Nvidia's Shrinking Guarantee Behind Ohio's Mega AI Campus

What happens when the world's biggest AI infrastructure guarantee gets cut by more than half in one week — does that make the deal safer, or does it just prove nobody's sure who's actually on the hook? Today on Concrete Compute: Nvidia, OpenAI and SB Energy make it official on the largest AI campus ever proposed, in Pike County, Ohio — but the financial guarantee behind it has shrunk from a discussed two hundred fifty billion dollars down to about one hundred five billion, and we're digging into why. Before that, in the headlines: New South Wales tells data center builders bring your own wind power or don't bother, a Texas gas-plus-nuclear plant takes its next step toward powering a Crusoe campus, and Kentucky taxpayers start asking hard questions about fifty-year tax breaks. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Monday, August 17, 2026. Let's get into it.

First up, New South Wales — the Australian state — just rolled out a new framework requiring new data centers to source at least forty percent of their energy from wind power. Developers will have to sign power purchase agreements — long-term contracts locking in renewable electricity — with wind providers, meet environmental and efficiency standards, and here's the part that matters most: impose no net cost to consumers. Comply, and the state promises to assess your application within seventy-five days. Don't, and NSW Treasurer Daniel Mookhey put it about as plainly as a politician can: 'Data centres that want to build in New South Wales need to bring power, not take it.' He went further, warning that any builder who skips the rules 'will find themselves almost alone.' There's also a water pricing review coming, run by the state's Independent Pricing and Regulatory Tribunal, aimed at protecting water users from drought and scarcity impacts as these projects scale. NSW already hosts ninety data center facilities, clustered mostly around Sydney, with players like AirTrunk, AWS, Equinix and NextDC — and the state projects data centers could reach eleven percent of its total energy demand by twenty-thirty. Now, this is a policy framework, not yet tested against a specific pending project, so we don't know what enforcement looks like beyond the threat of outright rejection. But as a statement of principle, it's about as close as I've seen a government come to writing a ratepayer-protection test straight into law.

Now, over in Texas — Blue Energy and GE Vernova Hitachi Nuclear Energy just signed an agreement to advance the next phase of a two-and-a-half gigawatt gas-plus-nuclear power plant in Victoria, Texas. This one's earmarked to power a Crusoe AI data center campus at the Port of Victoria. The phasing here is worth noting: Blue Energy plans to deliver about one gigawatt using two GE Vernova gas turbines starting in twenty-thirty, then layer on another one-point-five gigawatts from up to five GE Vernova Hitachi small modular reactors — the BWRX-300 design — beginning in twenty-thirty-two. Blue Energy calls its approach the 'Blue Way': building reactor components off-site at shipyards, then transporting and assembling them, rather than manufacturing reactors itself. But — and this is the nuclear rule I keep coming back to on this show — the whole project still hinges on a final investment decision that isn't expected until twenty-twenty-seven. An engineering and licensing agreement is real progress. It is not a permit, and it is not steel in the ground.

In Kentucky, taxpayers are starting to ask a pointed question about the state's data center incentives: is this actually worth it? Three hyperscale projects are moving forward under a sales-tax exemption that can run up to fifty years for any data center investing more than four hundred fifty million dollars — twenty-five years for smaller ones. Here's the catch critics point to: data center equipment typically gets replaced every three to five years, and each replacement can reset the clock, generating fresh tax breaks again and again until the exemption finally expires. The cautionary tale everyone points to is Virginia, which has the country's biggest concentration of data centers and, according to a legislative audit, tax exemptions costing the state more than one point six billion dollars a year — for an estimated return of just forty-eight cents in state revenue for every dollar spent. The Kentucky Center for Economic Policy is the one raising alarms; industry boosters counter that these projects bring six-figure jobs to rural, deindustrialized parts of the state. Worth noting: no projects have actually been approved under Kentucky's statewide incentive yet, so this is a projection, not a bill that's come due — for now.

