PJM's New Rule: Bring Power or Be Cut First

Today on Concrete Compute: PJM just told data centers waiting to plug in exactly what happens if they show up without their own power in tow — get in line, and if the grid gets tight, you're the first one that goes dark. Before that, in the headlines: the Cherokee Nation bans hyperscale data centers on its tribal land, Amazon says it's exploring turning its private Texas gas plant into a real grid resource, and Groq raises three hundred fifty million dollars at half its peak valuation — from an investor list that now includes the company that dismantled its leadership team. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Tuesday, August 18, 2026. Let's get into it — because today is one of those days where three different fights, across three different agencies, are really just one fight: who pays for the grid.

First up — picture your utility bill quietly padded by a grid operator's own math error, and multiply that across an entire region. That's what SemiAnalysis founder Dylan Patel argues happened, to the tune of twelve billion dollars, in a post on X this week about PJM's grid modeling. His argument, in his own words from that thread: 'Incompetence in grid modeling is causing US ratepayers to pay twelve billion dollars more than necessary,' and he says the auction system PJM uses to buy future power capacity is actively discouraging anyone from building more of it. Patel's call is for market-based reform — not just new rules aimed at data centers, but a rethink of how the whole capacity market is designed. Now, I want to be straight about sourcing: this twelve-billion number is Patel's own claim from a single thread, and we haven't independently verified it against a PJM or FERC record. But the timing is not nothing — this lands the same week PJM asked FERC for new authority over how data centers connect, and the same week FERC signed off on a separate, PJM-footprint transmission plan worth hundreds of millions more. Different filings, same question hanging over all of them: who's actually checking PJM's math?

No competitive bid, no open process — just a straight-up assignment. That's how one utility landed a nine-hundred-and-four-million-dollar transmission project, after FERC approved a cost allocation framework Friday for transmission that Midwest grid operator MISO wanted built, but which sits physically inside PJM's territory. Exelon's Commonwealth Edison gets that nine-hundred-four-million-dollar slice; Duke Ohio picks up a much smaller five-point-three-million-dollar piece. FERC explicitly rejected calls to put any of it out for competitive bid — the process critics say tends to produce lower prices. Two FERC commissioners, Judy Chang and David Rosner, defended the decision in a joint statement, arguing that transmission crossing grid-operator boundaries is too valuable to get stuck behind outdated tariff definitions that never anticipated this much interregional need. That's the story underneath the story: who gets to build the wires America's data-center boom requires, whether ratepayers get the benefit of competition, and who just eats the bill because the tariff language never planned for this.

If it ever ran flat out, this would rank as one of the largest single sources of pollution in the country — that's the ceiling on Amazon's new private power plant in Pecos County, Texas, feeding an eight-thousand-acre AI campus. The state's preliminary permit describes thirty-five gas turbines and five thousand megawatts of nominal output, capable of emitting up to thirty-three million tons of carbon dioxide equivalent a year at that maximum — though that's a ceiling assuming nonstop operation all year, not a forecast of what actually gets burned. What's new: Amazon now says it's 'actively exploring' turning that private plant into a front-of-the-meter resource — eventually connecting it to the real Texas grid so other customers could buy power from it too. No timeline, and Amazon hasn't said if it's filed anything with ERCOT yet. And here's a wrinkle nobody's explained: developer Pacifico Energy separately describes the broader project at up to seven-point-six-five gigawatts, which doesn't obviously square with the five-gigawatt, thirty-five-turbine plant the state actually permitted. University of Texas researcher Joshua Rhodes told Data Center Knowledge he couldn't yet say how that existing generation would affect a future ERCOT connection — 'this part is still unclear,' he said — and separately noted every turbine Amazon buys here is one some other project down the road doesn't get.

