Good morning,
This week we cover, Pope Leo XIV first encyclical focused on - you guessed it - AI, Alphabet new Gemini 3.5 flash models plus some other news from Google I/O (their annual dev conference). Plus the usual top news including Nvidia earnings and Meta layoff of 10% of its staff. But before that, we start with an update on our last edition’s big story on OpenAI-Anthropic-and-SpaceX IPOs.
Triple-IPO update
SpaceX went public-facing on its filing on 20 May (June listing targeted) at a ~$1.75tn valuation and a $75bn raise (the largest IPO in history). The filling gives us a first public view of the finance of the company. With $18.67bn total revenue: Starlink is the engine with $11.4bn of 2025 revenue (61% of the total) and the only profitable unit (~$4.4bn operating profit); the original launch business brought in $4.1bn (~22%) but lost money; and the newly merged AI arm, now "SpaceXAI" (xAI/Grok), generated under $1bn (~5%) but lost $6.4bn, single-handedly turning a profitable company into a $4.94bn net loss. The most unusual feature is the share allocation: SpaceX plans to steer up to ~30% of the offering to retail investors, with the remaining ~70% institutional, and retail buyers get the same price and timing as the big funds.
OpenAI is reportedly working with Goldman Sachs and Morgan Stanley to make a confidential IPO filing in the coming weeks, moving their target date from early 2027 to September or October 2026. Citi and JPMorgan are expected to also join the working group.
Finally, Anthropic closed an another new round for $65bn at $965bn valuation (x3 the last one from February). Some suggest this is meant to be ahead of OpenAI (currently valued at $850bn) at IPO.
Some interesting readings: A NYTimes paper by Jeff Sommer on why those massive IPOs are likely to deliver poor return for ordinary investors; a YouTube video from Casual Finance explaining the impact of the SpaceX IPO on the financial system and a recent NASDAQ rule update that facilitate those IPOs.
Google I/O: Gemini 3.5 Flash, Spark and Antigravity 2.0
On May 19 and 20, Google held its annual developers conference: Google I/O; this year, the focus was: cost. Sundar Pichai opened with the numbers that justify Google's roughly $180–190bn capex plan for 2026 (about six times its 2022 spend): the Gemini app now has ~900m monthly users, and Google processes ~9.7 trillion tokens a month. The strategically important announcements were not about a smarter model. They were about a cheaper, more autonomous one.
The headline launch was Gemini 3.5 Flash. In Google's line-up, "Flash" is the lightweight, cheap-and-fast tier — while the heavyweight "Pro" tier did the hard thinking. The news is that this Flash has largely closed that gap: it posts scores on demanding tasks that would have been flagship-tier just a generation ago, while running at roughly four times the output speed of rival frontier models and a fraction of the price: about $1.50 per million words of input and $9.00 per million of output. Speed matters more than it sounds: an AI "agent" completing a task may chain together dozens of steps, so a model that is four times faster turns a workflow that took minutes into one that takes seconds.
The consumer headline was Gemini Spark, pitched as a 24/7 personal agent rather than a chatbot you prompt one turn at a time. Spark is meant to work in the background across your Gmail, Docs and the wider Workspace suite — triaging an inbox, drafting replies, pulling scattered material into a single document. For developers, Google unveiled Antigravity 2.0, an "agent-first" coding environment very similar to Claude Code or Codex. And Gemini Omni is a video generator notable for being "grounded in physics" — water flows, objects fall and collide roughly as they would in the real world, sidestepping the dreamlike inconsistency of earlier AI video. The common thread across this I/O is the shift from "a chatbot you ask" to "an agent that acts."
And a 60% price cut. Google slashed its top AI Ultra subscription from $250 to $100 a month and moved from daily prompt limits to a compute-based model. Coming after last month's broader collapse of all-you-can-eat AI pricing.
Why it matters: Google's bet is that owning the distribution (billions of users across Search, Android and Workspace) plus winning on cost beats winning the benchmark. The planning assumption should be that raw model capability keeps getting cheaper and more interchangeable — so the durable advantage sits in proprietary data, workflow integration and distribution, not in which model you license this quarter.
The Pope takes on AI
On Monday 25 May, Pope Leo XIV personally presented his first encyclical, Magnifica Humanitas ("Magnificent Humanity"), a five-chapter document on artificial intelligence. The pontiff turned up to the press conference himself with a Silicon Valley AI executive. It is an interesting document, but more importantly a showcase that the Vatican has quietly become a genuine actor in AI governance.
What the document says. Leo's argument is not Luddite (i.e. opposed to technology). The encyclical insists technology is neither "a force antagonistic to humanity" nor inherently evil but that it is "never neutral," because it takes on the characteristics of those who build, fund, regulate and use it. From there it runs through the now-familiar list: job insecurity, manipulation of information, surveillance, bias, autonomous weapons, the concentration of technological power in a handful of corporations, even AI's carbon footprint; and lands on a deeper worry: that we may start to treat ourselves and each other as products to be optimised. The headline word is "disarm." On autonomous weapons in particular, Leo is blunt: "No algorithm can make war morally acceptable."
