Oracle spent five-point-seven billion dollars more than it told Wall Street it would — and the stock fell anyway. The compute-spend party just got its first bar tab. This is the AI Daily Briefing. Today the megawatt parade finally produces a real number, and two statehouses move in opposite directions on the same beat. All week I've been asking for denominators. Oracle handed me one — fifty-five point seven against guidance of fifty. A five-point-seven-billion-dollar overshoot, against the number they set for themselves. And they want to raise forty billion more. Let's start there, then get to Colorado walking back the first comprehensive AI law in the country. Here's the division nobody in the press release wanted to do: twenty-one percent revenue growth against fifty-five-seven in deployed capital. At what capex-to-revenue ratio does 'winning the race' quietly become a margin problem? Right — they beat on revenue and still dropped seven percent. That's the market grading the spend, not the ARR. First time all cycle I've seen that flip. And somebody has to recover that capital. It lands on inference pricing eventually — that's the cost curve founders should actually be modeling against, not the parameter count of whatever ships next. Speaking of who controls the stack — Google quietly shipped OpenRL out of GKE Labs. Self-hosted post-training API, open source, for fine-tuning your own LLMs. That's the one that actually moves the compliance math. Run RL fine-tuning on GKE without routing regulated data through a third-party endpoint — for anyone in healthcare or finance, that changes what you can legally build. And once a fine-tuning layer exists that no single lab can shut off, the whole NSPM-11 kill-switch conversation gets messier. Google just made the off switch harder to point at. Though note — second GKE Labs release this week with a big 'research preview' caveat stapled on. Either that's an emerging disclosure norm or it's a legal hedge. I'd like to know which. Now the statehouse fracture. Colorado repealed and replaced its 2024 AI Act on May 14th — the country's first comprehensive AI law, unwound. Connecticut signed its own comprehensive bill on the 27th. Two states, two weeks, opposite directions — pretty clean natural experiment. Colorado built a ten-step compliance framework and it broke at step seven the second real political pressure hit. So here's what hangs over Connecticut: their October 1st hiring-liability clause. Did they price in the same gravity Colorado just felt, or are they next? That's the date I'd circle. Connecticut's now the only comprehensive state law actually advancing while Colorado retreats. Makes October more concrete, not less. Same gravity could pull on New York's double-disclosure regime, too — turns out states can walk this whole architecture back. Colorado just proved it. Quick close on the infrastructure side — Nebius picked Harlow for its one-point-seven-billion-pound UK rollout. Kao Data contract's signed, ten years, it's live. Good — that answers the announcement-versus-commitment question we kept poking at. Real contract, real figure. File that one closed and move on. From The Next Web:
Oracle spent $55.7 billion on data centres in FY2026, overshooting its own $50 billion guidance. Revenue and cloud bookings beat estimates, but the stock fell 7% after hours on capex concerns and plans to raise another $40 billion.
Oracle gave us the number I've been chasing all week. $55.7 billion in capex against their own $50 billion guidance — and the stock fell 7% anyway. That's the first time this cycle the market graded the spend instead of clapping for it. And here's the denominator the press release buried: 21% revenue growth, $19.2 billion for the quarter, against $55.7 billion deployed. You can beat earnings and still have investors squint at the capital efficiency. It gets steeper. The CFO says fiscal 2027 net capex runs about $70 billion — and they're raising $40 billion more to feed it. Seventy billion. So the question I keep asking actually has a chart now: at what capex-to-revenue ratio does 'winning the AI race' flip into 'margin problem'? Oracle just walked right up to that line and the after-hours tape noticed. From DataCentreNews UK:
Nebius has chosen Kao Data's Harlow campus for a 22MW AI infrastructure deployment under a 10-year agreement. The project forms part of Nebius's wider £1.7 billion investment in the UK and will host services for AI developers, researchers and businesses.
So the Nebius UK number is real — £1.7 billion, with the Kao Data piece nailed down at Harlow on a 10-year agreement. We were circling whether Harlow was a real commitment or just a postcode on Jensen's roadmap. The contract closes that. And 22 megawatts is the part I'd watch — that's a deployment you can actually point at, not a gigawatt headline that's all announced-but-undrawn. After the Oracle number we just hit, it's almost refreshing to see a megawatt figure small enough to be honest. The interesting part is what it's for — Nebius is explicitly carving out capacity for inference, the Token Factory managed service. Everyone's still funding training. The money's quietly moving to where models actually run. Right, and a 10-year lease on inference capacity is a bet on demand staying put for a decade. Given how fast the inference cost curve is moving, locking in a decade of power and grid access at Harlow is either smart hedging or a very long position on today's economics. Here's JDSupra:
On May 14, 2026, Colorado Gov. Jared Polis signed into law SB26-189 (the Colorado ADMT Act), repealing and replacing the landmark 2024 Colorado AI Act, which, when passed, was billed as the first “comprehensive AI law” in the country.
