Oracle just told the SEC its AI bottleneck isn't only GPUs — it's memory, components, and the cost of sitting on capacity that isn't earning a dime yet. If you're just joining: hyperscale AI has been moving from big capex charts to named sites — Stargate's $500 billion, 10-gigawatt U.S. plan, with OpenAI, Oracle, and SoftBank tied to new data centers alongside Abilene. The question all along has been simple: can compute, power, and physical capacity actually show up on the same schedule? This is The Data Center Daily. Today — a filing that finally names the stranded-asset cost, and state regulators stepping into the FERC interconnection fight. Sarah, start with Oracle. We're staying on Hyperscale capex and capacity expansion — follow the show and you won't miss what comes next. Buried in Oracle's latest 10-K is a supply-chain disclosure that goes further than what most hyperscalers have said publicly — and it's not just about GPUs. Listen to how they frame the exposure.
For example, industry supply capacity for AI accelerators, including graphics processing units, as well as memory devices, is competitive, and we at times have to accept less favorable terms with suppliers to minimize supply constraints.
What jumps out to me is Oracle explicitly tying memory supply to contractual risk — meaning if the HBM or the accelerators don't land on time, they're sitting on powered, leased capacity that's bleeding cost with zero revenue offset, and maybe penalty clauses on top of it. The number to watch here isn't just Oracle's GPU allocation — it's whether memory and critical-component delivery are actually lined up with the data center bring-up schedule, because that gap is where the financial exposure lives. Here's National Association of Regulatory Utility Commissioners:
NARUC appreciates the ANOPR disclaiming any intention to assert jurisdiction over distribution interconnections, generation facilities, and retail sales. It is imperative that FERC, in any final rulemaking, make clear that it is affirmatively not asserting jurisdiction over these facilities, or end-use sales which fall squarely within the exclusive jurisdiction of state energy regulatory authorities.
NARUC's initial comments are in the docket — RM26-4-000, the large-load interconnection rulemaking I flagged as this week's structural event. So FERC didn't slam a final rule on the table by the end of June. The state commissioners are on record, which means we're still in comment phase. And NARUC's whole pitch is a fence. They want FERC to affirmatively disclaim jurisdiction over distribution, generation, retail sales — all the stuff the states already control. Right. They credit the ANOPR for not reaching into retail, then spend the filing making sure FERC says it out loud in the final rule. That's a state regulator who's seen this movie before. Here's what it tells me — when a data center load lands, who eats the generating cost is still a state call. NARUC is protecting the ratepayer's seat at the table before FERC writes the queue rules over their heads. That connects directly to the Oracle filing we just hit. The stranded-capacity costs Oracle disclosed don't disappear because FERC writes an interconnection rule — somebody at the state level still decides how those costs get assigned to retail load. Exactly. Oracle names operating costs on capacity that isn't earning yet. NARUC's saying: not on the ratepayer's tab without a state proceeding. Read those two filings back to back and the stranded-asset fight comes into focus. From Mayer Brown:
In late October 2025, at the direction of the Secretary of the US Department of Energy (“DOE”), the Federal Energy Regulatory Commission (“FERC” or the “Commission”) opened a preliminary rulemaking proceeding with the potential to result in significant changes to how data centers, advanced manufacturing facilities, and other energy-intensive projects are interconnected to the US bulk electric transmission grid.
Mayer Brown's read on FERC's preliminary large-load rulemaking — docket RM26-4-000 — is out, and it's the lawyers' answer to the thing I kept circling: does this actually define how a large load gets counted, or does it punt to the interconnection agreements? Remember how this started — the DOE Secretary used a Section 403 directive back in October to force FERC's hand. FERC didn't wake up one morning curious about data centers. Somebody handed them a mandate. And here's the part I care about — the NARUC comments we just hit mean the state commissioners are formally on the record in this same docket. So now we get to see whether FERC's standardized procedures protect the ratepayer or just grease the queue for the applicant with the biggest load. Right, and the timeline question is live. PJM's been quoting three to four years to connect in the zones where the demand actually is. The whole point of a federal standard is to compress that — Mayer Brown's takeaways are where you'd see if it does, or if the number survives untouched. My bet is that number survives untouched. You can standardize the paperwork all you want — it doesn't pour concrete on a substation any faster. This one's from Synergy Research Group:
New Q3 data from Synergy Research Group shows that the number of large data centers operated by hyperscale providers has climbed to 1,297 worldwide, nearly tripling since early 2018. Over the same period, total operational capacity has increased more than fourfold as average data center size continues to rise.
Synergy's Q3 numbers are in: 1,297 operational hyperscale facilities worldwide, nearly tripled since early 2018, capacity up more than fourfold. That's the operational count — built, energized, running. And the U.S. share of that operational capacity climbed to 55%, up from 52% three years ago. Quarterly capex hit $142 billion in Q3 alone, per Synergy. $142 billion in a single quarter. That's the spend. But Synergy's counting what's plugged in — none of those 1,297 are the 10-plus gigawatts of load sitting in the PJM and ERCOT queues, waiting three, four years to connect. So you've got a fourfold operational curve on one side and a queue backlog on the other, and right between them is where Oracle's stranded-capacity clause we just hit lives. The buildout's real. The connection isn't keeping up. Right — Synergy gives us the operational denominator. The five Stargate dots on a map don't move this number until they're energized. This is what energized actually looks like at scale. If The Data Center Daily helps you track the infrastructure behind AI, try Anthropic Pentagon Watch — a daily briefing on Anthropic’s fight with the DoD over Claude, military AI use, autonomous weapons, and AI procurement blacklisting. Find it wherever you listen to podcasts.
You’ll find links to every story we covered today in the show notes, so if one caught your ear, you can dig into the original reporting there. That’s The Data Center Daily for today. This is a Lantern Podcast.