Google I/O was the centrepiece. Google announced Gemini 3.5 Flash — a model that matches the intelligence of 3.1 Pro but runs at four times the speed and at Flash pricing — alongside Gemini Omni, a unified model that accepts text, image, audio, and video input and outputs video grounded in real-world knowledge. But the headline product was Gemini Spark, a 24/7 personal AI agent that runs on a dedicated Google Cloud VM, meaning it works even when your laptop is closed. You can email Spark directly via a Gmail address, track its progress through the new Android Halo system on mobile, and it integrates with Chrome, Canva, OpenTable, and Instacart. Sundar Pichai described it as the shift from an assistant that answers questions to an agent that does real work. Spark rolls out next week for AI Ultra subscribers — a tier Google slashed from $250 to $100 a month. Google also confirmed that Samsung Android XR smart glasses are coming this fall, built with Warby Parker and Gentle Monster, featuring live translation, navigation, and Gemini-powered notification summaries. No display — that's a 2027 story — but the companion-device model compatible with both Android and iOS could broaden reach fast.

On Tuesday, OpenAI announced that one of its reasoning models had autonomously disproved a central conjecture in discrete geometry — a problem posed by Paul Erdős in 1946 that had stood for 80 years. The model discovered an infinite family of point arrangements that yield significantly more unit-distance pairs than the classic square-grid approach, bringing ideas from algebraic number theory to bear on an elementary geometry question in a cross-domain leap human mathematicians hadn't explored. External mathematicians have verified the proof. OpenAI called it the first time a prominent open problem has been solved autonomously by AI, and the scientific community paid attention — Scientific American ran a feature calling it "amazing." The practical significance is still being debated, but as a signal of what reasoning models can do when pointed at unsolved problems, it's hard to dismiss.

SpaceX filed its S-1 with the SEC on Wednesday, targeting a Nasdaq listing under the ticker SPCX by mid-June at a valuation around $2 trillion — which would make it the largest IPO in history. The filing revealed that Starlink generated $11.4 billion in 2025 revenue with $7.2 billion in adjusted EBITDA, but also that SpaceX directed roughly 60% of its 2025 capital spending — around $20 billion — to its AI division, built largely around the Colossus data centre infrastructure. The filing also disclosed the terms of Anthropic's massive compute deal: $1.25 billion per month through May 2029, totalling roughly $45 billion across the contract term. Anthropic gets access to over 220,000 Nvidia GPUs at the Colossus 1 facility near Memphis, and announced an expansion to Colossus 2 the same day. The deal makes xAI — which operates the infrastructure — one of Anthropic's largest suppliers, a relationship neither company seems bothered by despite competing in the model layer.

The Musk v. Altman saga ended — or at least paused. A California federal jury in Oakland delivered a unanimous verdict on Monday, rejecting all of Elon Musk's claims against OpenAI and CEO Sam Altman. The jury deliberated for less than two hours after eleven days of testimony, finding that Musk had waited too long to bring his $150 billion lawsuit — the statute of limitations had expired before the 2024 filing. Musk called it "a calendar technicality" and said he'll appeal to the Ninth Circuit. Hours later he was in the news for a different reason: the SpaceX IPO filing. It was that kind of week.

Anthropic, meanwhile, quietly disclosed numbers that reframed the company's trajectory. Revenue is projected to hit $10.9 billion in Q2 — a 130% increase from the $4.8 billion reported in Q1 — with an expected operating profit of roughly $559 million, arriving two years ahead of internal projections. Claude Code alone has passed $1 billion in annualised revenue within six months of launch. The company cautioned that sustained profitability through the rest of the year is uncertain given coming compute and training costs, but for a company valued at $380 billion, showing it can generate cash — even episodically — changes the conversation around its eventual IPO. And on Thursday, Trump postponed the signing of a planned AI executive order that would have established a voluntary framework for companies to share frontier models with the government before release. "I didn't like certain aspects of it," he told reporters, adding he didn't want to do anything that might slow the US lead over China. The order was reportedly caught between pro-innovation factions wanting a genuinely voluntary framework and national security hawks wanting NSA-backed classified testing with teeth. For now, the US still has no federal pre-deployment review system for AI — and the Mythos problem that prompted the effort hasn't gone away.

Signals

Week in Review: May 22