API rates, so the 70x figure might not be precise. Second, this is a maxed-out stress test; most users won't hit weekly limits, so the actual subsidy per user is smaller. The core point still lands: when power users consistently max out their limits, the gap between subscription revenue and inference cost gets hard to ignore. Both Anthropic and OpenAI are effectively subsidizing heavy users with investor money right now, and that pricing structure won't hold long-term.

The sudden part is the blanket cutoff: all users, all sessions, Fable 5 just errors out. Elvis Saravia's take is don't panic — he already thought Fable 5 wasn't worth it for most tasks, with high cost and nerfed performance. His go-to combo is still Opus 4.8 for planning and GPT-5.5 for execution. The post doesn't spell out what the directive actually says or how long this lasts, so treat it as a temporary pull for now. If it's a security review, pricing or capability limits might shift when it comes back.