1. MAGNIFICA HUMANITAS: WHAT THE CHURCH'S AI DOCTRINE ACTUALLY DEMANDS
Pope Leo XIV published Magnifica Humanitas today, the Catholic Church's first encyclical specifically addressing artificial intelligence, framing it as a document about the defence of human dignity in the face of technologies capable of simulating — and in some domains surpassing — human cognitive capabilities. The title, rendered in English as "the magnificent humanity," is a deliberate rhetorical choice: the Church's response to AI is not primarily a warning but a statement of anthropological confidence, arguing that what makes human life valuable cannot be replicated by statistical pattern matching over training data, however capable those systems become. Christopher Olah, Anthropic co-founder and one of the field's most respected researchers on mechanistic interpretability, was among the lay speakers at the Vatican launch — an unusual pairing that signalled both the Church's seriousness about engaging the technical community and Anthropic's willingness to participate in a moral framework that extends beyond the lab's own safety discourse.
The encyclical addresses three domains with specific policy implications. The first is labour: Magnifica Humanitas argues that work is not merely an economic activity but a participation in human dignity, and that displacement of workers by AI systems without corresponding social investment — in retraining, in safety nets, in the redistribution of productivity gains — is a moral failure, not simply an economic adjustment. This is a stronger claim than the typical corporate AI ethics framing, which tends to treat workforce disruption as a transition problem rather than a justice problem. The second domain is data and consent: the encyclical takes a clear position that the extraction of human creative, cultural, and intellectual output to train AI systems, without consent or compensation, constitutes a violation of human dignity that cannot be justified by the aggregate social benefits of more capable models. The third domain is autonomous weapons: the Church's position is categorical — lethal autonomous systems that select and engage targets without human authorisation are incompatible with the moral requirement that the decision to take a human life remain a human decision, regardless of the military advantage such systems may provide.
The practical significance of Magnifica Humanitas is not that it will change the behaviour of AI labs or governments overnight. It is that the Catholic Church, as an institution with continuous presence in virtually every country on Earth, relationships with governments across the political spectrum, and a moral tradition that predates the nation-state, has now formally positioned these three questions — labour, data, and autonomous weapons — as matters of moral obligation rather than policy preference. That positioning changes the terms of advocacy for the billion-person constituency that takes the Church's moral guidance seriously. It also, notably, aligns with the positions of Anthropic on all three counts — the company has refused Pentagon contracts requiring removal of safety guardrails on autonomous weapons, has publicly supported data rights frameworks, and has made workforce transition part of its public mission. The encyclical does not mention Anthropic by name. It does not need to.
2. GOOGLE I/O 2026: GEMINI OMNI, ANTIGRAVITY 2.0, AND THE INTELLIGENCE PLATFORM
Google's I/O keynote on Tuesday delivered the clearest articulation yet of where the company is positioning itself in the AI layer stack. Gemini Omni, the centrepiece model announcement, is not a successor to Gemini 3.5 Flash in any straightforward sense — it is a different architecture designed around continuous multimodal reasoning rather than sequential input processing. Where previous multimodal models accept text, image, or audio input and produce text output, Gemini Omni processes and generates across all modalities in a single continuous reasoning stream, grounded in what Google describes as real-world knowledge rather than training data alone. The capability that best illustrates the distinction is video: Gemini Omni can accept live video input and generate a parallel audio commentary, translation, or analysis stream in real time, not by treating the task as a sequence of discrete image frames but by maintaining a coherent understanding of the scene across time. That difference in architecture has meaningful consequences for agentic applications where the agent needs to understand a changing environment rather than a static document.
Antigravity 2.0, Google's agent-first development platform, is the infrastructure layer that makes Gemini Omni practically deployable in agentic contexts. The core capability — launching a managed remote Linux environment with a single Gemini API call, then passing the agent the tools it needs to reason, plan, and act within that environment — removes the primary engineering friction that has limited agentic deployment to date. Building agent infrastructure today requires a developer to provision compute, manage state across agent turns, implement tool calling frameworks, handle errors and retries, and monitor agent behaviour in production. Antigravity abstracts all of that into a managed service, reducing the development surface for an agentic application to the model call and the tool definitions. Google is also introducing Managed Agents in the Gemini API as a production-grade offering — agents that can be deployed, monitored, and scaled through the same infrastructure as any other Google Cloud workload. The combination of Gemini Omni's reasoning capabilities and Antigravity's deployment infrastructure represents the most complete end-to-end agent development platform announced by any major lab to date.
