The grid finally handed us a number for what all these permit fights actually cost: sixty percent. If you're just joining us, PJM's capacity market was flashing red long before today. The auction has missed its supply target three years running, with the clearing price pinned at the $325-per-MW-day cap. The stress traces back to fast-rising load — AI data centers especially — and a generation pipeline that can't interconnect fast enough. This is The Data Center Daily. Today we're starting in New York, where 50 megawatts is the line; Texas is arguing over who pays; and Google says it has the biggest solar deal in the country. We'll see whose PPA is actually signed. This one's from Bloomberg:
PJM Interconnection LLC, which serves 13 states and Washington, DC, said Tuesday that its auction to procure power for the year starting June 2028 tied a $16.4 billion record set in late 2025. Data centers accounted for roughly $6.3 billion of that total, said Joseph Bowring, president of Monitoring Analytics, the grid’s independent market monitor.
PJM's 2028 auction cleared at a record-tying $16.4 billion, and the market monitor — Joe Bowring at Monitoring Analytics — puts $6.3 billion of that squarely on data centers. Six-point-three billion. Add the three auctions before it, and Bowring says data centers have loaded almost thirty billion onto PJM ratepayers. That's the invoice landing. And it still fell 6.8 gigawatts short of reliability. Third straight auction missing the target. Record cost, and it still didn't buy enough power. Bowring's fix is the part I'd underline: run a separate auction just for data centers, so the guy in Ohio who never asked for a server hall stops eating the difference. The watchdog is saying the quiet part out loud. Which is notable, because vendors keep telling us efficiency solves this. A monitor with a thirty-billion-dollar tab isn't persuaded by performance-per-watt slides. From Markus Kasanmascheff at WinBuzzer:
New York Governor Kathy Hochul has signed an order to halt incomplete environmental permits for new data centers using 50 megawatts or more, a cutoff aimed at facilities that can impose large water demands, for up to one year. The pause leaves already-approved developments outside its scope.
New York's put a number on the pause: 50 megawatts and up, incomplete permits frozen for up to a year while regulators write grid, water, air, and community rules. It's the floor number we didn't have when Executive Order 62 dropped. And here's why it matters: 12 gigawatts of large load was sitting in the interconnection queue as of May. New York just told a chunk of that to stop and prove it first. Note what's exempt: already-approved projects and smaller institutional sites. The freeze hits incomplete permits only; there's no claw-back. The stuff in the ground keeps going. Put it next to the 60% cost figure we just hit, and it's the other end of the same problem. PJM prices the gap; New York slams the door on new entrants while it writes the rulebook. And water quality is one of the four areas they're now required to define. Somebody's finally going to do the two-billion-gallon arithmetic before the permit, not after. Mark Segal over at ESG Today has the details. Google's signed what ESG Today is calling the largest solar project in the US. And it lands the same morning PJM prints the 60% supply-cost jump we opened with. Google's buying itself out of the grid-cost exposure the rest of the queue is about to eat. Largest in the US — great headline. Is the PPA signed and power flowing, or is this contracted-someday capacity dressed up for a press cycle? ESG Today says signed. What I don't see is an in-service date or the term. Announced and energized are different animals. Right — and Google self-insuring against a 60% cost bump is fine for Google. The ratepayer in PJM territory doesn't get to sign a solar deal to opt out of that invoice. Here's Matt Boms at Texas Energy and Power:
The scale explains why. ERCOT estimates up to 110 gigawatts of new large loads could seek to connect over the next five years, more than double today’s system peak of about 86 gigawatts. Before any of that generation arrives, about $37 billion in transmission costs are already baked into the system, pushing rates up roughly 3.5 percent a year for every customer.
After that Bloomberg number — PJM supply costs up 60% — Texas is going the other direction. The Texas Energy Buyers Alliance walked in and said, charge us. Tie the transmission bill to our approved capacity. And the scale is why. ERCOT's modeling up to 110 gigawatts of new large-load requests over five years — the system peak today is about 86. You're talking more than double, before a single new generator shows up. And there's already $37 billion in transmission baked in, pushing rates up 3.5% a year for everyone — grandma included. The buyers group is at least saying the new demand should carry the new bill. One state freezes the door shut at 50 megawatts; another tries to price the door correctly. Same week, opposite instincts. I'll take the tariff over the padlock. It's the statewide version of that Grimes County fight — write the rulebook and hand out the incentive check in the same room. Only here, somebody actually proposed who pays. DataCenterNews Canada writes:
According to NVIDIA, its GB300 NVL72 rack-scale system delivers up to 25 times the performance per watt of the Hopper generation on DeepSeek V4 Pro, up to 20 times on GLM5.1 and up to 10 times on Kimi K2.6. It attributed those gains to the move from an eight-GPU domain in Hopper systems to a 72-GPU rack-scale design in Blackwell NVL72.
NVIDIA says its GB300 NVL72 rack does up to 25 times the performance per watt of Hopper on DeepSeek V4 Pro. Twenty on GLM5.1, ten on Kimi K2.6. Nice benchmarks. The whole pitch is 'fixed power budgets' — serve frontier models without blowing past electricity and cooling limits. Vendor touts power discipline the same week PJM's watchdog prints a 60% cost bump. Read the room. Twenty-five times per watt against Hopper — sure. But per watt of what? A rack that pulls more total power than the eight-GPU box it replaced. The efficiency gain doesn't shrink the interconnection request. 'Fixed power budgets.' Man. Somewhere a site's running turbines with no air permit and NVIDIA's out here selling discipline slides. Put a signed PPA or an in-service date behind the number and I'll listen. The giveaway for me: even NVIDIA says a single figure can't capture inference trade-offs. Low latency versus throughput versus cost. So 'up to 25x' is the best-case corner; it doesn't describe the operating reality. If you follow infrastructure, you'll probably want the software side too. Check out AI Daily Briefing: top AI news for engineers, founders, and investors, every weekday. We separate real capability from demo hype, fast. Find it wherever you listen to podcasts.
Links to every story we covered today are in the show notes, so if one caught your attention, you can dig into the original reporting there. That's The Data Center Daily for today. This is a Lantern Podcast.