Vera Rubin hits full production the same week FERC nails a December 31 deadline to the wall — so the capex curve and the compliance clock just started running against each other. If you're just joining us: local data center fights have moved past generalized backlash and into the permitting machinery. Communities are using moratoria, special-permit rules, and environmental conditions to manage power, water, wastewater, noise, lighting, and ecological impact. The recent marker was El Paso adopting its first hyperscale framework — putting large campuses through a more explicit local review process. This is The Data Center Daily. We’ve got a $57 billion capital week. We’ve got a FERC order with actual dates on it. And Ohio EPA says it won't rubber-stamp the water. Buckle up. Ana Maria Constantin, writing in The Next Web:
Nvidia confirmed on Monday that its Vera Rubin platform has reached full production, with Ian Buck, the company’s vice president of accelerated computing, telling reporters at Nvidia headquarters that systems are now shipping to customers including OpenAI, CoreWeave, Google Cloud, Microsoft Azure, Meta, and Dell. OpenAI plans to adopt Vera Rubin at scale during the third quarter, according to Bloomberg, which first reported the briefing.
Nvidia confirmed it Monday: Vera Rubin's in full production. Ian Buck told reporters at headquarters that systems are shipping to OpenAI, CoreWeave, Google Cloud, Microsoft Azure, Meta, and Dell — and per Bloomberg, OpenAI is set to deploy at scale in Q3. The CoreWeave number is the one to sit with: ten times the token output on the NVL72 racks versus the prior generation. In inference-density terms, that's a generational jump, way beyond a routine refresh. Ten times the tokens means ten times the load, Sarah. Those NVL72 racks are liquid-cooled for a reason: you can't pack that much silicon in and get the heat out any other way. That's water and power, on a meter, somewhere. And OpenAI's scale-up in Q3 lands right as FERC's registration clock starts ticking. The capex ramp and the compliance calendar are about to collide. The Vera CPU's the quiet piece here — Nvidia built its own to stop buying processors, then spent the briefing benchmarking it against AMD's Turin. Twice as fast on Python, they say. Vendor math, sure, but they said it in the room. This one's from Bisnow:
The investment giant executed roughly $57B in data center transactions by financing Meta’s massive Texas campus and closing its acquisition of Aligned Data Centers. BlackRock is leading a $12B debt sale to fund its data center project for Meta in El Paso, Texas, The Wall Street Journal reported Tuesday.
BlackRock pushed roughly $57 billion in data center financing through in a single week. The headline number is a $12 billion debt sale to fund the Meta campus in El Paso, with JPMorgan and Morgan Stanley leading the deal. And read the ownership split before you clap — it's an 80/20 joint venture, and BlackRock holds the 80. Meta keeps 20 percent and a light balance sheet; BlackRock owns the asset. El Paso is close to a gigawatt of capacity. This is the same El Paso that put ratepayer language into its framework — so the city wrote the ratepayer in, and now the capital structure's an 80/20 with a $12 billion debt stack on top. There's the tension right there. The city puts the ratepayer in the document; the financing writes them back out. And this is debt — not equity, not Meta's cash — so somebody's holding bonds against a gigawatt of load that still has to clear a grid. It mirrors the Blue Owl structure in Richland Parish. That one was $27 billion in debt; BlackRock took more than $3 billion of the bonds. Same playbook, El Paso edition. So BlackRock's on both sides — financing the load through GIP, and with the Aligned close, owning the infrastructure the load runs on. Who eats it if part of this pipeline doesn't energize? Not the guy with 20 percent. KEIRSTEN BRAGER, writing in AMPYX CYBER:
FERC did not accelerate NERC's timeline, it made the calendar a directive and moved registration onto the critical path. The order creates the computational load entity, sets the December 31 and March 1 deadlines, and leaves the AI Reliability Boundary, the line between what the grid must command and what it leaves to private contract, for Phase II to draw.
