Pope Leo XIV's First AI Encyclical, OpenAI's IPO Filing, Anthropic at $900B

Top AI News Highlights — May 25, 2026

• Pope Leo XIV publishes the first-ever AI encyclical, Magnifica Humanitas, a 42,300-word document framing AI's military use and human dignity as central concerns (Vatican / Washington Post).
• OpenAI files a confidential S-1 with the SEC, reportedly targeting a valuation range of $852B to $1T with a September roadshow on the calendar (CNBC / Fortune).
• Anthropic's $30B round at a $900B+ valuation is "closing as soon as next week," according to Bloomberg, with a Q2 revenue outlook of $10.9B and a first-ever quarterly operating profit of $559M (Bloomberg).
• SoftBank announces its "AI Data Center GPU Cloud" for an October 2026 launch, powered by the in-house Infrinia AI Cloud OS, deploying NVIDIA GB200 NVL72, with a group-internal beta beginning today (SoftBank).
• Leaks pile up — Claude Mythos 1, Claude "Memory Files," GPT-5.6, and Gemini 3.5 Pro all reportedly inching toward release. None are officially confirmed; the full breakdown is at the end of the article.

Pope Leo XIV Publishes the First-Ever AI Encyclical: Magnifica Humanitas

On May 25, 2026, Pope Leo XIV released Magnifica Humanitas ("The Greatness of Humanity"), the first encyclical of his pontificate and, far more notably, the first papal encyclical in the two-thousand-year history of the Catholic Church to take artificial intelligence as its central subject (Vatican). For readers outside the Catholic tradition, the genre matters here: an encyclical is the highest form of ordinary teaching document a pope can issue. It is addressed not only to the bishops of the Church but, in modern practice, to "all people of good will." The decision to make AI the subject of such a document is itself a statement about where the Vatican now places the issue in the moral hierarchy of the contemporary world.

The document is dated May 15, 2026 — precisely 135 years after Pope Leo XIII's 1891 encyclical Rerum Novarum ("Of New Things"), the foundational text of modern Catholic social teaching, which addressed the dignity of industrial-era workers, the moral status of private property, and the relationship between labor and capital. The dating is not an accident. Leo XIV's choice of name — taken in part to honor Leo XIII — and his choice of date together draw an explicit line: just as Rerum Novarum spoke into a world being reshaped by industrial labor, Magnifica Humanitas is being positioned as the equivalent intervention for a world being reshaped by machine intelligence (Vatican News).

That historical anchoring is worth pausing on, because it tells you how the Vatican wants the document to be read. Rerum Novarum was not, in 1891, a piece of niche religious writing. It is now widely credited with shaping a century of labor law, with influencing the development of Christian democratic parties across Europe and Latin America, and with providing intellectual scaffolding for the welfare state. The Holy See is signaling, by dating Magnifica Humanitas exactly 135 years later, that it expects this text to do similar long-arc work for the AI era.

It is also worth noting that the document arrives at a moment when the AI policy debate in the West has fragmented along several different axes. The United States has pulled back from the comprehensive regulatory posture that the Biden-era executive orders established, and is now relying mainly on sector-specific rulemaking and voluntary safety frameworks. The European Union has the AI Act on the books but is still working through enforcement guidance for general-purpose models. The United Kingdom continues to pursue a "pro-innovation" stance that resists hard rules in favor of regulator-led principles. China is moving in the opposite direction, tightening both content controls and deployment licensing. Into this fragmented landscape, Magnifica Humanitas introduces something none of these regimes provides: a unified moral framework that can be cited across jurisdictions without renegotiation. That portability is precisely what gives it the potential to influence boardroom discussions in companies operating across all of these regulatory regimes simultaneously.

42,300 Words on AI's Military Use and Human Dignity

Running roughly 42,300 words, Magnifica Humanitas centers on three interlocking themes: the militarization of AI, the future of human labor in an environment of widespread automation, and the impact of advanced systems on the very definition of personhood (TIME). The National Catholic Reporter notes that the text contains language amounting to a de facto call for disarmament around lethal autonomous weapons systems (LAWS) — autonomous weapons that select and engage targets without meaningful human control (NCR).

