Today on Concrete Compute: Georgia just cleared a three-point-two gigawatt power contract for OpenAI's Effingham County data center — and the utility is promising it'll actually lower your bill, not raise it. That's the kind of claim I want to stress-test today. Before that, in the headlines: Nvidia may be about to buy the 'GitHub for AI' for nearly thirteen billion dollars, that supposedly hundred-twenty-billion-dollar Google-Marvell chip deal turns out to be a lot slipperier than the headlines made it sound, and data centers are becoming an actual campaign issue in states from Utah to Pennsylvania. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Friday, August 28, 2026. Let's get into it.

Let's start with a number that stopped me cold: nearly thirteen billion dollars. So why would Nvidia pay that kind of multiple for a company that barely monetizes its own platform? One catch worth flagging: neither Nvidia nor Hugging Face has confirmed this. TechCrunch reported the talks 'had not yet produced a signed agreement and could still atomize.'

Remember that hundred-twenty-billion-dollar Google-Marvell custom chip deal from a few weeks back? The Next Platform went and actually read the filing, and the number is a lot squishier than the coverage suggested. So how much of that headline figure is actually locked in? Not much, yet. Now, here's what The Next Platform reports: Marvell issued Google a warrant to buy up to fifty-nine million Marvell shares at two-hundred-six dollars and fifty-eight cents apiece — worth about twelve-point-two billion dollars if fully exercised. Only about one-point-four million of those shares vest automatically in year one. The rest unlock in tranches of roughly two-hundred-forty-thousand shares for every five-hundred-million dollars of qualifying chip revenue Google actually generates for Marvell, through fiscal 2033. Only if Google buys the full hundred-twenty billion in silicon does the warrant fully vest. And here's the kicker The Next Platform points out: Marvell's stock jumped enough on the announcement alone to cover the warrant's value before a single chip shipped. That's not a done deal you can bank on — that's a very long maybe with a nice press release attached.

A genuinely rare sight this week: OpenAI's account posted an open letter co-signed by more than one hundred organizations — Anthropic, AWS, Google, Microsoft, and Oracle among them — calling for urgent, coordinated action to shore up the cyber defenses of critical infrastructure as AI capabilities keep advancing. CEO Sam Altman amplified it himself on X, writing, quote, 'this is a critically important moment for cyber defense with AI; there is not much time to act. we are happy if you want to work with us or any of our competitors or partners, but please take this moment seriously. only an urgent and intense collective response will work,' end quote. That's rivals agreeing to work together, which doesn't happen often in this industry. Now, what it isn't, yet, is a funded program — the letter is a call to action, not a check being written or a specific initiative with a budget attached. I'd like to see what concrete resources actually follow before calling this more than a strong statement.

Here's a story with real electoral teeth. Data Center Knowledge has a piece out documenting how data center projects are turning into a straight-up campaign liability. In Florida, gubernatorial candidate David Jolly is pledging a statewide data center moratorium 'on Day One' if elected, pointing to Escambia County's ordinance banning large-scale data centers over water, power, and community concerns, and contrasting himself with rival Byron Donalds, whom he says is, quote, 'listening to the big tech interests backing him,' end quote. On X, Jolly put it this way: quote, 'Floridians are saying loud and clear: we don't want massive data centers threatening our water, raising concerns about our power grid, and changing our communities,' end quote. Meanwhile in Georgia — where our main story is happening today — Governor Brian Kemp has called a proposed statewide data center moratorium from Democratic gubernatorial nominee Keisha Lance Bottoms 'insane' and 'irresponsible.' Same industry, same week, two very different political postures. So is this a partisan wedge? Doesn't look like it. Now, over on X, Pennsylvania State Treasurer and gubernatorial candidate @GarrityForPA posted, quote, 'While you and I struggle with rising energy costs, Josh Shapiro is bragging about building "energy intensive" data center projects financed by his campaign donors. I need you to take action now,' end quote. Republicans and Democrats are running against these projects with nearly identical language — the wedge isn't party, it's who's paying for the power. That tension is about to show up again, in a much more concrete way, in our main story.