Elsewhere, an X thread was making the rounds this weekend that pairs neatly with our main story today. The account at MelvinInvests — who posts AI infrastructure analysis as an analyst for Milk Road Pro — argued that power and grid infrastructure, not chip supply, is now the real ceiling on how fast the AI buildout can move. The thread's framing: transformer lead times have stretched out dramatically over the past few years, and grid operators like PJM are already warning of years-long delays for new interconnections — which, if true at the scale the thread claims, would mean billions of dollars of GPUs sitting in finished buildings that simply can't power on yet. I'll flag that these are one analyst's own numbers and framing, not a peer-reviewed study, so treat the specifics as his read rather than settled fact. But the core argument keeps showing up everywhere in this space, and it landed the same day Nvidia and SB Energy touted a four-point-two-billion-dollar grid investment in Ohio — which tells you every serious player in this buildout already treats power, not silicon, as the scarce resource. This is a running theme in his broader work, too — Milk Road's AI coverage regularly makes the case that grid and power constraints, not chip supply, are what's actually capping the buildout, and today's thread was just the latest version of that argument.

One more before we get to Ohio: Sunrun and Voltus announced a deal to feed capacity from residential solar-plus-storage systems into Voltus's 'Bring Your Own Capacity' program, which lets hyperscalers use aggregated home batteries to help satisfy grid interconnection requirements. The catch — and Data Center Knowledge's own reporting flags this directly — neither company disclosed how much capacity is involved, how many homes are participating, or which hyperscaler projects benefit. And there's a physics problem underneath the marketing: Grid Strategies founder Rob Gramlich and Persistence Analytics' Neil Osnato both note in that reporting that a battery sitting hundreds of miles from a data center can help regional resource adequacy while doing nothing for the local transformer or substation constraint holding up that specific facility. Capacity, in other words, is not the same thing as deliverability. It's the same lesson from a similar Sunrun virtual power plant announcement last year that promised sixteen-point-eight gigawatts and, on closer inspection, worked out to roughly four gigawatts of resources analysts actually consider firm.

Our main story today: the guarantee that keeps shrinking — and what that shrinkage actually tells us about who's holding the risk on the biggest AI campus ever proposed. Nvidia, OpenAI, and a company called SB Energy made it official this morning: they're building what they're calling the world's largest planned AI campus, at a site called PORTS-Pike, in Pike County, Ohio. The location has a story of its own — this is a redeveloped former uranium-enrichment site, the old Portsmouth Gaseous Diffusion Plant, sitting on private and federal land in Appalachian Ohio, developed in partnership with AEP Ohio, the Department of Energy, and the Department of Commerce. Here's the shape of the deal. Nvidia is securing what the industry calls 'land, power and shell' capacity — basically the site, the electricity, and the building itself, before a single chip goes in — to exclusively host Nvidia AI compute. The initial commitment is four-point-two-five gigawatts of what they call IT capacity — the actual usable computing load, as opposed to the bigger number you get once you count cooling and overhead — with an option to expand to eight gigawatts total. SB Energy will build, own, and operate the data center, then lease it to OpenAI for twenty years. OpenAI is the customer using the compute. Capacity is expected to come online in phases starting in twenty-twenty-eight. Now here's the number that's actually been moving, and moving fast. According to an SEC filing cited by Axios, Nvidia has agreed to guarantee up to one hundred five billion dollars in conditional lease and power payment obligations to SB Energy — a financial backstop that kicks in if things go sideways. But CNBC had previously reported the number under discussion was as high as two hundred fifty billion dollars, and it apparently passed through a stop around one hundred twenty billion before landing at one-oh-five in the actual filing. That's three different numbers reported in about a week for essentially the same guarantee. Nvidia is also putting one-point-five billion dollars directly into SB Energy — down, notably, from a previously reported three billion. On the power side, SB Energy and SoftBank say they'll build at least ten gigawatts of new energy generation and invest at least four-point-two billion dollars in regional grid infrastructure, in partnership with AEP Ohio, explicitly framed as protecting ratepayers from footing the bill. And on jobs — OpenAI says the project will support thirty-five thousand construction jobs through twenty-thirty-two, and two thousand five hundred permanent positions once it's running. There's also an eighty-million-dollar community benefits fund, combining SB Energy's original forty million with an additional forty million OpenAI is adding on top. The technical backbone here is Nvidia's full-stack DSX AI factory platform — GPUs, CPUs, and networking built as one integrated system, designed to make the whole thing upgradeable generation after generation without ripping out the shell every time. And this wasn't put together on a handshake: Goldman Sachs and JPMorgan advised SB Energy, Morgan Stanley advised Nvidia.

