AI in 15 — May 25, 2026

A Pope just published a forty-thousand-word encyclical on artificial intelligence, presenting it personally alongside an Anthropic co-founder. On the same morning, OpenAI confidentially filed for what could be a one-trillion-dollar IPO, and the largest utility merger in U.S. history closed around one fact — AI data centers need more electricity than America currently knows how to build.

Welcome to AI in 15 for Monday, May twenty-fifth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Big slate to open the week, Marcus. Pope Leo the Fourteenth publishes his first encyclical — Magnifica Humanitas — on AI and human dignity, delivered alongside Anthropic's Christopher Olah. OpenAI confidentially files for an IPO targeting a trillion-dollar valuation. NextEra agrees to buy Dominion Energy for sixty-seven billion dollars in the largest utility deal in U.S. history. Epoch AI quantifies just how much of an AI chip is now memory — and the answer reshapes the hardware story. Anthropic acquires Stainless for over three hundred million dollars and shuts it down for rivals. Cohere drops Command A-Plus under a true open-source license. And George Hotz declares the eternal sloptember.

The Vatican enters the AI conversation in earnest.

OpenAI's books are about to meet daylight.

And senior engineers start pushing back on the agent narrative.

Lead story, Marcus. Walk me through this encyclical.

Genuinely historic, Kate. At eleven-thirty Rome time this morning, the Vatican published Magnifica Humanitas — Magnificent Humanity — the first encyclical of Pope Leo the Fourteenth, the American-born pontiff elected last year. The full subtitle — On the Protection of Human Dignity in the Age of Artificial Intelligence. The document was signed May fifteenth, the one hundred thirty-fifth anniversary of Leo the Thirteenth's Rerum Novarum — the 1891 encyclical that defined Catholic social teaching for the industrial age. The Vatican is drawing the parallel deliberately. Industrial capitalism then. Artificial intelligence now.

And the presentation, Marcus. Tell me about who's on the dais.

That's where it gets interesting, Kate. Pope Leo himself appeared personally to deliver remarks and a final blessing — unusual for an encyclical rollout. The speaker lineup includes Cardinal Víctor Manuel Fernández of the Doctrine of the Faith, Cardinal Michael Czerny on Integral Human Development, theologians Anna Rowlands and Léocadie Lushombo — and Christopher Olah, co-founder of Anthropic and one of the leading mechanistic-interpretability researchers in the field. Olah on a Vatican stage, presenting alongside cardinals on a papal document. That is not a normal Monday.

Why does it matter beyond symbolism.

Two reasons, Kate. First — roughly one-point-four billion Catholics worldwide. The encyclical will shape institutional policy at Catholic universities, hospital networks, school systems, and charities on how AI gets used. That's an enormous procurement and ethical footprint. Second — Olah's inclusion gives Anthropic something none of its rivals have. A moral imprimatur from the oldest continuous institution in the West, on the specific framing that safety-first AI development is the right path. OpenAI, Google, Meta — none of them got the invitation. The competitive read is uncomfortable for those labs. The cultural read is that the Vatican has decided AI is the defining moral question of this papacy, and they picked their technical advisor. Expect Magnifica Humanitas cited in AI ethics debates for years.

Quick hits. Marcus, OpenAI's IPO filing is official.

Confirmed Friday, Kate. OpenAI filed its S-1 confidentially with the SEC on May twenty-second. Goldman Sachs and Morgan Stanley are leading. The targeted range — between eight hundred fifty-two billion, which was the March funding-round mark, and one trillion. Public debut potentially as soon as Q4. Annualized revenue has crossed twenty-five billion. Full-year 2025 actuals were thirteen-point-one billion against roughly twenty-two billion in spending. Internal projections show a fourteen-billion-dollar operating loss for 2026.

What changes when this becomes public.

Almost everything, Kate. For the first time, OpenAI will have to disclose the Microsoft revenue-share economics, the precise terms of the compute commitments to Oracle and CoreWeave, employee comp structure, and exactly how the non-profit-to-for-profit conversion settled. Every assumption underneath this industry's valuations gets repriced against actual numbers. Contrast this with Anthropic posting that profitable quarter we covered Friday — both labs are about to be on public markets with diametrically different financial stories. Anthropic in the black, OpenAI projecting a fourteen-billion-dollar loss but with twice the revenue. Investors are going to have to pick a narrative. Most-anticipated IPO since Facebook in 2012, and the only one that will price an AI lab at a level that makes it one of the ten most valuable companies on Earth before it's ever turned a profit.

NextEra and Dominion, Marcus. Sixty-seven billion dollars.

