Seventeen months. That's how long Chris Malone spent building OpenAI's data center empire — the actual buildings, the actual power contracts, the actual relationships with utilities that make Stargate real instead of a slide deck. And now he's gone, with no reason given, in the same twenty-four-hour window OpenAI put out benchmark numbers claiming its homemade chip beats Nvidia's best. So today on Concrete Compute: is OpenAI's infrastructure story as buttoned-up as the marketing suggests, or are we watching a company sell confidence with one hand while it loses institutional memory with the other? Before that, in the headlines: a coal-and-gas plant that Pennsylvania already agreed to retire is getting kept alive again, and a Virginia startup just raised a hundred fifty million dollars betting that the power crunch itself is the business. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Wednesday, August 26, 2026. Let's get into it.

Let's start in Pennsylvania, where the Department of Energy has again ordered the Eddystone Generating Station to stay available for dispatch. Data Center Knowledge, which broke down the order, reports it keeps the plant online through November 20th, and describes it as the sixth such DOE intervention since Eddystone's planned 2025 retirement. You'd think at some point a plant either retires or it doesn't, but here we are again. Per that reporting, the order covers 760 megawatts across two units that Constellation Energy and grid operator PJM — the company that runs the wholesale power market for a huge chunk of the mid-Atlantic — have to keep economically dispatchable, meaning available to run when needed, even though the order specifically says these units don't count as guaranteed capacity. That distinction matters more than it sounds: the plant has to be ready to fire up, but it isn't part of the official reserve margin PJM counts on for planning purposes, which is exactly the kind of gray zone that makes critics uneasy. Why does this keep happening? Data Center Knowledge cites DOE's own reliability analysis pointing to roughly twenty-five gigawatts of projected load growth coming to PJM, with about fifteen gigawatts of that from data centers, running up against roughly seventeen gigawatts of fossil plants already scheduled to retire. That's a supply gap you can see coming from years away, and Pennsylvania's consumer advocate, Darryl Lawrence, is quoted in that reporting saying he hasn't seen demand forecasts like this in his twenty-plus years in the business — in his words, "We haven't seen this kind of forecasted demand in the PJM system since the days when they were building steel mills." Now, here's the question nobody's answering yet: who actually pays for keeping an old plant on standby? The order requires PJM to minimize costs to ratepayers through economic dispatch, but minimizing a cost isn't the same as eliminating it, and the final bill — and who eats it — is still an open question nobody in this reporting has settled.

Now, sticking with the power theme — Emerald AI just raised a hundred fifty million dollars in an oversubscribed Series A, valuing the company at just over a billion dollars, with Energize Capital and DCVC co-leading. Their pitch is genuinely interesting: instead of building more power plants, make AI data centers flex their electricity use to match what the grid can actually deliver right now. They're already proving it out with a serious partner list — Digital Realty, Nvidia, Dominion, and PJM itself — on a nearly hundred-megawatt project in Manassas, Virginia, called the Vera Rubin AI Research Factory, slated to come online later this year. Energize Capital's John Tough summed up the thesis this way: quote, "The binding constraint on AI is no longer chips or capital; it is power." Here's my read: that's a smart bet on a real bottleneck, but let's not get ahead of ourselves — the hundred-plus gigawatts of grid capacity Emerald says it could unlock is the company's own framing of theoretical potential, not a signed commitment from anybody, and the Manassas pilot hasn't been energized yet either. It's also worth remembering that software flexibility only helps if the underlying grid infrastructure — the wires, the substations, the transformers — actually exists to deliver power when demand shifts; Emerald's tool manages the timing, it doesn't build the poles and wires. Promising software, real partners, still pre-launch.