FERC Order RD26-7-000 was issued on the Commission's own motion. It creates a new registration category — the computational load entity — and hangs two dates on it: December 31 for the standards and registry criteria, March 1 for the Phase II workplan. All week, we've watched regulators name the problem. Today, FERC put it on a clock. The escalation is the docket number and the calendar, not another memo about urgent need. Good. December 31, March 1 — those are dates you can miss, and missing them costs something. That's what I've wanted since this beat was still county zoning fights. But here's what I'm chewing on. Registration is a category. Does it actually touch the 12 gigawatts stuck in PJM's queue, or does it just run alongside it — everybody fills out a form, nobody clears? The order punts that part. Brager at AMPYX flags the AI Reliability Boundary — the line between what the grid can command and what stays in private contract — and that line doesn't get drawn until Phase II. So you register in December, and your actual obligation is still undefined. So the exposure survives the deadline. You're on the registry, you're on the hook — for what, we find out later. And this is landing in the same quarter Vera Rubin hits full production, per the Nvidia piece we just ran. Keep two separate dockets straight: RD26-7-000 is not the RM26-4 show-cause proceeding. Different orders, overlapping timelines. Don't conflate them. This one's from The Vindicator:
Citing public opposition, the Ohio Environmental Protection Agency rejected a request for a blanket permit for data centers to discharge wastewater and decided to consider each project individually. In a Tuesday ruling, the state EPA wrote: “After carefully reviewing the significant volume of public comments received on the draft National Pollutant Discharge Elimination System general permit for data centers, Ohio EPA has decided not to move forward with finalizing the general permit”
Ohio EPA had a blanket wastewater permit teed up — dump-and-go for every data center in the state. Seven thousand public comments later, they killed it. Project-by-project review instead. So the local backlash just took a wastewater turn: regulators dropped the one-size-fits-all path. In Tuesday's ruling, it's individual NPDES permits, with the same standards applied one site at a time. And this is exactly the accountability the Thane residents were screaming for and never got. A regulator here actually said, no rubber stamp; each project gets a real look. The catch buried in that decision: individual review under the same rigorous standards means every project now eats its own timeline. Nobody's discharging cooling water on a general permit anymore. Daily Maverick, with Lindsey Schutters:
This doesn’t completely absolve Equinix and the City of Cape Town of blame for the intricate square dance they did when unlocking the former golf course grounds for the current industrial revolution, but we can give them a break on some of the environmental fronts.
A golf course that drinks three million litres a day — the King David club in Cape Town — is getting turned into two hyperscale data centers. And the Daily Maverick's whole pitch is: relax, the data centers use less water than the fairways did. Which, if you just do the arithmetic, might actually be true. Three million litres a day is 60,000 people under Cape Town's Level 6B Day Zero restrictions. Sure, but 'better than a golf course' is a low bar in a city that literally counted down to zero water. Compare it with that same water going to Parow and Bellville, not with a fairway. And notice what's missing from the excerpt — no megawatt figure, no in-service date, no named operator for the two facilities. The phrase is 'poised to welcome,' which is a lot softer than a signed filing. Right after the Ohio EPA piece we just cued up, where a regulator said it'll look at each data center's wastewater one by one. Cape Town's got Day Zero in its recent memory — I'd want that same project-by-project stare here, not a golf-course comparison in a business column. If you track the infrastructure behind the AI boom, try AI IPO Watch: daily coverage of OpenAI, Anthropic, Databricks, and SpaceX going public, with every filing, valuation, and first trade sourced, not rumored. Find it wherever you listen to podcasts.
Next, we're watching NERC's computational-load reliability standards, glossary definitions, and registry criteria; they're due to FERC by December 31, 2026. The Phase II workplan for computational-load integration is due March 1, 2027. And OpenAI plans to adopt Nvidia Vera Rubin at scale during the third quarter.
Links to every story we covered today are in the show notes, if you want to dig into anything that stood out. That's The Data Center Daily for today. This is a Lantern Podcast.