A policy move this week touches every one of these projects, whether they know it or not. Utility Dive reports President Trump signed an executive order declaring a national emergency and directing the Department of Energy to write rules restricting foreign-made transformers, inverters, batteries, and related grid software from the bulk power system — the high-voltage backbone that feeds everything, including data centers. Utility Dive reports the order targets equipment tied to entities in roughly two dozen countries, China chief among them, over sabotage and remote-access concerns, and it even lets DOE impose new conditions on foreign equipment already installed. Now, here's the catch: nothing is actually banned yet. DOE has to write the rules, and grid equipment manufacturers told Utility Dive they still need clarity — especially on software, where a product's actual country of origin isn't simple to pin down. In an industry already fighting a transformer shortage, adding a new compliance layer to that supply chain could turn a two-year lead time into something longer — we just don't know by how much yet.

From federal policy to a state fight over the exact same question our main story answers differently. Florida's new data-center ratepayer law, SB 484, is getting its first real test, and it's not going well for Duke Energy Florida. The state's public counsel, Walt Trierweiler, told the Florida Public Service Commission that Duke's compliance proposal, quote, 'doesn't attempt to comply with the most basic provisions' of the law, end quote — a law written specifically to keep data center costs off ordinary customers' bills. Duke's plan includes a twenty-year minimum service term and upfront cost-advance requirements for large-load customers, but it stops short of setting an actual new rate for them. An Earthjustice attorney representing Florida Rising went further, warning that if a data center bubble bursts, quote, 'the general body of customers will be left holding the bag for billions and billions of dollars' of infrastructure, end quote. Now, Duke's attorney countered that the proposal 'protects customers today, preserves the commission's authority tomorrow.' No ruling yet — but this is the pair worth comparing, because Georgia just tried to answer the same question in a completely different way.

Our main story today is the one I teased at the top, and it's the biggest 'who pays' data point we've gotten yet: the Georgia Public Service Commission has approved Georgia Power's contract to supply OpenAI's planned Effingham County campus — locally called Project Camellia — with three thousand two hundred megawatts of new generation. Let's be clear about what this is, because the deal-stage matters: this is not a memorandum of understanding, not a handshake at a conference. It's a signed contract that Georgia's utility regulator has actually reviewed and cleared. Here's what Atlanta News First and the Effingham Herald report about the terms. The site sits near Savannah, and OpenAI is reported to be investing more than thirty billion dollars in a four-building, four-point-four-million-square-foot campus there. Georgia Power says OpenAI will cover the full infrastructure cost to serve the site — that's the substations, the transmission, all the dedicated equipment a normal residential customer never has to think about, but that a three-gigawatt customer absolutely requires. OpenAI has also committed up to one thousand megawatts of what's called flexible, curtailable demand — a mechanism that lets it dial back its own power draw at peak times, so Georgia Power doesn't have to build generation sized for the campus's absolute worst-case moment, which is exactly the kind of overbuilding that tends to land on everyone else's bill. And here's the number Georgia Power is putting its name on: it's projecting the OpenAI contract, along with other large-load deals, will produce about nine hundred fifty million dollars in annual customer savings starting in 2029, rising to two-point-eight-four-seven billion dollars cumulative from 2029 through 2031. That would lift the typical residential customer's expected savings to one hundred eighty dollars a year, up from a prior commitment of one hundred two dollars. Power delivery is expected to begin sometime between 2028 and 2032. Now, why does any of this matter beyond Georgia? Because this same question — does making the data center pay its own way actually protect the ratepayer, or is that promise only as good as the assumptions behind it — is playing out in at least two other state dockets right now. We just talked about Florida, where Duke Energy's SB 484 compliance proposal is getting torn apart by the state's own public counsel for not setting a real rate at all. And Utility Dive has reported that PJM — the grid operator across a big chunk of the mid-Atlantic and Midwest — says data-center load made up roughly nine percent of that grid's wholesale costs this year. So you've got three different regulatory approaches to the same underlying problem, all surfacing in the same stretch of days: Georgia's negotiated contract with disclosed savings projections, Florida's statutory fight over whether a proposal even complies with the law, and PJM's wholesale market quietly absorbing the cost of the buildout in the background.