The competitive context of these announcements is important. Google I/O happened the same week that OpenAI filed its S-1, that Anthropic crossed $30 billion in ARR, and that the financial services sector made its most explicit statement yet about which AI vendors it trusts with its most sensitive workflows. Google is not behind in this competition in any obvious technical sense — its models are competitive, its infrastructure advantages are substantial, and its distribution through Android, Search, Workspace, and Cloud gives it touchpoints with more users and enterprises than any competitor. What I/O signals is that Google has decided the competition is no longer primarily at the model quality layer, where benchmark differences between Gemini Omni, Claude, and GPT variants are increasingly marginal. The competition is at the platform layer — which company's development infrastructure, deployment services, and integration standards become the default substrate on which the next generation of AI applications is built. Antigravity 2.0, paired with MCP support for Gemini Spark, is Google's answer to that question. Whether it arrives in time to establish the platform advantage before OpenAI's post-IPO capital and Anthropic's enterprise momentum further consolidate their respective positions is the strategic question that will define the next eighteen months.
3. OPENAI'S S-1: THE QUESTIONS A TRILLION-DOLLAR VALUATION CANNOT DODGE
OpenAI filed its S-1 confidentially with the Securities and Exchange Commission on Thursday, targeting a public listing in September 2026 at a valuation above $1 trillion, with Goldman Sachs and Morgan Stanley as lead underwriters. The filing follows the company's conversion from a non-profit to a public benefit corporation structure earlier this year, which was a necessary prerequisite for a public offering but also the most structurally contested decision OpenAI has made — triggering the Elon Musk lawsuit that a California jury dismissed in under two hours on May 19, and generating ongoing scrutiny from the original non-profit board about the governance implications of a permanent-capital structure backed by public shareholders. The revenue numbers are substantial: $25 billion in annualised recurring revenue, growing at approximately 60% year-on-year, with a consumer subscriber base that has crossed 700 million monthly active users. The loss numbers are equally substantial: approximately $14 billion in annual operating losses, driven by compute costs, research investment, and a sales motion that is increasingly dependent on the enterprise deployment arm OpenAI launched this spring with McKinsey, Bain, Goldman Sachs, and SoftBank as partners.
The unit economics question is the one the S-1 cannot avoid and which the prospectus will be read most carefully for. OpenAI's $25 billion in revenue is generated at a gross margin that the company has not publicly disclosed, but which analysts following the company closely estimate at between 55 and 65% — substantially below the 70-80% margins typical of software-as-a-service businesses, because the compute costs of serving inference requests at ChatGPT's scale are not optional infrastructure but a direct cost of goods. The $14 billion annual loss figure implies that OpenAI is spending approximately $39 billion annually — on compute, research, go-to-market, and infrastructure — to generate $25 billion in revenue. Closing that gap requires either a substantial increase in revenue per unit of compute (through better model efficiency), a substantial decrease in compute costs (through hardware improvements and architectural innovation), a dramatic acceleration in revenue growth (through enterprise penetration), or some combination of all three. The S-1 will be read to determine which of those levers management believes is most credible, and on what timeline. Public markets have been willing to price high-growth technology companies at extended loss multiples when the path to profitability is legible — the question is whether the OpenAI prospectus can make that path legible for a business that has no direct historical comparator.
The governance arrangements are likely to generate as much scrutiny as the financials. OpenAI's public benefit corporation structure gives the board authority to override shareholder interests in service of the company's stated mission — the safe development of artificial general intelligence for the long-term benefit of humanity — and the S-1 will need to define how that authority is exercised, how conflicts between mission and shareholder value are resolved, and what protections exist against the mission being redefined by future management. These are not theoretical concerns. The non-profit restructuring dispute, the Musk lawsuit, and the ongoing tension between safety commitments and commercial imperatives have all surfaced governance questions that the pre-IPO OpenAI could manage privately. A public company answers those questions to shareholders, regulators, and the press in real time. Whether the governance structure OpenAI has designed — one without precedent in public company law — can sustain that scrutiny, and whether public markets will accept the valuation premium implicit in a trillion-dollar listing of a company losing $14 billion a year, are the questions that will determine whether the September target holds.
4. ANTHROPIC AT $900 BILLION: THE CHALLENGER THAT BECAME THE MARKET
Anthropic this week confirmed it is closing a new funding round that sources say will price the company above $900 billion valuation — a figure that would, at the moment of closing, make Anthropic the most valuable private technology company in history and place it ahead of OpenAI's most recent secondary-market valuation ahead of the latter's IPO. The revenue trajectory underpinning that valuation is extraordinary by any measure: Anthropic crossed $30 billion in annualised recurring revenue on 1,400% year-on-year growth, having reached $4 billion in ARR in mid-2025 and doubled every six weeks since the start of 2026. The driver is enterprise: the company's Claude models are now deployed in production across financial services, healthcare, legal, and technology sectors, with the Claude API processing more tokens per day than any competing model in enterprise contexts. The Gates Foundation partnership announced this week adds a development-sector deployment dimension that no competitor has matched at comparable scale or commitment.
