Meta hired $145B in capex and fired 8,000 people  May 19th 2026

Quick Hits

• CISA's own cloud keys turned up on GitHub. A contractor for the U.S. Cybersecurity and Infrastructure Security Agency pushed live AWS GovCloud credentials to a public repo. KrebsOnSecurity had the exclusive.
• The Pope's first AI encyclical has a co-author from Anthropic. Christopher Olah will appear with Pope Leo XIV at the Vatican on May 25 to launch Magnifica Humanitas, the first papal text on AI labor since Rerum Novarum in 1891. AP News
• Meta started firing ~8,000 people the same week it committed $145B to AI infrastructure. Zuckerberg's internal memo frames it as the company recalibrating to "AI reality." CNBC
• OpenAI shipped GPT-5.2-Codex at 55.6% on SWE-Bench Pro. Long-horizon context compaction, native cybersecurity tooling, and the coding model is closing on its own benchmark. OpenAI
• NextEra is buying Dominion in a $66.8B all-stock deal to become the world's largest regulated utility. The pitch is the megawatts AI data centers need. CNBC
• A jury dismissed every Musk claim against OpenAI on a statute-of-limitations technicality. The $134B disgorgement and the play to remove Altman are dead. CNBC
• Standard Chartered will cut more than 7,000 jobs, saying out loud what most won't. The official release describes them as "lower-value human capital" being replaced by "AI investment." Reuters
• NVlabs released SANA-WM on arXiv: a 2.6B open-source world model that produces a minute of 720p with full 6-DoF camera control on a single GPU. arXiv

The capex math finally hit payroll

What changed this week isn't the size of any single number. It's that the capex and the layoff numbers landed on the same balance sheets, in the same days.

Meta committed $145B to AI infrastructure on Tuesday. By Thursday it had begun cutting ~8,000 jobs, with the internal memo framing the firings as a recalibration to "AI reality." Standard Chartered did the same arithmetic on Wednesday with louder words: "lower-value human capital" replaced by "AI investment." Anthropic closed in on a $30B raise at a $900B valuation, which is bigger than OpenAI's last public mark. Google and Blackstone announced a $25B AI cloud joint venture running on custom TPUs. NextEra bought Dominion for $66.8B, openly because the data centers need the megawatts.

The Vatican picked the same week to weigh in. Pope Leo XIV's first AI encyclical, Magnifica Humanitas, launches May 25 with Anthropic's Christopher Olah as co-author. It's the first papal text on AI labor since Rerum Novarum in 1891.

For nine years this newsletter has treated AI as a technology story. The last five days suggest it's no longer just that. The numbers Wall Street rewards in AI capex are the same numbers showing up in HR termination batches, and the Vatican has weighed in too.

Key Takeaways

• The capex-vs-payroll gap stopped being abstract. Meta and Standard Chartered both put numbers on it this week, in the same earnings cycle.
• Wall Street is now pricing AI capex and headcount cuts as one trade. Spend $145B on data centers, fire 8,000 engineers, watch the stock react in unison. The narrative flipped from "AI growth" to "AI replacement."
• Pope Leo XIV picked Anthropic, not OpenAI, as his AI co-author. The first papal labor encyclical since 1891 will carry a Chris Olah byline. The reputational stakes for frontier labs just moved.
• Frontier AI security incidents span four vectors this week. A U.S. cyber agency leak, a 314-package npm worm (Mini Shai-Hulud), a Claude Code RCE via deeplinks, and Anthropic Mythos exploiting 21 of 41 real V8 vulnerabilities. The "AI as attacker, AI as target" thesis now has weekly receipts.

Worth Reading

• arXiv: NVlabs SANA-WM — Open-source 2.6B world model that produces a minute of 720p video with full 6-DoF camera control, running on a single GPU. The first credible open-source counterpunch to the closed video-model labs.
• Cloudflare: Project Glasswing technical post — Cloudflare's engineering write-up of what Anthropic's Mythos model revealed about frontier cyber capabilities. Primary source for what has mostly been press coverage.
• Bank of England, FCA, HM Treasury: joint statement on frontier AI cyber resilience — The first time UK financial regulators have issued a board-level AI cyber-risk mandate. Required reading if you sit on a financial-services CISO desk.
• Fortune: Mustafa Suleyman says all white-collar work automated in 18 months — The Microsoft AI chief said it on the record. Read the reasoning, not the number.
• Tom's Hardware: China bypasses US GPU bans with 1.54-exaflops LineShine supercomputer — CPU-only monster packing 2.4 million Huawei-designed Armv9 cores. The sovereign-compute story that doesn't fit a Quick Hit but should be read by anyone modeling Nvidia's TAM.

This week's poll

Meta committed $145B to AI infrastructure and laid off 8,000 people the same week. Where does that math go from here?

Thanks for reading AI Weekly. Forward this to one person whose company just did the same math.

Browse all newsletters · Archive