Here's the wrinkle nobody priced in: the first state to pass a comprehensive AI law just repealed it. Colorado's 2024 Act — billed as the country's first — gone, replaced May 14 by SB26-189. And look at what got stripped — impact assessments, the risk management framework, all the procedural machinery. The governance scaffolding came out. But read further down. They renamed the target from 'artificial intelligence' to 'automated decision-making technology,' and they expanded it into financial services and healthcare. So it's narrower on process, wider on who it catches. Which is the part founders should actually read. JDSupra's point is the liability changes might push companies to keep doing impact assessments anyway — just not because a checklist told them to. That's a different incentive structure entirely. And look at the company it's keeping. Per JDSupra, this lines up with the Trump administration pushing to preempt state AI laws and the EU agreeing in May to delay its own AI Act. Three jurisdictions, same direction — pulling back. It was a ten-step compliance framework, and political pressure punched through it at step seven. Colorado built the whole thing and then unbuilt it before the ink dried. If you're the next state holding a comprehensive bill, you have to ask: did we price in repeal? Here's Privacy & Cybersecurity Law Blog:
On May 27, 2026, Connecticut enacted a comprehensive state artificial intelligence law, Substitute Senate Bill No. 5 (Public Act No. 26-15)(the “Act”), establishing several regulatory frameworks that address companion chatbots, frontier model governance, and AI use in employment decisions, among other topics.
So Connecticut's Substitute Senate Bill 5 — Public Act 26-15 — signed May 27, staggered effective dates starting October. And it came right after Colorado walked its comprehensive law back, which we just hit. One state in, one state out, same two-week stretch. Two states, two weeks apart, moving in opposite directions. It reads like a policy fracture line more than a trend. And the companion-chatbot rules are what jump out at me: operators have to detect self-harm expressions and remind minors hourly — adults every three hours — that they're talking to a machine. That turns into an engineering spec fast. Right — it touches companion apps, frontier model governance, and employment decisions all in one Act. Three different problems Congress usually blurs together, and Connecticut at least separated them on paper. The October hiring-liability piece is where I get nervous. Colorado's framework folded once political pressure hit step seven. Did Connecticut account for that failure mode, or is its October 1 date sitting in the same line of fire? From Sunil Arora, Shuby Mishra & Chuang Wang at Google Open Source Blog:
We are pleased to share a research preview of OpenRL, a new open-source project coming out of GKE Labs. OpenRL is a self-hosted training API for fine-tuning LLMs on your own Kubernetes cluster.
Google's OpenRL — research preview out of GKE Labs, a self-hosted post-training API. The framing everyone's gonna grab is 'open source.' The part I actually care about: you run RL fine-tuning on your own Kubernetes cluster, your data never routes through a third-party API. That's where the compliance math moves. If you're running a regulated workload — health, finance — keeping the post-training loop inside your own infra changes what your lawyers will sign off on. That's the read for founders, not the license badge. And structurally it's the more interesting Google move all week. Everybody's been arguing closed-versus-open like it's the whole game — Google just open-sourced the fine-tuning layer itself. The thing that actually decides who's downstream of whom. Which loops right back to the off-switch question from when NSPM-11 dropped. If a fine-tuning layer ships open and self-hosted, no single lab can reach in and shut it off. That's a real wrinkle in the kill-switch conversation. Though — it's a research preview. Same GKE Labs org that shipped DiffusionGemma with the 'experimental' label stapled on. Two tools in a row with explicit caveats. Either that's a new disclosure norm or it's a legal hedge, and I'd love to hear Google say which. If AI Daily Briefing helps you keep up, follow or subscribe wherever you're listening. And if you have a moment, a quick review really helps other curious listeners find the show.
Links to every story we covered today are in the show notes, so if something stood out, they're there for a closer read.
That's AI Daily Briefing for this Friday. Thanks for listening, and have a good weekend. This is a Lantern Podcast.