The Washington Post argues that the encyclical's most consequential move is rhetorical: it elevates AI ethics from a "technical and industrial question" to what the Post calls a "religious imperative" (Washington Post). For the first time in recent memory, religious moral authority is being placed on the same negotiating table as corporate self-regulation, the EU AI Act, NIST's AI Risk Management Framework, and the patchwork of U.S. state-level legislation that has emerged over the past two years. That convergence — secular law and religious doctrine speaking to the same problem — is rare, and it changes what AI governance discussions look like inside multinational organizations.

On the military section specifically, several commentators have observed that Leo XIV's framing extends earlier Vatican positions on autonomous weapons, but with a different emphasis. Previous Vatican statements on LAWS tended to focus on international humanitarian law and the principle of meaningful human control. Magnifica Humanitas, according to early academic responses, reframes the issue around what it calls the "non-delegable judgment" — the claim that decisions over life and death cannot, even in principle, be transferred to a system that does not bear moral responsibility. That argument has political consequences: it puts pressure on member states of the UN Convention on Certain Conventional Weapons to move beyond voluntary codes and toward binding restrictions.

On labor and dignity, the encyclical takes care to avoid two predictable framings. It does not adopt the technophobic line that AI is intrinsically destructive of work, and it does not adopt the technophilic line that automation will resolve itself through productivity gains. Instead, it proposes a third frame: that the moral test of an AI deployment is whether it expands or contracts the worker's capacity to act with judgment and dignity. That formulation is portable. It maps relatively cleanly onto the kind of governance language that European labor councils, U.S. unions, and Japanese corporate HR departments are already using.

One additional element of the labor section deserves attention. The encyclical reportedly devotes substantial space to what it calls the "intermediate technologies" — the tools, processes, and management systems that mediate between a worker and a deployed AI system. These include performance monitoring, workflow scheduling, and algorithmic task assignment. The text's claim, in essence, is that the moral status of an AI deployment cannot be evaluated in the abstract; it must be assessed in light of the mediating systems that shape how workers actually encounter the technology day to day. This is a more sophisticated argument than most religious commentary on automation has historically made, and it is the kind of analytical move that gives the document credibility with secular ethicists who might otherwise dismiss it.

On personhood, Magnifica Humanitas reportedly engages — without naming it directly — the question of whether sufficiently sophisticated AI systems should be treated as bearers of moral consideration in their own right. The Church's position, unsurprisingly, is no. But the argument it offers is not the simple one that personhood requires a soul. It is the more nuanced claim that personhood is constituted in relation: human persons are persons because they exist within webs of recognition, responsibility, and obligation with other persons. Machine systems, on this view, cannot enter those webs in the same way, regardless of behavioral sophistication. Whether or not readers find that argument convincing, it is at least an argument that can be engaged with on philosophical grounds rather than dismissed as theological assertion.

Anthropic Co-Founder Christopher Olah Speaks at the Vatican Launch

According to Vatican News, the launch event in Rome included a presentation by Christopher Olah, the Anthropic co-founder widely regarded as a leading voice in mechanistic interpretability — the research field that attempts to explain a model's internal behavior in terms humans can actually reason about, rather than treating the model as a black box (Vatican News). The presence of a frontier-lab researcher on a Vatican stage is, in itself, news. It marks one of the first occasions on which an active AI lab co-founder has spoken in an official Holy See setting in connection with a magisterial document.

Olah's research history matters here. His work on interpretability has long been organized around the premise that understanding what a model is doing internally is a precondition for assigning responsibility for its outputs. That is, in a different vocabulary, a moral claim — and it is unusually close to the encyclical's insistence on "non-delegable judgment." The two traditions are arriving at the same problem from opposite directions. The Church reaches it through theology of the human person; interpretability reaches it through reverse engineering. The convergence is what made the Rome appearance more than ceremonial.

The broader conceptual resonance is that Anthropic's "Constitutional AI" framework — training models to follow a written set of principles, with rejection sampling and reinforcement learning anchored to those principles — and the Church's notion of human dignity as a foundation that cannot be traded off are, in different idioms, both attempting to inscribe normative claims into systems that increasingly act on the world. Neither approach is sufficient on its own. But the fact that they can be discussed together in a single room, in front of the press, is a signal that the boundary between "technical AI safety" and "applied ethics" is thinning faster than most industry observers expected.