The revenue dispute with OpenAI is worth noting for what it reveals about the commercial structure of frontier AI. OpenAI has privately argued that Anthropic's $30 billion ARR figure is overstated by approximately $8 billion, pointing to questions about whether revenues generated through Anthropic's cloud partnerships with AWS and Google Cloud should be reported at gross value — the total payment from the enterprise customer — or net of the cloud provider's cut. This is not a trivial accounting question: Anthropic's Amazon and Google partnerships involve substantial capacity commitments from both cloud providers in exchange for preferred deployment status, and the revenue sharing arrangements that accompany those commitments affect how the $30 billion figure is calculated. The dispute is unlikely to be resolved before the funding round closes, but it matters for the same reason the OpenAI S-1 unit economics question matters — the gross margin of a frontier AI business is the variable that determines whether the valuations being assigned to these companies reflect durable commercial positions or optimistic projections about a cost structure that does not yet exist.
What the Anthropic valuation trajectory confirms, regardless of how the revenue accounting is resolved, is that the frontier AI market has bifurcated in a way that the field's competitive dynamics did not clearly predict two years ago. In 2024, the consensus view was that OpenAI's lead in consumer adoption, enterprise deployment, and model capability was durable enough that no competitor could close the gap without a multi-year runway. The Anthropic story of the past eight months — a company that went from $4 billion in ARR to $30 billion in less than a year, that is approaching OpenAI's valuation before OpenAI has completed its IPO, and that has established dominant enterprise positions in the sectors where AI deployment is most consequential — is a systematic falsification of that consensus. The competitive dynamic in frontier AI is not winner-take-all. It is a race between two companies with genuinely differentiated products, different missions, different commercial structures, and different risk profiles — and the outcome of that race will be determined not by which lab builds the better model next month, but by which commercial and governance infrastructure proves more durable over the next five years.
5. NEXTERA BUYS DOMINION: THE $67 BILLION WAGER ON AI'S ELECTRICITY APPETITE
NextEra Energy announced on Monday that it would acquire Dominion Energy in an all-stock deal valued at approximately $67 billion — the largest utility merger in United States history. The strategic rationale is explicit and unusual in its directness: NextEra's management stated that the primary driver of the acquisition is the projected electricity demand from AI data centre buildout. Dominion has connected more than 450 data centres in its service territory, with electricity demand in its coverage area rising faster than at any point since the post-war industrial expansion. NextEra, already the world's largest producer of wind and solar power, is acquiring the generation, transmission, and distribution infrastructure to service the specific geography — Northern Virginia, which hosts the highest concentration of hyperscale data centres on Earth — where AI-driven power demand is most acute. The combined company will become the largest regulated utility in the United States by asset base, with a rate base that will justify sustained capital investment in generation and transmission capacity for the decade following the merger's close.
The financial logic of the deal reflects a specific thesis about AI infrastructure that goes beyond a single acquisition. AI data centres are not ordinary commercial customers from a utility planning perspective. They are large, predictable, long-duration load additions with 20-year power purchase agreements, financed by companies — Amazon, Microsoft, Google, Meta — whose creditworthiness exceeds that of most sovereign governments. A utility that can acquire the right to serve that load base, and that can demonstrate the generation capacity to meet it reliably, gains a rate-base expansion opportunity that is both larger and more predictable than traditional industrial or residential demand growth. NextEra's $2.25 billion in bill credits committed to Dominion's existing customers is the political cost of winning regulatory approval for that expansion — a form of rate-smoothing that makes the consumer side of the utility merger more palatable to the state regulators whose approval is required. The combined company expects to spend the next decade building the generation and transmission infrastructure needed to meet projected AI power demand, using AI-driven planning and asset management tools to optimise the capital allocation. A utility using AI to build the infrastructure that powers AI is not a paradox. It is the logical endpoint of AI becoming infrastructure.
The broader significance of the NextEra-Dominion deal is what it says about the market's confidence in the AI-driven power demand thesis. Utilities are among the most conservative capital allocators in the American economy — regulated monopolies whose returns are set by state public utility commissions and whose investment decisions are subject to years of regulatory review. When the largest renewable energy utility in the world makes the largest utility acquisition in history on the explicit rationale that AI data centres will require more electricity than the existing grid can supply, it is not making a speculative bet. It is making a decades-long infrastructure commitment based on contractual evidence — signed power purchase agreements, committed capital expenditures from hyperscalers, and engineering assessments of the load growth already visible in existing data centre campuses. The 15-to-25% share of US electricity consumption that AI data centres are projected to represent by 2030 is not a forecast derived from AI optimism. It is a projection derived from capacity that is already under construction, funded by capital that has already been committed, and contracted against load that has already been signed. The grid is not preparing for an AI future. It is catching up with an AI present that the electricity infrastructure was not designed to serve.