The DOE's Quiet Rulebook for Efficient Data Centers

Today on Concrete Compute, we're doing something a little different. It's a quiet day on the wires, so instead of chasing headlines, we're going deep on a document that almost nobody reads outside of federal facilities managers: the Department of Energy's own best-practice guide for building efficient data centers. Before we get into it, two things worth flagging for later this week — a federal judge in Arkansas is weighing whether to block two newspapers from publishing leaked details of a Google power contract, a genuine First Amendment fight we'll dig into soon, and out in Texas, regulators are starting to ask how many of the gigawatts on their interconnection queue are actually real. We'll come back to both. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Wednesday, September 2, 2026. Let's get into it.

So who actually writes the rulebook for how a data center gets built efficiently? Inside the federal government, that's the Department of Energy's Federal Energy Management Program — FEMP for short, the office that pushes federal agencies to cut energy waste — working alongside the National Renewable Energy Laboratory, NREL, the government's main clean-energy research lab. Back on July 26, 2024, the two of them put out a best-practice guide for data center design, and the framing is the interesting part: they didn't just list a grab bag of tips. According to the guide itself, FEMP and NREL developed an approach for optimizing data center sustainability that's explicitly ordered — listed, in their words, in order of importance. Now, that might sound like a small editorial choice, but think about what it actually implies. It means somebody at NREL sat down and decided that not every efficiency measure pulls the same weight, and that if a facility manager only has the budget or the time to do a few things, some moves matter more than others, rather than treating a checklist as one flat pile of equally-weighted tasks. The stated goal, straight from the document, is simple: FEMP encourages federal agencies and organizations to improve data center energy efficiency because it can offer, in their phrase, tremendous opportunities for energy and cost savings. That's a government agency telling other government agencies — and by extension, anyone reading a public document like this — that the money on the table from doing efficiency right is substantial, not marginal. Now, why does a federal design guide matter to you if you don't run a data center? Because federal agencies operate a lot of their own computing infrastructure, funded by taxpayers, and every watt wasted there is a watt somebody's covering on the back end, whether that's a line in an agency budget or, indirectly, your own tax bill. And there's a broader signal here too — when the government that regulates power markets and hands out federal incentives publishes its own internal standard for what efficient actually looks like, that standard tends to ripple outward into how the rest of the industry talks about the same problem, even without a single regulation forcing anyone's hand. This is the same efficiency conversation that keeps surfacing whenever a state fights over who pays for a new substation or a new gas plant to serve a data center campus — the underlying question is always some version of: how much power does this thing actually need to use, and who's checking the math? It's worth sitting with that for a second, because it's easy to treat a design guide as bureaucratic paperwork and miss that it's answering the exact same question the big fights over interconnection queues and rate cases are answering, just at a much smaller scale — one building instead of one grid.

Here's where I have to be straight with you about the limits of what we know today. The excerpt we have access to from this DOE and NREL guide gives us the framing — order of importance, tremendous opportunity for savings — but it doesn't hand us the actual ranked list. No specific measures, no numbers, no line-by-line priorities. And I'm not going to invent them for you just to sound more complete, because that's exactly the kind of thing this show exists to push back against — a guide that gets summarized into something punchier than what it actually says on the page. So what can we responsibly say? We can say the structure itself is the lesson. A prioritized list is a bet that some interventions carry outsized returns, and that spreading your attention evenly across everything is actually a worse strategy than nailing the top few. That's a very old engineering idea — the eighty-twenty instinct — showing up in federal energy policy of all places. And it matters right now because the AI buildout has turned data center energy use from a back-office line item into a front-page fight. Every gigawatt announced this year gets scrutinized for how much of it is real demand versus speculative interconnection requests, and every new substation gets scrutinized for who's actually paying for it. A best-practice guide that ranks efficiency measures by importance is, in a sense, the same discipline applied one level down — instead of asking which projects deserve priority in a queue, it's asking which design choices inside a single building deserve priority on a spreadsheet. Think about what that ordering could mean at scale: if the top-ranked measure in a document like this turns out to be something structural — cooling design, airflow management, the kind of thing baked in before a single server rack goes online — then a facility that skips straight to the flashier line items further down the list is leaving real savings on the table, and doing it at the size of a federal computing footprint, not a single server room. I've made this point before on this show, and today's guide reinforces it rather than changing it: efficient design is the one lever in the whole buildout chain a developer controls completely, on their own timeline, without waiting on a utility or a grid operator to say yes. You can't will a transformer into existence faster, and you can't shortcut a two-year interconnection study. But you can absolutely design a building that needs fewer transformers in the first place, or that sizes its cooling correctly the first time instead of overbuilding and then discovering the load never materializes. My own standing view on this buildout is that the companies spending the most on these campuses can afford to be the most transparent about how efficient their designs actually are — that's my bar, not a legal requirement, and it applies whether the guide in question comes from a hyperscaler's sustainability report or, like today, from the federal government itself. The honest caveat is that a guide like this, aimed at federal agencies, doesn't bind a single private hyperscaler to do anything — it's persuasion, not regulation, and there's a real difference between a best-practice document and an enforceable standard. Whether it shapes design decisions at the scale of a five-hundred-megawatt campus is genuinely unclear from what we have in front of us today, and I'd rather leave that unresolved than pretend the reach is bigger than the evidence supports. Time for the Hype Check. I'm giving this one a three. It's a real document from a real federal program with real intent behind it — cutting energy waste at government-run facilities is a legitimate goal, and the order-of-importance framing is a smart idea worth remembering next time someone hands you an undifferentiated checklist. But the actual substance, the list itself, isn't something we could verify today, so I don't want to hand you specifics we can't back up just to make the segment feel fuller. The number I'd want next isn't a bigger claim — it's just the list.

So the belief worth challenging today isn't a hyperscaler's press release — it's the assumption that a best-practice guide only matters if it comes with a flashy number attached; sometimes the structure of the advice is the story. If today's dive into a quiet federal document was useful to you, follow Concrete Compute wherever you're listening, so tomorrow's episode — whatever the news decides it should be — 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!