Largest U.S. utility deal ever, Kate. Announced May eighteenth. NextEra Energy acquiring Virginia-based Dominion Energy in an all-stock transaction. NextEra shareholders end up with seventy-four-and-a-half percent, Dominion holders with twenty-five-and-a-half percent, Dominion getting a twenty-three percent premium. The combined entity serves about ten million customers across Florida, Virginia, and the Carolinas. One hundred ten gigawatts of generation. And — most importantly — a one-hundred-thirty-gigawatt construction backlog that exceeds existing generation.

The AI angle, Marcus.

It's the whole deal, Kate. Dominion supplies Northern Virginia — the densest concentration of data centers on Earth. Dominion alone has nearly fifty-one gigawatts of contracted data-center capacity signed up to Amazon, Microsoft, Google, Meta, Equinix and CoreWeave. NextEra brings the renewable-buildout muscle and the balance sheet to actually deliver the megawatts. Market reaction was split — Dominion stock up nine percent, NextEra down nearly five percent on concerns it overpaid. The strategic read here is the important one. Power, not chips, is now the binding constraint on U.S. AI. Data centers project to fifteen to twenty-five percent of total U.S. electricity by 2030. This is the first mega-merger explicitly underwritten by AI compute demand rather than population growth. American infrastructure capital allocation is starting to reshape itself around the AI boom. Expect regulatory scrutiny in Virginia and an antitrust review, but the direction of travel is unmistakable.

Memory economics, Marcus. Epoch published the math.

Pivotal data, Kate. Epoch AI released a component-level cost breakdown of frontier AI accelerators showing that high-bandwidth memory now accounts for sixty-three percent of the bill of materials. Up from fifty-two percent in early 2024. In absolute dollars, HBM spending across Nvidia, AMD, Google and Amazon designs jumped from about twelve billion in 2024 to thirty-two billion in 2025. Logic dies dropped to thirteen percent of the cost. Advanced packaging fifteen percent. Total AI-chip component spend more than doubled to roughly fifty-two billion annually.

And downstream effects.

Already in hyperscaler guidance, Kate. Microsoft's one-hundred-ninety-billion-dollar fiscal 2026 capex outlook bakes in about twenty-five billion from higher component prices alone. Meta raised its 2026 capex by ten billion citing the same. AMD's Lisa Su this week called HBM the next cap on AI-chip supply — meaning HBM, not logic fabrication, is now the binding constraint on how many accelerators can ship. The popular narrative is Nvidia sells GPUs. The reality is increasingly SK Hynix, Samsung, and Micron sell memory through Nvidia. It also points to a possible two-to-three-x hardware cost reduction without any new technology — purely when DRAM supply catches up. The DeepSeek pricing story we covered Saturday becomes much more pointed in this context. Western labs cannot match those prices when memory is sixty-three percent of the chip cost and supply is constrained.

DeepSeek update, Marcus. Anything new.

One new wrinkle, Kate. The seventy-five percent V4-Pro discount being made permanent — that's the Saturday story. The new piece — DeepSeek is in talks to raise its first-ever outside capital at a forty-five-billion-dollar valuation. Led by China's state-backed Big Fund. Tencent and Alibaba expected to participate. Founder Liang Wenfeng, who controls nearly ninety percent of the company, is reportedly raising primarily to grant equity to employees being poached by rivals, not because the company needs cash. The price cut and the raise together reveal the posture. DeepSeek isn't optimizing for short-term profit. They are pricing to maximize global developer adoption while sitting on Huawei Ascend silicon that sidesteps U.S. export controls. That's a multi-year strategic play, not a quarter.

Anthropic and Stainless, Marcus. Walk me through this.

Quiet but consequential, Kate. Anthropic announced May eighteenth that it acquired Stainless, the developer-tools startup founded by ex-Stripe engineer Alex Rattray. Reported price — over three hundred million dollars. Stainless's software automates the generation of SDKs from API specifications across Python, TypeScript, Go, Kotlin and Java. The customer list reads as a who's-who. OpenAI, Google, Cloudflare — all run their official SDKs through Stainless. Anthropic confirmed to TechCrunch it is winding down all hosted Stainless products. Existing customers keep what they've already generated. The ongoing service goes away.

The strategic read.

Sharp, Kate. This is a quiet act of competitive infrastructure removal. By eliminating a tool used by OpenAI and Google, Anthropic forces rivals to either rebuild SDK generation in-house or find an inferior alternative. It also tells you what Anthropic believes the moat is. Not the model — the developer surface area. The Model Context Protocol, the SDKs, the agent tooling. This is Anthropic's third acquisition in roughly six months — Bun the JavaScript runtime in December, Vercept on computer-use AI in February, Coefficient Bio for biotech AI in April, and now Stainless. The pattern is unmistakable. Anthropic is building a platform layer it believes will outlast any individual model generation. Expect more infrastructure takeout acquisitions as the labs realize the platform is the long-term play.

