Chris Olah's Vatican remarks are striking because he said the quiet part in the courteous register. Anthropic published the full text of his comments on Pope Leo XIV's AI encyclical, including his warning that every frontier AI lab, Anthropic included, operates inside incentives and constraints that can conflict with doing the right thing. [1] He named commercial viability, frontier pressure, geopolitics, pride and ambition. [1] It was a moral confession from inside the machine.

The paper's Tuesday account of Olah saying frontier labs' incentives conflict with doing right treated that confession as the story. Wednesday's follow-up has to put the confession beside the bill. Perplexity AI Magazine reported that a SpaceX IPO prospectus disclosed an Anthropic compute arrangement worth $1.25 billion a month through May 2029, about $45 billion over the term. [2] The source is not a traditional financial outlet, so the claim should be read as a trade-media report pending primary filing review. It is still the public receipt in today's stack.

The juxtaposition is the point. Olah told the Vatican audience that outside critics are needed because the people building the systems live inside incentives. [1] A reported $45 billion compute obligation is not merely an incentive. It is gravity. Once a lab carries infrastructure costs at that scale, every safety discussion competes with utilization, revenue, customers, model cadence and strategic dependency.

Anthropic's own text is unusually rich. Olah said AI systems are not engineered like bridges or airplanes; they are grown on structures roughly modeled after the brain and remain mysterious even to those who train them. [1] He also said the questions raised by AI are too large for computer scientists alone and belong to the humanities, religion, philosophy and society. [1] That is not boilerplate corporate responsibility language. It is an argument for external moral jurisdiction.

The compute report gives the jurisdiction a harder reason to exist. Perplexity AI Magazine described the alleged arrangement as access through SpaceX capacity and linked it to the scarcity of frontier-scale compute. [2] It also framed the disclosure as central to understanding AI infrastructure costs. [2] Even if future filings refine the exact terms, the broad logic is already visible across the industry: frontier labs cannot preach restraint while buying the right to burn astonishing quantities of electricity, chips and capital without explaining the governance that follows.

X's reaction to this pairing is predictable and not entirely wrong. Some read it as hypocrisy: a safety-branded lab speaking piously while attaching itself to a giant compute bill. Others read it as realism: moral AI will still require industrial infrastructure, and pretending otherwise is childish. The better reading is less satisfying. The moral voice and the compute bill are not contradictions. They are the same problem seen from two rooms.

Olah's three questions for discernment underline that. He asked about the global poor, human flourishing and the nature of AI models, including internal states that may mirror joy, fear, grief and unease. [1] Those are not questions a procurement department can answer. But procurement can determine which questions get time. A lab under immense compute pressure may still ask noble questions. It may also answer them on the schedule imposed by its suppliers.

The predecessor on SpaceX's reported S-1 compute disclosure gave the paper the financial side of the ledger. Wednesday's article binds it to the moral side. Frontier AI's public language has matured. The money behind it has grown even faster. The danger is not that moral language is fake. The danger is that it becomes a decorative surface over obligations so large that only growth can service them.

That is why Olah's call for outside critics should be taken literally. If the Church, civil society, governments and scholars are needed, they are needed not only to read encyclicals. They are needed to read contracts, filings, capacity deals and power arrangements. The soul of the model may remain mysterious. The invoice will not.

The next receipt should be primary. A prospectus line, a supplier agreement or an Anthropic confirmation would either harden the reported number or correct it. Until then, the moral fact still stands: a lab asking society to trust its intentions must also explain the dependencies that shape its choices. Ethics without a balance sheet is theater. A balance sheet without ethics is the industry default.

-- THEO KAPLAN, San Francisco