Business Impact — AI Ethics Becomes a Boardroom Agenda

For executives outside the technology industry, the practical question is whether any of this affects a quarterly plan. The honest answer is yes, but indirectly. Expect European multinationals to begin citing Magnifica Humanitas in AI governance committee discussions, particularly in sectors where moral framing already shapes procurement: defense contractors and their software suppliers, workforce-monitoring SaaS vendors, surveillance-adjacent computer vision providers, and any vendor selling AI into healthcare, education, or social services. In several of those sectors, the encyclical will not be the binding instrument. It will, however, be cited as the moral reference point against which corporate behavior is evaluated.

The broader shift is that AI adoption decisions are no longer purely an ROI conversation. They now require internal consensus on a second axis — whether a given use case is compatible with the organization's stated stance on human dignity. That doesn't mean slowing deployment; it means building the governance scaffolding to defend deployments when challenged by employees, regulators, the press, or institutional investors. Over the next two to three years, designing that scaffolding will be a recurring item on the C-suite agenda, especially for companies with significant European exposure.

There is a separate, narrower implication for AI vendors. Whether or not a buyer's procurement team formally cites Magnifica Humanitas, the document gives them a new rhetorical tool to push back on aggressive sales motions. Vendors that have not invested in interpretability, model cards, evaluation transparency, and clear data-handling commitments will find those omissions raised more often, and earlier in the sales cycle. The vendors best positioned to navigate the change are the ones that have already organized their public communication around language compatible with the encyclical's frame — most obviously Anthropic, which is part of why Olah's appearance in Rome is being read as strategically as well as substantively significant.

It is also worth noting how the encyclical will likely interact with the broader institutional investor community. Large pension funds, endowments, and sovereign wealth funds have, over the past three years, expanded their ESG and responsible-investing screens to include AI-specific criteria. Most of those screens currently rely on a patchwork of frameworks: the OECD AI Principles, the EU AI Act risk categories, the NIST AI RMF, and a handful of industry-specific standards. Magnifica Humanitas does not replace any of these, but it provides a reference text that fund managers can cite in stewardship engagements without appearing to take an idiosyncratic stance. Expect to see the encyclical referenced in proxy voting guidance and stewardship reports from European and Latin American asset managers within the next twelve months, with knock-on implications for how listed AI vendors describe their governance practices in annual reports.

OpenAI Files Confidential IPO, Anthropic's $30B Round at $900B Nears Close

The final week of May 2026 will be remembered as the week the two largest generative AI companies moved their capital structures in parallel. Each story stands on its own; together, they reshape the competitive map at the top of the industry, and they do so in a way that will be felt by enterprise buyers within the next two quarters.

OpenAI's Confidential S-1 Filing (May 22) — $852B–$1T Range, September Roadshow

CNBC reported on May 20 that OpenAI had entered the late-stage preparation phase for a confidential S-1 filing with the U.S. Securities and Exchange Commission (CNBC). Fortune followed on May 22 with the additional detail that the target valuation has been structured in a band of $852B to $1T (Fortune).

Goldman Sachs and Morgan Stanley are reportedly serving as joint lead underwriters, with the roadshow penciled in for September. Because the filing uses the SEC's confidential submission process — which allows an issuer to advance through review without immediately disclosing the prospectus publicly — full audited financials are not expected to enter the public record until shortly before the roadshow begins. For readers unfamiliar with the mechanics: confidential submission was originally designed for emerging growth companies under the JOBS Act and has since been broadened. It lets the issuer iterate with SEC staff while keeping the prospectus out of competitors' and journalists' hands until the company is ready to market.

Two strategic motives are commonly cited for the confidential route in this case. First, it limits how much detail competitors — Anthropic in particular — can glean from OpenAI's financial disclosures during the pre-IPO window. Compute costs, gross margins on inference, the split between API and consumer revenue, and the economics of enterprise contracts would all be visible in an open S-1. Each of those is something Anthropic would prefer to study before its own public communications. Second, the confidential path allows OpenAI to release headline numbers as a surprise close to the roadshow, maximizing investor attention rather than letting the story dissipate over months.

