Investment Grade

On Thursday, July 9, S&P cut Oracle's issuer credit rating from BBB to BBB-. That's one step above junk, which, for a company as old and big and staid as Oracle, is ... not good. The same note named OpenAI as a key credit risk sitting within Oracle's $638 billion backlog and projected a free operating cash flow deficit of roughly $42 billion for fiscal 2027, on top of about $167 billion in total debt already on the books. Having this much debt and FCF deficit is also... not good. SMASH CUT to Friday, July 10, over in Madison: Wisconsin's Public Service Commission let a deadline expire without putting the item on the agenda, which is the regulatory version of getting ghosted. The item it declined to revisit was a rule the commission approved back in April, part of We Energies' new "very large customer" rate structure. Why? Turns out that any customer in that class rated below A- has to post financial guarantees before the utility will sell it electricity. Add those two together, and the Oracle subsidiary co-developing the Port Washington campus (roughly a gigawatt on 672 acres, alongside Vantage and OpenAI, if you haven't been following it) is staring at more than $100 million a year in cash deposits or letters of credit before they can move forward. To reiterate, $100 million a year for the privilege of being ALLOWED to buy electricity. Oracle had already sued the commission in Ozaukee County Circuit Court on June 19, arguing that the PSC acted outside its authority, on insufficient evidence, and that the A-line isn't needed to prevent harm to anyone. Maybe! But when they filed, it was two notches under the threshold. Three weeks later, it was three. That's not exactly what I would call "getting closer." What's funny here is nobody in this fight is arguing the stuff people usually argue about: megawatts, interconnection queues, or transformer lead times, or how much water the DC is going to take (hint: VERY VERY LITTLE). The site exists, the demand is real, everything lines up! Even We Energies wants the customer so badly that it asked its own regulator to stand down. The only thing in the way is a letter grade. The utility's argument is good, which is the problem We Energies asked the commission to back off the rule on June 10, warning it would push investment out of the state, and its lawyers put the credit case about as plainly as anyone could: "tens of billions of dollars in Oracle's value would need to be destroyed before creditors or counterparties, such as Wisconsin Electric and its other customers, could experience losses." Which is true! If I were Oracle's counsel, I'd make the same argument. And for the record, I am not a credit analyst; I have never rated so much as a parking-garage bond. But you don't have to be a financial genius to notice that this is a statement about a company, and the commission is not underwriting a company. Simply put, We Energies now bills large data centers directly for the generation built to serve them, and that generation includes the proposed Red Oak Ridge plant in the town of Paris, a build that runs north of a billion dollars. If you draw this all the way out, a gas plant is a thirty-to-forty-year asset, and the demand it's being built against is maybe three years old. So, ten to one, asset life over demand history, on a billion-dollar bet. If the tenant walks, the plant stays put, and the residual lands on schools, small manufacturers, and every residential meter in southeastern Wisconsin. So I think the PSC, whether it would ever phrase it this way or not, is pricing a duration mismatch, and Oracle's position with the credit markets doesn't really factor in. Even though the ratings agency picked the same week to agree, Oracle's borrowing to fund its AI position had already pushed debt-to-equity past 400% as of May, and the stock shed more than $50 in the month before the deadline. Whatever you think of the A- line, the commission's worst-case scenario got measurably more plausible while the reconsideration request sat on somebody's desk. And Wisconsin isn't some rogue outlier, much as Oracle's lawyers might want it to be. As of May, twenty-three states had approved at least one large-load tariff, with another seven pending. Virginia's version of Dominion's GS-5 tariff starts automatically applying to customers at 25 MW and up in January 2027, with minimum 14-year terms. I wrote in early July about FERC ring-fencing the wire while the scarcity leaks out anyway, and this is the other half of the same coin: cost allocation is about who pays; collateral is about who gets stuck with the plant when nobody does. The Northwest ran this experiment and kept the receipt When I started pulling on this thread, I found a strong historical record leading back to the Pacific Northwest. The Bonneville Power Administration began selling enormous volumes of cheap hydroelectricity to aluminum smelters in the 1940s, and by the end of the century, forty percent of