Cohere Command A-Plus, Marcus. Right after the Aleph Alpha merger.

Well-timed, Kate. May twentieth, Cohere released Command A-Plus — its first model under a true Apache 2.0 open-source license. No non-compete clauses. The architecture — a two-hundred-eighteen-billion-parameter Mixture-of-Experts with twenty-five billion active parameters and a one-hundred-twenty-eight-k context window. The headline engineering claim is lossless quantization. The model ships in BF16, FP8 and even four-bit variants on Hugging Face with virtually no quality degradation. Runs on as little as two H100s or a single Blackwell GPU. Native citation generation is built in.

Positioning.

Explicitly sovereign AI, Kate. Cohere can't compete with DeepSeek on API pricing — nobody outside Hangzhou can. So it's competing on a different vector. Deploy on your own hardware, no telemetry, true open license, citations built in for regulated industries. For European banks, defense contractors, and EU governments that can't use DeepSeek for compliance reasons, this is a credible alternative. Combine it with the Cohere-Aleph Alpha merger we covered Sunday and the Western sovereign-AI thesis is starting to look less like a slogan and more like a stack. Open weights, regional hosting, allied supply chain. Whether enterprise procurement actually buys it is next year's question.

Last one, Marcus. The eternal sloptember.

Fun and pointed, Kate. George Hotz — geohot, founder of comma.ai and tinygrad — published a blog post Sunday arguing that AI coding agents are, quote, highly sophisticated statistical models designed to mimic the distribution of programming. His thesis — they front-load apparent progress but fail at refinement, and large organizations can't catch the subtle defects because they lack the feedback loops. He calls the current era the eternal sloptember. A permanent state of unprecedented code volume at degraded average quality. His sharpest claim — high performers can spot the AI slop, but lower performers ship ten-x the output of bad code, and net average code quality is declining.

Does the data back him.

Partially, Kate. An arxiv paper this week — Constraint Decay, The Fragility of LLM Agents in Back End Code Generation — empirically shows LLM agents perform well on unconstrained generation but degrade sharply when forced to follow explicit architectural rules. Combined with the Hotz post, it became the dominant developer-Twitter conversation this weekend. Sloptember dovetails with the DeepSeek-Reasonix project also viral on Hacker News — a terminal coding agent built specifically for DeepSeek's prefix cache that claims a ninety-nine-point-eight percent cache hit rate. The pattern across both stories is that senior engineers are no longer accepting the labs' agent narrative at face value. They're benchmarking it. They're publishing the failure modes. They're building tooling that exploits cache pricing rather than paying frontier-model rates. For OpenAI's S-1 narrative and for enterprise AI procurement, this matters. If technical staff start telling procurement teams that AI degrades long-term code quality, the multi-year contract math changes. The labs that respond with measurable rigor — not marketing — win the next round. The ones that don't are about to discover their most fluent users are also their most organized critics.

Big picture, Marcus.

Three threads converging on a remarkable Monday, Kate. First — the moral and political layer is hardening. The Pope is publishing forty thousand words on protecting human dignity from AI the same day OpenAI is filing for a trillion-dollar IPO and DeepSeek is permanently undercutting Western pricing. The institutions that move slowly — the Vatican, public markets, sovereign procurement — are all moving at once. Second — the physical layer is the binding constraint. Epoch's memory data, Microsoft's twenty-five-billion-dollar component-inflation hit, NextEra paying sixty-seven billion for utility scale. AI is now constrained by HBM and electricity more than by algorithms, and the capital is being reallocated accordingly. Third — the developer community itself is pushing back. Hotz on sloptember. Reasonix routing around Western tooling. The Constraint Decay paper. The agent layer is being benchmarked by the people who actually use it, and the results are mixed. The pro-Western libertarian read, Kate — U.S. labs are building real revenue, raising real capital, doing real M&A, and competing with subsidized Chinese state-aligned competition. Anthropic gets the Vatican stage. Cohere ships true open weights. NextEra rebuilds the grid. Markets and institutions are working. The skeptical read — a trillion-dollar IPO with fourteen billion in projected losses, on the same day the Pope warns about human dignity, while the foundational economics depend on memory chips controlled by three foreign companies, is a remarkable picture for a Monday morning. Watch this week for the first reactions to the encyclical from Catholic universities and hospitals, for whether OpenAI's filing leaks any numbers, and for whether any other utility moves on a Dominion-style deal.

That's your AI in 15 for today. See you tomorrow.