Fortune notes that, if the deal prices at the upper end of its range, it would rank among the largest technology IPOs in U.S. history. It would also crystallize a set of questions that have been hanging over the AI industry since 2023: what gross margin generative AI companies actually achieve once compute costs are netted against revenue, how durable the consumer subscription base really is, what proportion of revenue is concentrated in the top few enterprise contracts, and how the relationship with Microsoft is structured in financial terms. Each of those is a question a confidential filing eventually has to answer, and the answers will set the comparable baseline for every other AI company in the public markets for the next several years.

There is one further wrinkle worth flagging for international readers. A confidential filing does not commit OpenAI to going public on the targeted timetable. If market conditions deteriorate, or if the September roadshow does not draw the demand the underwriters expect, the company can delay or restructure the deal with relatively little public consequence. The probability of an actual September listing is high but not assured; the probability that the filing itself reshapes private-market expectations is essentially one.

A second wrinkle concerns governance. OpenAI's corporate structure — a nonprofit parent overseeing a capped-profit subsidiary, with Microsoft holding a contractual claim on a significant share of future profits up to a defined ceiling — is one of the most-watched structural questions for the IPO. The S-1, when it eventually becomes public, will need to disclose the precise mechanics of that relationship, including any modifications to the cap, any preferences attached to Microsoft's stake at IPO, and the role of the nonprofit board in the listed entity's governance. Each of these elements has been the subject of reporting and speculation for the past two years; the filing will collapse much of that uncertainty into specific contractual language. For investors, the financial numbers will be the headline. For governance specialists and regulators, the structural disclosures will be at least as consequential.

Anthropic's $30B Round at $900B+ Valuation "Closing Next Week"

On the same May 22, Bloomberg reported that Anthropic is on track to close a roughly $30B funding round at a post-money valuation above $900B, with the closing expected "as soon as next week" (Bloomberg). The lead investors are reportedly Sequoia, Dragoneer, Greenoaks, and Altimeter, each committing in the neighborhood of $2B. The remainder is being filled by a syndicate that has not been fully disclosed but reportedly includes sovereign wealth participation alongside existing strategic investors.

The financial picture supporting the valuation is striking. Bloomberg cites a Q2 revenue outlook of $10.9B and, importantly, a first-ever quarterly operating profit of $559M. If both numbers hold through quarter-end, Anthropic would become the first frontier AI lab to combine hyperscaler-tier revenue with positive quarterly operating margins. In private-market terms, the $900B+ valuation would also exceed the lower end of OpenAI's IPO valuation range, which is itself a meaningful milestone: it is the first time a private generative AI company has been priced above the public-market expectation for its closest competitor.

The contrast with OpenAI's path is unusually clean. Anthropic is pursuing profitability and a premium private valuation in parallel; OpenAI is moving toward public-market liquidity in a structure that will pressure-test the assumption that private investors have been pricing AI labs correctly. The fact that Sequoia is leading is also a governance signal. Sequoia's traditional posture — long holding period, light operational interference, board representation but not control — gives Anthropic room to scale while keeping a degree of independence in its boardroom that a near-term IPO would foreclose. That matters because Anthropic's strategic positioning is, in part, premised on being able to refuse certain commercial opportunities (defense work, content moderation engagements, regions with limited rule-of-law guarantees) without short-term shareholder pressure to revisit those decisions.

It is worth being precise about what is and is not confirmed at this point. The Bloomberg report is detailed and sourced, but Anthropic has not made an official statement, and final terms can shift before signing. The most likely scenario, based on what is currently public, is a closing in the last week of May or first week of June, with a public announcement that confirms the headline numbers and possibly adjusts the disclosed list of investors.

A related question that the reporting does not yet settle is what Anthropic intends to do with the capital. The most plausible uses are some combination of expanded inference capacity (which directly supports the Q2 revenue trajectory), continued investment in training infrastructure for next-generation models, and deeper enterprise go-to-market spending in geographies where the company is currently underweight. There has also been speculation about strategic acquisitions, particularly in interpretability tooling, evaluation infrastructure, and verticalized AI applications, though no specific deal has been confirmed. The general pattern across recent frontier-lab rounds is that the disclosed use-of-proceeds language tends to emphasize compute and research; the actual deployment of capital tends to include a meaningful enterprise sales component that is less visible from the outside.

Side-by-Side: OpenAI vs Anthropic