Who Foots The Grid Bill For AI's Power Habit?

Today on Concrete Compute: the House just voted four hundred seventeen to three to stop utilities from sticking you with the bill for AI's power habit — so why did the Senate kill the fast track less than a day later? Before that, in the headlines: an AI infrastructure builder just closed a raise that values it north of thirty billion dollars, OpenAI and Anthropic are apparently hunting for much smaller data center deals, Oracle locked down over a gigawatt and a half of Texas wind, and Nvidia and Google are teaming up to make data centers play nice with the grid. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Friday, September 18th, 2026. Let's get into it.

First up, money. Crusoe, the AI infrastructure builder, just closed the initial round of a three-point-nine-billion-dollar Series F, and according to Crusoe's own newsroom, that values the company at thirty point nine billion dollars post-money. Crusoe says it now has more than a hundred and forty billion dollars in what it calls total contracted value, and more than six gigawatts of gross contracted capacity — gigawatts under contract, not gigawatts running. That gap matters, because Crusoe itself says only about one gigawatt of that six-plus is actually delivered and operational today; the rest is commitments waiting on power and construction. Crusoe's own newsroom describes an energy-first strategy behind that number — in-house power plant development paired with partnerships spanning grid, battery, nuclear, thermal and renewable providers. Now, what's interesting is where the new money's headed: not just more giant campuses, but smaller modular units the company calls Crusoe Spark, faster to deploy than a gigawatt-scale build. My read: that's the same instinct you're about to hear from Anthropic and OpenAI — when the giant sites take years to power up, going smaller and faster starts looking a lot more attractive to everyone in this business.

Sticking with that theme of speed over scale — CNBC reports that Anthropic and OpenAI, the two labs that just spent the past year signing multi-hundred-megawatt and gigawatt-scale megadeals, are now also hunting for much smaller data center deals, twenty to thirty megawatts, in the UK, the Nordics, and possibly here at home. So why go small now? An analyst told CNBC it's about speed to usable capacity — you can get a twenty-megawatt site energized while you're still waiting years for a single giant power block, especially with land and power growing scarce in Europe and pushback mounting here at home. Now, I want to flag something: this is sourced to people familiar with the conversations, not confirmed contracts — OpenAI told CNBC it doesn't comment on commercial talks, and Anthropic didn't comment at all. So treat this as exploratory for now, not signed paper. But directionally, it lines up with what you just heard from Crusoe: the frontier labs are hedging their bets on getting compute online, because one giant campus at a time isn't fast enough anymore. And it tells you something about where the bottleneck actually sits — it's not that these companies can't find compute demand, it's that they can't get enough of it energized fast enough on any single site, mega or otherwise.

Next, Oracle. DCD reports Oracle has signed wind energy supply agreements across ten Texas projects, adding up to more than one point seven gigawatts of ERCOT-delivered capacity — ERCOT being the grid operator for most of Texas. Suppliers include Clearway Energy, Engie, RWE and Scout Clean Energy; the Engie piece alone covers up to five hundred sixty-eight megawatts. Oracle says the power is expected to support its Abilene AI campus, and frames the deals as progress toward matching all of its AI data center electricity use with carbon-free power by twenty thirty-five. Worth being precise here: these are signed supply agreements for wind, which is intermittent — it blows when it blows. So 'carbon-free matching' is an annual accounting exercise, not Oracle actually running on windmills at three in the morning. Real contracts, real megawatts — just not the same thing as firm, round-the-clock power, and that distinction matters if you're the one relying on the grid staying up.

Last headline before today's main event: Nvidia, Google, and a startup called Emerald AI have launched a new coalition — Quartz reports it's the AI Energy Management Alliance, a twenty-member coalition — built around getting data centers to flex their power use. The idea: instead of a data center demanding a fixed, guaranteed block of power around the clock, it agrees to throttle back during rare peak-demand moments — shifting workloads, tapping batteries, running on-site generation — in exchange for faster interconnection, meaning a quicker hookup to the grid. Emerald AI's CEO makes the case that new data centers can wait a decade or more to connect because utilities have to reserve capacity for those rare peak moments. Google separately committed one gigawatt of power demand it says it can reduce on request through utility agreements nationwide. Now, this is a voluntary coalition pitching regulators and utilities on a new way of doing business — it's not a rule anybody's adopted yet, and its claim about unlocking a hundred gigawatts of capacity depends entirely on utilities actually saying yes. Still, notice the through-line: Crusoe going modular, the labs chasing smaller deals, Nvidia and Google asking for flexibility instead of guaranteed megawatts — everybody in this industry is trying to solve the same problem, getting power fast. And that brings us to who's actually going to pay for all of it.