"Leo". None of this is improvised. Robert Prevost — the first American pope, from Chicago — chose the name Leo to invoke Leo XIII, whose 1891 encyclical Rerum Novarum was the Church's response to the upheavals of the Industrial Revolution and the rights of workers. Leo XIV signed Magnifica Humanitas on 15 May, the 135th anniversary of that document, to the day. He has repeatedly called AI "another industrial revolution" and ranked it alongside migration and climate as a defining challenge of the age. The Church is consciously updating 130 years of social doctrine for the algorithmic era.
This is a decade in the making. The surprise guest at the unveiling was Chris Olah, co-founder of Anthropic, whose line "we need informed critics who will tell the labs when we are failing" captured the unusual sight of an AI executive endorsing the Church's critique. But the Vatican's tech diplomacy long predates this moment. Informal "Minerva Dialogues" between Church officials and tech leaders date back to around 2016 — convened, with no small irony, in the very Roman church where the Inquisition once forced Galileo to recant; in 2020 the Vatican convened the Rome Call for AI Ethics, signed first by Microsoft and IBM, later by Cisco and Qualcomm; it published a doctrinal note, Antiqua et Nova, in January 2025; and Pope Francis became the first pontiff to address a G7 on AI. As the initiative's architect, Archbishop Vincenzo Paglia, has put it: "The Vatican is not an expert on the technology but on values." That is precisely the gap it is filling.
Interested to learn more? Elias Wachtel in The Atlantic explains the relationship between the Pontiff’s state and Silicon Valley in much details.
Why it matters: AI governance now has three poles: industry, government, and a values-based pole with no commercial or national stake, anchored by a 2,000-year-old institution that has turned itself into a convening power. For anyone with European exposure, watch how this moral framing reinforces the regulatory one: it gives the "responsible AI" camp a louder voice in exactly the markets where the EU is already the strictest.
More top news
NVIDIA shows (again) a new record quarter with $81.6bn revenue (+85%), data-centre up 92% — yet the shares slipped on "sell the news". China data-centre sales were essentially zero (against $4.6bn a year ago). Guidance now assumes no China contribution at all: the export-control standoff has gone from blockade to a market NVIDIA no longer bothers to forecast.
Meta cuts ~8,000 jobs, then pledges "no more." About 10% of staff went on 20 May (Asia first, 4am notices) to help fund a $125–145bn capex year. Zuckerberg told the survivors he does not "expect other company-wide layoffs this year" — wording employees were quick to note he has used, and then broken, before.
Amazon scraps its AI scoreboard. It quietly killed an internal leaderboard ranking staff by AI usage after engineers gamed it with pointless tasks ("tokenmaxxing") that ran up Amazon's own compute bill. A clean lesson in Goodhart's Law for anyone setting AI-adoption targets: the moment a measure becomes a target, it stops measuring anything.
The White House drafts an "FDA for AI" — then stalls. Officials briefed Anthropic, Google and OpenAI on an order letting agencies vet frontier models before release (triggered by Anthropic's withheld Mythos model), then postponed the signing on 20 May, with the President reportedly unhappy with parts of it after a few calls with Silicon Valley’s Executives.
Anthropic wire to Elon Musk. SpaceX's S-1 shows Anthropic paying $1.25bn a month (~$15bn a year, up to ~$40bn over the term) for the Colossus clusters through 2029 — roughly 80% of SpaceX's entire 2025 revenue. Then, mid-roadshow, Musk took to X to recast the deal as a six-month lease. An odd thing to downplay when it underwrites your IPO.
China keeps pace on capability, leads on price. Alibaba's Qwen 3.7 Max (agent-first, 1m-token context, API-only) tops the Chinese models on the public leaderboards — still behind the US frontier, still cheaper, still shipped without open weights.
Anthropic poaches Karpathy. OpenAI co-founder and Tesla former AI VP, Andrej Karpathy joined Anthropic's pre-training team on 19 May. A new high-profile talent move, and one more arrow pointing the same way.
Person to Know: Demis Hassabis
The Financial Times recently revealed that the CEO of Google DeepMind, was personally an early investor in Anthropic — a previously undisclosed stake in a direct rival. This CEO is Demis Hassabis, a human any person interested in AI should know.
Demis Hassabis is the co-founder and CEO of Google DeepMind, the lab behind AlphaGo and a string of scientific and engineering breakthroughs that helped turn “AI research” into “AI power.” Born in London in 1976, he was a child chess prodigy, studied computer science at Cambridge, worked briefly in neuroscience, then built games before returning to research — a career arc that explains his obsession with intelligence as both a product and a system.
Why he matters is not just that he runs a flagship lab. Hassabis has been one of the most effective “institution builders” in modern AI: he assembled DeepMind’s early talent, sold the company to Google in 2014, and has since helped steer the industry’s centre of gravity toward large-scale, research-driven model building. His worldview is now the default playbook for the frontier.
And the twist in the FT story is what makes him newly relevant outside the lab: an early, personal stake in Anthropic suggests that even the leaders of the biggest incumbents are hedging. In an era where today’s rival can become tomorrow’s platform, Hassabis is both a symbol of AI’s first wave and a clue to how the second wave is being financed and positioned.
Interested to learn more about a man who contributed to so much of the current AI development: read his latest bibliography by Sebastian Mallaby The Infinity Machine (summary in The Economist) or listen to his last discussion with Lex Friedman.