So how does this actually shake out — is this the circular financing everybody's been worried about, or isn't it? Jensen Huang didn't wait for skeptics to pile on — he went straight to X to shut it down: 'Is this circular financing? No. OpenAI will pay the lease,' Huang said on X. Circular financing, for anyone just tuning in, is the worry that a chipmaker's own investment dollars end up counted as proof its customers actually want the product — Nvidia funds the buyer, the buyer buys Nvidia chips, and everyone points to the resulting revenue as organic demand. Huang's argument is that this isn't that: OpenAI is the one paying rent on the lease, independently, as the tenant. Fair enough — but Axios, reporting on the same filing, put it plainly: this deal is likely to reignite fears of a circular AI bubble anyway. Both things can be true. OpenAI paying its own lease is exactly the kind of independent, paying customer that should reassure people. Nvidia backstopping that lease with a guarantee that's moved from two-fifty to one-twenty to one-oh-five billion in the space of a week is exactly the kind of unsettled deal structure that should make people ask more questions, not fewer. NVIDIA's own announcement quotes SB Energy's co-CEO, Rich Hossfeld, framing the project as something bigger than a data center: 'Infrastructure is vital for the AI economy. With SoftBank Group, OpenAI and NVIDIA, SB Energy is building power-first infrastructure at unprecedented scale while strengthening the communities that make it possible — protecting ratepayers, creating tens of thousands of well-paying jobs, and investing in infrastructure to revitalize Southern Ohio,' according to that release. The same announcement quotes OpenAI's Sam Altman leaning into the human scale of it: 'This is going to be a huge site, with enough computing power to help millions of people use AI to do things we can only start to imagine today, from finding new medicines to starting businesses and solving hard problems.' He's also quoted adding, 'We want the people who live here to feel the benefits too, through good jobs, more opportunity for local businesses, and investment in the community for years to come.' Here's my read. On the ratepayer question — my standing test for these deals is simple: whoever creates the incremental cost should pay it. A four-point-two-billion-dollar grid investment funded by SB Energy and SoftBank, explicitly structured with AEP Ohio to keep the cost off residential bills, is the shape of a good deal on paper. The part nobody's published yet is the actual cost-allocation filing that would prove Pike County ratepayers see zero net impact — until that document exists, 'designed to protect ratepayers' is a claim, not a receipt. On jobs — put both numbers on the table, because they're different questions. Thirty-five thousand construction jobs through twenty-thirty-two is real work for real people, for about six years. Two thousand five hundred permanent positions is the number that actually tells you what this site looks like in twenty-thirty-three. Anyone spending what Nvidia spends on this buildout can afford to be precise about which number is which — that's my standard, and I'd apply it to any company doing the same thing. And on the guarantee itself: what actually changed is that Nvidia's exposure phases in only as capacity gets placed into service, not upfront. That's a real distinction — it's not a blank check sitting on Nvidia's balance sheet today. But three different guarantee figures reported in one week, for the same underlying deal, tells you the lawyers were still negotiating the number as the press release was going out. Time for the Hype Check. I'd put this one at a six. The land is real, the site's history is real, the jobs commitments are specific enough to hold someone accountable to later — but the headline financial guarantee has moved three times in seven days, and 'guarantee' is doing a lot of unearned work when it only kicks in as gigawatts actually get energized, years from now. So the open question heading into twenty-twenty-eight, when the first phase is supposed to come online: if Nvidia's exposure keeps shrinking every time the filing gets updated, whose balance sheet is actually left holding the risk if this doesn't go the way the press release says it will?

If you're the kind of person who wants the receipts on that guarantee number before you believe any of it — same here, and that's exactly why this show exists. Follow Concrete Compute wherever you're listening, so tomorrow's follow-up on Ohio lands in your feed 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!