Our main story today: call it who pays for the plug — the fight over whether ratepayers, meaning you and your neighbors on the regular residential rate, get stuck covering the grid upgrades AI data centers need. On Wednesday, the House passed the Ratepayer Protection Act, four hundred seventeen to three. That's about as close to unanimous as Congress gets these days, and it's the first data-center bill this Congress has actually passed. So what does it do? At its core, the bill tells state utility regulators to 'consider' adopting standards for big electricity users — data centers pulling a hundred megawatts or more — that would make those users cover the 'full, incremental' cost of any generation, transmission or distribution upgrade needed to serve them. In plain terms, if a data center forces a utility to build new power plants or wires, the data center pays for that, not the family down the street. The bill also requires large loads to put up financial assurances before upgrades happen, and to guarantee cost recovery if they bail on a power contract early. Now, here's the important qualifier, and it's right there in the bill's own language: it tells states to 'consider' these standards. It doesn't mandate them. And that word is doing a lot of work — 'consider' is a suggestion, not a command, which leaves states holding the actual decision no matter what Congress says here. The only three no votes came from progressive Democrats: Summer Lee of Pennsylvania, Delia Ramirez of Illinois, and Rashida Tlaib of Michigan. Their objection wasn't that the bill goes too far — it's that 'consider' isn't 'require.' Then, a day later, the story got messier. CNBC reports Senate Republicans tried to fast-track a companion bill through unanimous consent, meaning it would pass instantly unless a single senator objected — and Senate Democrats blocked it Thursday, arguing the framework leans on voluntary commitments from states and developers rather than binding rules. So where does that leave things for you? Not dead — a failed unanimous-consent push just closes the fast lane, it doesn't kill the bill — but it's stuck, with the Senate's calendar getting tighter as midterms approach in November.

So how real is this? Let's hear from the people actually in the room. Republican Rep. Gabe Evans, R-Colo., the bill's co-sponsor, said in a press release that the bill, quote, 'ensures large data centers pay for the infrastructure they require while giving states the flexibility to determine what works best for their communities,' end quote. That's the sales pitch — protection with local control. Not everyone's satisfied that local control is the right frame. The Center for Data Innovation argued the bill should be broader, not stricter. Michelle Lopes Maldonado, the group's associate director of AI policy, said in a statement, quote, 'Congress should apply the standard uniformly to any customer at a hundred megawatts or more, establishing a separate rate class for these large electricity users. This preserves the cost-recovery principle while eliminating sectoral favoritism and ensuring residential users are not left paying the bill,' end quote. Her point: why single out data centers when any hundred-megawatt factory or crypto mine creates the same cost-shifting problem for you? Then there's the research shop actually paid to grade bills like this for investors. ClearView Energy Partners, in a client note, called the House bill, quote, 'somewhat behind the regulatory curve,' end quote, and said it would, quote, 'largely reinforce' end quote a shift already happening at the state level. That's worth sitting with: according to a database from the Smart Electric Power Alliance, only thirteen states currently lack any utility tariff setting requirements for data centers and other large loads, and at least three of those thirteen were already considering proposed requirements as of July. Most of the country wasn't waiting on Congress. Now, on the Senate side, Senator Martin Heinrich, D-N.M., has his own alternative, the GRID Savings Act, which would have FERC — the federal energy regulator — set rules on large-load interconnection, including cost allocation and financial security requirements. According to Politico, Heinrich has said the House bill doesn't go far enough. Here's my read, and it's genuinely mixed. I like that Congress moved at all — four hundred seventeen to three tells you the politics here have converged, nobody in either party wants to be caught defending higher power bills for a data center three towns over. That's the beneficiary-pays principle, the idea that whoever creates the cost should pay it, picking up real bipartisan legs. But 'consider adopting' is a suggestion wearing a mandate's clothes, and ClearView's right that it's mostly catching up to what states were already doing. The one genuinely new wrinkle is folding in power-supply costs, not just transmission and distribution — ClearView flagged that interconnection agreements typically only cover the wires, not the power plant itself, so if this actually got enforced with teeth, it could force data centers to build or contract their own new generation before they connect, which would slow construction down, not speed it up. That's a real trade-off nobody's talking about loudly enough. Time for the Hype Check. I'm putting this one at a four out of ten on substance: the vote count is real and the beneficiary-pays principle getting near-unanimous House support is genuinely notable, but the operative word is 'consider,' the Senate companion just stalled, and most states are already doing the work this bill claims credit for. If the Senate lets this die quietly before midterms, the states keep writing their own rules — which means whether your bill goes up or down still depends entirely on which state you live in, not on Washington.

If today helped you make sense of who's actually paying for the AI buildout, follow Concrete Compute wherever you listen — new episodes land every weekday morning. 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!

I also host Quickly Quantum: a daily quantum computing briefing you don't need a physics degree to follow. The breakthroughs, the funding rounds, and how much substance is really under each claim. Find it wherever you get your podcasts.