Our first lead today: OpenAI's homemade chip, and whether the numbers behind the applause actually hold up. At Hot Chips this week, OpenAI published its first real benchmark results for Jalapeño, the inference chip it co-developed with Broadcom — inference just means running an already-trained AI model to answer questions, as opposed to training it in the first place, and it's where most of the day-to-day compute spending actually happens. The numbers themselves are striking: on SemiAnalysis's public InferenceX benchmark suite, Jalapeño delivered one and a half to nearly two times more throughput per kilowatt, and up to three point six times lower latency, than Nvidia's GB200 and GB300 rack systems — and it did it running at 700 watts against Nvidia parts rated at 1,200 and 1,400 watts. Richard Ho, OpenAI's head of hardware, put it directly on a press call: "Jalapeño can serve more AI work per unit of power, while also returning responses more quickly." OpenAI says it plans to start deploying the chip in its own data centers later this year, at gigawatt scale over multiple generations, with data center partners. Now, the reaction was immediate and it was loud. Greg Brockman quoted the release approvingly. And Dylan Patel of SemiAnalysis, after his own on-site analysis, called it huge news — his read is that first-generation silicon is already competitive on total cost of ownership against Nvidia's current lineup, and maybe even its upcoming Rubin chips. So is this the moment OpenAI stops needing Nvidia? Not so fast. Here's the catch that got buried in the celebration: every one of those numbers is measured against Nvidia's GB200 and GB300 — chips that are currently shipping, not Nvidia's next-generation Rubin hardware. And when SemiAnalysis ran its own separate comparison against Rubin, the two came out roughly even on cost per token — practically a tie, producing almost the same number of output tokens per dollar. Part of the reason, according to SemiAnalysis, is a technique called speculative decoding, which lets a chip guess ahead and verify in batches: Jalapeño's results didn't use it, Rubin's did, and SemiAnalysis says that technique is worth roughly a three-to-five-times reduction in cost per token when applied — though that's SemiAnalysis's own estimate, not an independently confirmed industry figure. Strip that context away and OpenAI's celebratory framing — beats Nvidia, full stop — gets a lot more complicated, because the comparison that actually matters for OpenAI's future buying decisions is against the chip Nvidia hasn't even shipped yet, not the one it's already selling. This is also, worth saying plainly, a benchmark OpenAI ran itself, on a suite it chose, published by OpenAI, with actual deployment starting at what the company itself calls small volumes by year's end. That's not third-party-audited production performance yet, and SemiAnalysis's own deeper read is careful to note the results establish an inference advantage specifically — not a claim that Jalapeño beats Blackwell at anything else, and OpenAI hasn't claimed that either. It's a genuinely impressive first showing for homegrown silicon — and it's also a press release with a capex story to tell.

Our second lead: the man who spent seventeen months building OpenAI's physical infrastructure is out, and OpenAI isn't saying why. Chris Malone, who joined as head of data centers in March of 2025 after building infrastructure at both Meta and Google, has left the company — the Wall Street Journal first reported it, and CNBC has since confirmed it. In a statement to TechCrunch, OpenAI said it had, quote, "recently reorganized" its infrastructure organization "to support the scale and pace of our work," adding: "We have a strong, deeply experienced data center team in place, with clear leadership and the technical expertise to execute our plans." Here's the structural detail worth noting: as part of that reorganization, Malone had stopped reporting directly to OpenAI President Greg Brockman and started reporting to Vice President Sachin Katti, who took over leadership of the group — and according to the Journal, OpenAI is now reviving plans to lease entire data center facilities outright, with others stepping in to lead that effort in Malone's place. And this isn't happening in isolation. Malone's exit adds to a run of senior departures this year: revenue chief Denise Dresser announced she was leaving after less than a year in the role, that came shortly after longtime executive Brad Lightcap said he was ending an eight-year run at OpenAI to, in his words, start something new, and Fidji Simo stepped down from her role to focus on managing a chronic illness. Now, does one departure without a stated reason mean anything? On its own, maybe not — companies replace revenue chiefs inside a quarter and the business survives fine, and OpenAI has now done exactly that twice this year. But data center programs don't run on org charts, they run on relationships — with utility executives, planning boards, construction contractors, grid operators — and those relationships sit with specific people, not with a title on a company website. Interconnection queues, the waiting lines to physically connect a new project to the grid, run for years, and so do the underlying contracts; you can't hand those off in a press release the way you can reassign a reporting line. It's also the one function where a mistake is the most expensive kind: a site that can't get power turned on when scheduled strands capital that's already been spent on steel and concrete, sitting idle while the clock on financing keeps running. No reason for Malone's departure has been disclosed by either OpenAI or Malone himself, and that silence is doing a lot of work in how this story gets read.

Put these two stories next to each other and here's what jumps out at me. In the space of one day, OpenAI told the market two very different things about its infrastructure: one message says we're moving so fast we just leapfrogged Nvidia's currently shipping chips, and the other says the executive who spent seventeen months building the actual buildings quietly walked out the door with no explanation. My read is that neither story cancels the other out — a company can genuinely have strong chip engineering and a leadership churn problem in data centers at the same time, and Stargate is big enough to contain both truths without either one being the whole picture. What would move me is straightforward: on the chip side, an independent third-party benchmark against Rubin, not GB200, run outside OpenAI's own chosen suite; on the personnel side, a named successor with a public track record on the megawatt-and-permitting side of the business, not a statement about reorganizing the org chart. Time for the Hype Check. I'd put this pairing at a five out of ten on substance — the chip result is real and independently discussed by SemiAnalysis, but it's measured against last-generation Nvidia hardware and, per SemiAnalysis's own estimate, excludes a technique that would erase most of the gap, while the executive exit is confirmed but completely unexplained. Real news, wrapped in more confidence than the underlying evidence quite earns.

If OpenAI names Malone's permanent successor in the next few weeks with real authority over sites and power contracts, that tells us this was reorganization, not retreat — if the seat stays vague past Stargate's next big groundbreaking, that tells us something else. Follow Concrete Compute wherever you get your podcasts so you don't miss how this one plays out. 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!