AI compute has two new choke points: the rulebook and the neighborhood. Here’s how we got here: Equinix has laid out a $5 billion to $7 billion annual data-center investment ramp, aimed at markets where it already operates and sees interconnection density and customer demand. The company is tying that spend to enterprise hybrid-cloud and AI demand, while acknowledging that power, labor, and community constraints can slow even well-capitalized buildouts. This is The Data Center Daily. A regulator is moving while factories stall and local approvals get shaky. Let’s start with the rule that doesn’t exist yet. We're staying with this story: Equinix AI buildout capex ramp. Follow the show and you won't miss what comes next. Mara Voss, writing in Btw:
The preliminary vote matters because it narrows the chance that large computational loads will stay outside the reliability regime. It does not settle the text, approve a rule or start a compliance clock. Developers should distinguish those milestones before turning a standards signal into a capital budget.
NERC’s September 19 result is only preliminary. The ballot still needs validation and a comment report, followed by FERC approval, an effective date, and registration decisions. Nobody’s on a compliance clock yet. And yet some CFO is absolutely putting “NERC standards passed” on a capital slide. Slow down. A preliminary ballot won’t get a data center interconnected or operating, and it doesn’t waive reliability obligations that aren’t final. The change is narrower: large computational loads now look less likely to remain outside NERC’s reliability perimeter. Developers can keep design flexibility, but they shouldn’t bake recurring compliance costs into the base case as if the rule already exists. Exactly. Start keeping better operating evidence. Don’t call a regulatory hurdle cleared. The grid doesn’t accept PowerPoint compliance. Latitude Media writes:
Fluence revised its 2026 fiscal year guidance down for a second time last week, citing ongoing supply chain issues that the company expects will pull its annual revenue down by around $1 billion from original estimates — and that analysts worry could hamper its burgeoning work with data centers.
Fluence planned to make 11 Gridstack Pro units a day in August and September. Houston made fewer than one a day in August, then roughly three in early September. That’s a factory-ramp miss, not a rounding error. And this is the domestic storage backstop behind all those “clean, reliable” data-center power plans? The automated welder fell over, so they’re welding by hand, hiring more people, inspecting more, and eating $56 million in late penalties. Fluence says Houston, with Bergstrom, can eventually produce 15 gigawatt-hours a year. But “eventually” doesn’t help this year’s schedule. For the current fiscal year, the delayed ramp accounts for much of a roughly $1 billion revenue cut. A battery container on a vendor roadmap doesn’t help a data hall. Until that plant is clearing one unit a day by a wide margin, operators should treat storage delivery dates like a weather forecast: useful, but don’t build your reliability plan around them. Andrew Larson, writing in Hartford Business Journal:
No hyperscale data center has been formally proposed in Connecticut. None has come before a local land-use board, and none has signed an agreement to connect to the electric grid. But data center developers are interested in Connecticut and have been quietly exploring whether such large-scale projects could work here.
More than 50 Eversource study requests, about half tied to Connecticut—and zero signed grid agreements. Developers are paying for feasibility studies, which is a very long way from putting 100 megawatts on the system. Eversource says the projects currently under study add up to just over 1,000 megawatts. At round-the-clock operation, Jacob Lucas puts that at roughly a million typical Connecticut homes’ worth of electricity—while no project has even reached a land-use board. Two years ago, Eversource had received zero transmission-level study requests for 100-megawatt-plus data centers. Now Connecticut is its hottest territory. Fine—study away. But a paid study is not a substation, a transmission upgrade, or a customer that’s signed up to cover the bill. After the NERC ballot we just covered, Connecticut makes the local point clear: a large computational load can be under study long before anyone has approved, contracted, or connected the project. Lindsey Schutters, writing in Daily Maverick:
The Municipal Planning Tribunal’s (MPT’s) 14 July approval for the Equinix data centres has been legally suspended and cannot be acted upon after a formal internal appeal was lodged on 6 August by the Housing Assembly social movement and the UK non-profit Foxglove (represented by the Legal Resources Centre).
Back to yesterday’s Equinix buildout, with a permitting wrinkle: Cape Town’s July approval is suspended pending appeal. The Municipal Planning Tribunal’s July 14 decision can’t be acted on while the Housing Assembly and Foxglove appeal is live. And Equinix says the 327,000-square-meter site is a long-term land bank with no immediate development plans. Fine—but the rezoning and bulk allocation stay attached to the land. That’s a useful head start while everyone else gets told to join the queue. Cape Town is now drafting refined guidelines for large data-center applications. On the capacity ledger, approved-but-frozen should sit between a controlled site and an actual build—especially when the developer hasn’t submitted a planning application. The appeal came from a social movement and a nonprofit, not a rival developer with a better lawyer. Communities have figured out that the planning process can stop a project after the victory lap. Latitude Media writes:
So Rune, which last week announced a $40-million Series A led by Spark Capital, designed a modular data center attached to a power electronics block. The system plugs directly into utility-scale projects — without expensive substation upgrades, transformers that are in short supply, and yearslong interconnection queues.
Texas threw away nearly 10,000 gigawatt-hours of renewables last year, mostly because the wires couldn’t move it. Rune sees that and says: park a GPU box at the solar farm, behind the inverter, and eat the curtailment. RELIC is an interesting physical workaround: direct connection at the project, custom power electronics, liquid cooling. Rune says it can energize in six weeks. That bypasses the substation-upgrade queue we just heard is holding up conventional load in Connecticut. But keep the ledger honest. Rune raised $40 million and has roughly 1 megawatt deployed; the hundreds of megawatts in its pipeline are still a claim. Curtailment is real, and the 10,000 GWh is real. The hard part is whether intermittent spare solar can support customers who want compute on demand. And it arrives as Fluence struggles to ramp storage output. If batteries can’t absorb surplus generation at scale, flexible compute beside the generation gets more compelling. It doesn’t, however, turn a curtailed-hour resource into a 24/7 data-center power contract. If you follow data centers, you might also like AI Daily Briefing: top AI news for engineers, founders, and investors, with real capabilities versus demo hype explained fast every weekday. Find it wherever you listen to podcasts.
We’re watching for NERC’s validation of the preliminary ballot and its report of comments on the computational-load standards. Links to every story are in the show notes, so check out the pieces you’d like to explore further. Thanks for listening, and we’ll be back tomorrow. That’s The Data Center Daily for today. This is a Lantern Podcast.