The Cherokee Nation has banned hyperscale data center development on land it owns or holds in trust. Principal Chief Chuck Hoskin Junior announced the policy after a task force surveyed nearly sixteen hundred Cherokee Nation citizens and found sixty-four percent opposed hyperscale data centers, against just fourteen percent in favor. Hoskin's own words: these projects, quote, 'affect our resources and cultural lifeways,' end quote, and the tribe won't back them on the reservation without what he called robust consultation. Here's the nuance: the ban itself only covers tribally owned and trust land — across the rest of the reservation, in northeastern Oklahoma, the tribe can demand consultation but can't block a project outright. And two are already moving inside that footprint: a billion-dollar campus in Tulsa County called Project Clydesdale, and another, Project Mustang, near Claremore. The Cherokee Nation joins Oklahoma's Seminole Nation and the Kickapoo Tribe, who've already banned this kind of development, while the Muscogee Nation is still studying the question. Add it to the growing list of host communities deciding the honest answer is no.

Half of what it was worth eleven months ago — that's where Groq's new funding round lands the company. The AI inference chipmaker just raised three hundred fifty million dollars at a three-and-a-half-billion-dollar valuation, down from six-point-nine billion last September. Leading the round: Disruptive, the same firm that led Groq's prior raise — and joining in with 'planned participation' is Nvidia, the company that licensed all of Groq's chip technology for twenty billion dollars back on Christmas Eve and hired away Groq's founder and CEO in the process. A company spokesperson insists this isn't a down round, just a fresh valuation for what's now a genuinely different business — Groq's pivoted from building rival chips to running an Nvidia-powered inference cloud, and it's aiming to scale from fifty-four megawatts of data center capacity to over two hundred by 2027. My honest take: call the framing whatever you like, the math is the math — the valuation's been cut in half, and the company that hollowed out Groq's leadership and built a competing product is now writing checks into the survivor. That's a strange relationship, and how it evolves is worth watching closely.

Our main story today: call it Last In, First Cut. PJM filed a proposal with FERC on August 13th that changes the deal for any new data center big enough to matter. Here's the mechanics, and stay with me because they matter. PJM wants to define a 'New Large Load' as any single site pulling fifty megawatts or more that comes online after June 1st, 2027 — facilities within a mile of each other can get bundled together and counted as one site, so campuses can't dodge the threshold by splitting into smaller-looking pieces. Under the plan, the utility serving that load has to line up qualifying new capacity — new generation or contracted capacity — equal to the load's peak demand. If they don't, the unsupported chunk of that data center's demand gets curtailed, meaning PJM can order it to cut back, before PJM even reaches for its other emergency tools, like asking regular businesses to power down under demand-response programs. In plain terms: a data center that shows up without its own power plan goes to the front of the line when the grid gets tight, not the back. Now, why is PJM doing this at all? Because the math has stopped working. PJM has projected roughly seventy gigawatts of new large load arriving by 2038, against just fifteen gigawatts of retired power plant capacity that's actually been replaced since 2022. Two consecutive PJM capacity auctions — the process that's supposed to line up enough generation years in advance — have failed to clear enough supply to meet the region's needs. This filing is PJM's attempt to keep saying yes to new data centers without pretending it can also promise them, or anyone else, that the lights stay on. There's an escape hatch built in, too, called Bring Your Own New Capacity, or BYONC — a data center can avoid curtailment risk almost entirely by lining up its own generation or storage ahead of time, and blend that with other backstop options PJM offers. And starting with the 2029-2030 delivery year, PJM says it won't even count excluded large-load demand in what it procures through its main capacity auction — that load becomes the developer's problem to solve, not the grid's. PJM also wants to stand up a Large Load Registry — basically a database tracking every big new customer's location and expected demand — and that piece is already contested. Joseph Bowring, who runs PJM's Independent Market Monitor, told Data Center Knowledge the tariff doesn't actually require data centers to back up what they report with evidence, and he wants monthly validation, not PJM's vaguer promise to update things 'periodically.' PJM's own spokesman, Jeffrey Shields, says the grid operator has already tightened its vetting of large loads. Bowring's response was blunt: quote, 'We have not seen any evidence of that,' end quote. One more thing worth flagging plainly: this is a filing, not a final rule. The Illinois Attorney General, the Maryland Public Service Commission, New Jersey's Rate Counsel, and an industrial customer coalition have all filed motions to intervene — meaning they want a seat at the table before FERC decides anything, and none of them have shown their hand yet on where they'll land. And the curtailment-first mechanism only applies to large loads connecting after June 2027 — every data center already sitting in PJM's queue today is grandfathered out of this entirely.