U.S. smelting capacity was located in the Pacific Northwest. That arrangement lasted fifty years, longer than any computing platform I have ever heard of. But in 1996 and 1997, the smelters traded away chunks of their long-term BPA contracts for access to a wholesale market that, at that particular moment, was cheaper. But when 2000 arrived, and the West Coast power crisis came with it, wholesale prices rose by factors of ten and twenty. All of a sudden, the contracted power was worth more than the metal it was made of. What's the rational move? Don't smelt any more. The smelters paid their workers to stay home and resold their Bonneville electricity on the spot market. Paying people NOT to make aluminum penciled out better than making it, which is, before you ask, totally legal, exactly what the contracts allowed. By the summer of 2001, all ten of the region's smelters had shut down or dropped to token production, and most of them never came back. The dams stayed, though. The transmission built to feed Troutdale and The Dalles is still standing, still in the rate base, sized for an industry that left. Nobody defaulted, nobody committed fraud, and the region still wound up holding fifty years of single-purpose infrastructure after the single purpose moved away. I am fairly sure THAT is the scenario sitting in a spreadsheet in Madison right now, and I want you to notice it is not "will Oracle go bankrupt." Credit risk is an architecture problem Look, everyone is treating the collateral fight as a financing question, and I get why, but I think underneath it there's a granularity question which is more interesting. A $100 million annual guarantee only makes sense as an ask because the underlying asset is indivisible. One plant, one campus, one tenant, and no way to lose eight percent of Red Oak Ridge, because there is no eight percent to lose. Project finance figured this stuff out something like a century ago: break the huge commitment into pieces whose failures aren't perfectly correlated, and suddenly it's bankable. That is the entire reason nobody underwrites a single mortgage the size of a neighborhood. We built the AI buildout backward from all of that. I don't think anybody decided to, exactly. But everything got concentrated: one counterparty, on an indivisible thirty-year asset, justified by three years of demand history. On the other hand, if you could split it up into tranches of investable assets, things would change quite a bit. A portfolio of facilities spread across several interconnections, several utilities, and a mix of tenants doesn't need a letter of credit the size of a small utility's annual revenue, because no single tenant walking out strands a billion dollars of steel. Note that nobody got more creditworthy in that scenario! The blast radius just shrank until the counterparty stopped mattering so much, the same reason a bank will happily lend against a hundred small loans it can't individually assess and won't touch one enormous loan it can. (NOTE: This, and credit default swaps, were part of the root of the 2008 financial crisis - caveat emptor) This sort of distributed STUFF is what years of my life have been spent working on. Kubernetes at Google, and now Expanso, are variations on the same idea: stop trying to make any single component heroic, and make its failure boring instead. Small failure domains, loosely coupled, add up. I've been arguing for years that compute should sit closer to its power and its data, mostly on physics and cost grounds, and occasionally because the grid flatly said no. What I did not expect, and probably should have, was the balance sheet independently arriving at the same place, in a language utility commissions already speak. The market has half-conceded all of this, by the way. Last Sunday, the Journal reported that Nvidia is in talks to guarantee financing of around $250 billion so OpenAI can lease a 10-gigawatt campus that SB Energy is building in Piketon, Ohio. Strip off the zeros and squint, and it's the identical instrument the Wisconsin PSC asked for: a counterparty that can't carry the obligation alone, and a third party stapling its own balance sheet to the lease so the deal can close. One of those requests got a quarter-trillion-dollar term sheet. The other one got the regulator sued in Ozaukee County. The commissioners in Madison are asking, out loud, the question the industry has spent the past year answering privately and refusing to answer in public: if these assets can only be financed with a co-signer, what the hell do we think we've been building? Want to learn how intelligent data pipelines can reduce your AI costs? Check out Expanso. Or don't. Who am I to tell you what to do?. NOTE: I'm currently writing a book based on my observations of real-world challenges in data preparation for machine learning, focusing on operational, compliance, and cost issues. I'd love to hear your thoughts! Originally published at Investment Grade. Top comments (0)