So does the Georgia deal actually hold up? Let's take it apart piece by piece, because there's real substance here and real gaps. The strongest thing going for this deal is the stage it's at. A signed, PSC-approved contract with disclosed megawatt and dollar figures is a stronger evidentiary position than almost anything else we cover in this space — most of what crosses my desk is an announcement, a groundbreaking photo, or a memorandum that could evaporate in a year. This one has actually been through a regulatory review. But here's what I'd flag before anyone declares this settled. First: Georgia Power hasn't disclosed what it's actually going to build to generate three-point-two gigawatts. Gas plants, more renewables, something else entirely — we don't know yet, and that's not a small omission, because the generation mix determines the real environmental and cost footprint of this deal for the next several decades. Second: those savings figures — the nine hundred fifty million a year, the two-point-eight-four-seven billion cumulative — are Georgia Power's own forward-looking projections. They are not audited outcomes. They're the utility's math, about a deal the utility itself structured, being used to justify the utility's own request. That doesn't make them wrong, but it does mean 'ratepayers save money' is currently a promise, not a receipt. And third: 'OpenAI pays the full infrastructure cost' is Georgia Power's characterization of contract terms that, as far as I can tell, haven't been made fully public. Local reporting notes the project is still in early planning, with substations and transmission lines yet to actually be built. Here's the beneficiary-pays test I keep coming back to on this show: whoever creates the incremental cost should pay the incremental cost. Dedicated infrastructure, reserved capacity, stranded-asset risk — that belongs to the developer. Shared regional benefits can be shared. On paper, this Georgia deal is built close to that spec: full-cost infrastructure recovery from OpenAI, plus a curtailable-demand mechanism specifically designed so Georgia Power doesn't overbuild generation on the ratepayers' dime. That's the shape of a good deal. My standard for the companies spending the most money on this buildout — and it's my standard, not a regulatory requirement — is that they can afford to make that math checkable in public, not just characterized by the utility that negotiated it. And that's exactly the piece that's missing right now. We have Georgia Power's word on the cost allocation. We don't have the contract language itself, and we don't have the generation-mix disclosure that would tell us what this three-gigawatt campus is actually going to run on. Compare that to what's happening in Florida this week, where regulators are publicly fighting, on the record, line by line, over whether Duke's proposal even complies with the ratepayer-protection statute. Georgia's process looks tidier specifically because less of it is happening in public view — and tidier isn't automatically better. So where do I land? This is a real step forward for the developer-pays model, and Georgia Power deserves credit for structuring a deal with an actual curtailment mechanism and disclosed savings targets, instead of just asking regulators to trust that a giant new customer will work out fine. But the savings promise and the generation-mix question both remain open, and I'd want to see the actual contract language and the fuel mix before calling this settled. Time for the Hype Check. I'm putting this one at a seven. Real regulator-approved contract, real disclosed megawatt and dollar figures — that's rare and it's earned. But the generation mix is undisclosed, the savings numbers are projections rather than audited results, and 'full-cost infrastructure recovery' is still the utility's own characterization of terms we haven't seen in full. Close the gap on those three things, and this becomes an eight or a nine.

So the next time you see a headline about a data center 'paying its own way,' the question worth asking is the one Georgia hasn't fully answered yet: whose numbers are those, and has anyone outside the utility checked them? We'll follow whether the actual contract language and the generation mix ever surface. If today's episode was useful, follow Concrete Compute wherever you get your podcasts, so tomorrow's episode finds you automatically. 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!