So does this actually make hyperscalers pay their own way? That's the question the filing exists to answer, and the honest answer is: partially, and only for the newest arrivals. Here's what I like about it, using my own yardstick for these deals — the party creating the incremental cost should be the one who pays for it. This proposal pushes that idea further than PJM has gone before. If a data center wants to skip the line and connect before there's enough generation built to serve it, fine — but the tariff now says that risk sits with the load, not with everybody else on the grid. Bring your own power, or accept you're first to go dark. That's a real shift from the arrangement ratepayers have been living under, where new demand gets connected and existing customers quietly absorb the reliability and transmission costs that follow. But — and this is a big but — the mechanism only bites for large loads connecting after June 1st, 2027. Every gigawatt of data center demand already sitting in PJM's interconnection queue today, the stuff that's actually driving those two failed capacity auctions PJM just had, skates through completely untouched. So the industrial customer coalition, the Illinois Attorney General, and the rate counsels from Maryland and New Jersey who filed to intervene this week represent something specific: everyone whose bills are still exposed to power that connected before this rule existed. Then there's the enforcement question, and this is where Joseph Bowring's fight with PJM over the Large Load Registry actually matters more than it sounds. A rule that says bring your own capacity or get curtailed is only as good as PJM's ability to verify that the capacity a data center claims to have actually exists and shows up when needed. Bowring wants monthly validation and real evidence behind every registry entry; PJM's Jeffrey Shields says vetting has already improved. Bowring's blunt response, which you heard in our setup a minute ago, tells you this isn't settled — it's an open fight between the grid operator and the one office whose entire job is to catch PJM being wrong. And remember, this is the same week Dylan Patel put a twelve-billion-dollar price tag on a separate PJM modeling failure. I'm not saying these are the same mistake — they're not — but they rhyme: a grid operator making consequential decisions off numbers that its own watchdogs say they can't fully verify. Here's a parallel I'd draw myself, not something either filing states directly: Amazon building its own gas plant behind the meter in Pecos County is voluntarily doing exactly what PJM's new rule would make mandatory for everyone else on its grid — bring your own power, don't lean on the shared system. The difference is Amazon chose that path in ERCOT, a separate grid with its own rules. PJM's proposal is an attempt to make that the default for anyone who doesn't want to build their own plant and connects after June of next year. Time for the Hype Check. I'm putting this one at a six. The mechanism is real, it's specific, and the direction — shifting reliability risk onto new load instead of onto ratepayers — is the right one by my own standard. But it doesn't touch the load that's already causing the problem, the enforcement teeth depend on a registry fight that isn't resolved, and four separate parties just filed to intervene because they don't trust the details yet. This is a serious proposal, not a settled solution. So here's how this actually plays out: if FERC approves this close to as written, the hyperscalers who've already lined up their own generation or storage under BYONC keep building on schedule, barely touched by the new rule. The ones who assumed they could connect first and sort out power later are the ones whose unsupported megawatts get curtailed ahead of every other emergency tool PJM has — that's the mechanism, spelled out in the filing itself. And the customers this rule was supposed to protect? They're still on the hook for every megawatt that connected before June of next year, because that's exactly the load this rule doesn't touch.

If tonight's episode helped you understand why your utility bill might move before a single chip gets plugged in, that's exactly what this show's for — follow Concrete Compute wherever you listen, and if you've got a friend who thinks data centers are just big boring buildings, send them this one. This has been Concrete Compute, an AI-voiced podcast, created and built by a real human using today's cutting-edge technology. Nothing you heard on this show is financial advice. I'm Brian Lampert, and I'll catch you all tomorrow — take care!