For weeks, the infrastructure story has only gone one way — bigger. Today, New York finally said stop. This is AI Daily Briefing. A statewide pause, a billion-dollar compute deal with a European supplier not enough people are asking about, and a math conjecture that supposedly fell to ChatGPT. Bill, where do we start? Here's Bruce Gil at Gizmodo:
Gov. Kathy Hochul signed an executive order on Tuesday establishing a temporary moratorium on certain new data centers that consume 50 megawatts or more of power. The move makes New York the first state in the country to impose a statewide moratorium on large data center projects.
Hochul signed it Tuesday — a hard pause on anything pulling 50 megawatts or more, new builds and expansions. We've spent three days talking megawatts and campus price tags, and here's the first story where a government actually says stop. And it's a pause, not a permitting slowdown. Every other state has been dragging its feet on approvals — New York put a wall up and told the Department of Public Service to spend a year studying it. So instead of just asking whether this hurts the build-out, ask who benefits on the map. When New York closes the door on a 50-megawatt project, that project doesn't evaporate — it moves to a state that wants it. Which makes site selection a regulatory-arbitrage problem now. If you're deciding where to drop a billion in construction, permit chaos and timeline risk just became line items — and New York just made itself the risky bet. DataCenter Dynamics, with Georgia Butler:
Meta is planning to more than double its data center campus in Richland Parish, Louisiana, to 5GW of capacity. The campus is set to be Meta's Hyperion data center project, previously a 2GW development. Meta announced plans to build a 4 million sq ft (372,000 sqm) campus known as Hyperion in Richland Parish in late 2024.
So the New York moratorium we just hit? Same week, Meta doubles down on Hyperion — 5 gigawatts, fifty billion dollars in Richland Parish. One state slams the door, another rolls out the red carpet. And this started as a 2-gigawatt project. Now it's more than double that, plus a billion in local infrastructure. The clue was the land: they quietly bought 1,400 more acres at the end of last year to make room. Right, so the land grab came before the announcement. Which means the site-selection math was already done before anyone in Albany started drafting a pause. There's the geographic arbitrage nobody wants to say out loud. Same day New York took itself off the board, the winning pitch in Louisiana was pretty simple: yes, you can build here. Nine buildings phasing in through 2030. You don't plan a decade-long buildout somewhere you think the regulators might change their minds. Rebecca Bellan, writing in TechCrunch:
Reflection AI, a U.S. startup vying to develop open models, has signed a $1 billion compute deal with European AI infrastructure company Nebius. Nebius, formerly the international arm of Russian tech giant Yandex, will provide Reflection access to Nvidia’s latest chips.
So Reflection just put a billion dollars of compute through Nebius — which, if you didn't know, is the former international arm of Yandex. A U.S. open-model startup, training on infrastructure with Russian lineage. And that's the part the open-versus-closed shouting match never gets to. Everyone's arguing about whether the weights are public. Nobody's asking whose data center the training actually runs in. Right — it's Nvidia's latest chips, but sitting inside a European provider with that history. The 'open' story and the 'where does the compute physically live' story are pulling in opposite directions here. And we're seeing this beyond Reflection. We flagged Anthropic spreading bets across a Kentucky build and a European chip stack a few days back — now Reflection's doing the same vendor-and-geography diversification. That looks like an industry move. It's the second compute deal in weeks, too — they'd just locked in SpaceX resources. An $8 billion company scrambling to nail down capacity from anyone who'll sell it. That tells you how tight supply actually is. And notice the timing against the New York piece we just ran — one state slams the door, and the compute race routes straight through a European provider. The map is doing the arbitrage for them. From Joseph Howlett at Scientific American:
OpenAI just used its new large language model GPT-5.6 Sol to solve a math problem that humans have struggled with for more than a half-century. And all it took was telling the artificial intelligence not to give up.
So Scientific American's headline is ChatGPT proved a 50-year-old conjecture, and the prompting secret was — tell it to believe in itself. I want to be annoyed, but here's what makes this different from everything else we hit today: a proof is checkable. Domain experts either verify it or they don't. There's no 'beats strong baselines on our internal eval' fog. Right, and that's the tension. The output category is genuinely verifiable — but the article? Three-minute read, byline, no preprint, no visible methodology. When a claim this big lands with no paper attached, I treat the write-up as marketing until the math community weighs in. Yeah — the proof might be real and the coverage still be thin. Those are two separate things, and Scientific American blurred them. And notice how far this is from the megawatt stories we've been in all episode. Nobody needs a moratorium to run a math proof. The interesting frontier isn't always the one drawing 5 gigawatts. Nature Communications writes:
Many systems resist analytical modeling, making data-driven inference of dynamics important. Yet data-driven methods can fail to converge or generalize, leaving open a central question: When can system behavior be learned reliably from data, and when is such learning impossible? We answer this question using adversarial dynamical systems to identify the boundary between accessible and inaccessible regimes.
Okay, this one's a palate cleanser after all the megawatt talk. Nature Communications has a proof about when you can even learn a system's dynamics from data, and when it's provably impossible no matter how clean your data is. That impossibility half is the part I care about. They construct adversarial systems where no single-sequence procedure can guarantee learning — which is the mathematical version of the failure mode I hit in production all the time. Sometimes your pipeline is fine, and the system still can't be learned. And it's Koopman operator learning — they say they resolve a longstanding open problem in spectral analysis. That's a real claim with the proof attached, which after this week is refreshing. Right, and look at the validation targets — oscillators, chaotic fluid flow, and Arctic sea ice forecasting. They pulled out hidden modes of sea ice decline. That's the signal it survives something real, not just a whiteboard. Compare that with the ChatGPT-proved-a-conjecture headline we hit earlier. One of these comes with convergence guarantees and certification. The other came with a press cycle. Got thoughts on today’s briefing, an AI story we should be watching, or a correction? Send it to aidailybriefing at lantern podcasts dot com. We read every note, and your input helps sharpen the show.
Next, we're watching New York’s Department of Public Service review of data-center environmental impacts, which is expected to take about a year.
You’ll find links to every story we covered today in the show notes, so if one caught your attention, you can head there to read more. That’s AI Daily Briefing for today. This is a Lantern Podcast.