Latest / The Operations Podcast with Fexingo: Process, People, and Profit in Modern Business / How a Data Center Cut Energy Use by 40 Percent
Transcript
- Lucas: So there's this data center in Ashburn, Virginia — right in the heart of the world's largest internet hub — and over about 18 months, they dropped their power usage effectiveness from 1.8 down to 1.08. That's a 40 percent reduction in energy waste. Luna: 1.8 to 1.08 — for anyone who doesn't live in data centers, what does that actually mean? Lucas: PUE, or power usage effectiveness, is the ratio of total energy consumed by the facility divided by the energy consumed by the IT equipment alone. A perfect score is 1.0 — meaning every watt that comes in goes straight to the servers, nothing wasted on cooling, lighting, or power distribution. Luna: And the industry average is around 1.6, right? So 1.8 is pretty bad, and 1.08 is basically best-in-class. Lucas: Exactly. And this wasn't a brand-new hyperscale facility built from scratch. This was a 15-year-old colocation data center that had been running the same way for a decade. They didn't rip out the servers. They changed how they managed air. Luna: So what was the first thing they did? Lucas: They started with hot-aisle containment. In most older data centers, cold air from the CRAC units — computer room air conditioners — gets pushed into a raised floor, and then it comes up through perforated tiles in front of the server racks. But the hot exhaust from the back of the racks mixes with the cold air, so the cooling system has to work harder to maintain inlet temperatures. Luna: Right, because the thermostat is reading mixed air, not the actual intake temp. Lucas: Exactly. So they installed physical barriers — basically plastic curtains and ceiling panels — to seal off the hot aisles. Now the hot exhaust goes straight to the return vents, and the cold supply stays cold. That alone dropped their PUE from 1.8 to about 1.4 within three months. Luna: That's a huge gain from just plastic sheeting and some duct tape. Lucas: It sounds simple, but the discipline is in the sealing — any gap, any missing grommet, and you lose the containment. They spent two weeks just walking every row with a thermal camera and patching leaks. Luna: So after containment, what came next? Lucas: They raised the set point. Most data centers keep server inlet temperatures around 68 to 72 degrees Fahrenheit. That's a holdover from the days of mainframes and magnetic tape. Modern servers from Intel and AMD can run safely at 80 degrees or even higher, per ASHRAE's updated guidelines. Luna: But there's a psychological barrier there, right? No one wants to be the person who lets the servers get hot and then they crash. Lucas: Absolutely. The operations team was nervous. So they did it in stages — one degree per week, monitoring inlet temps and CPU utilization. They found that at 78 degrees, the fans inside the servers started to ramp up a bit, but the actual processor temperatures stayed well within spec. Luna: And how much did that save? Lucas: Raising the set point from 68 to 78 degrees reduced the cooling load by about 20 percent, because the chiller doesn't have to work as hard. Combined with the containment, they were now at a PUE of about 1.25. Luna: Still not at 1.08 though. What was the third big move? Lucas: They implemented free air cooling. In Northern Virginia, the outside temperature is below 70 degrees for about 60 percent of the year. So they installed economizers — basically big louvers that bring in outside air when it's cool enough, and exhaust hot air out. When it's 50 degrees outside, you don't need the chiller at all. Luna: But doesn't that introduce humidity and particulate issues? Lucas: It can. So they added evaporative cooling for humidity control — basically misters — and MERV-13 filters for particulates. The filters need to be changed monthly, which is an operational cost, but it's much smaller than running the chiller. That brought them down to about 1.12. Luna: So how do you get from 1.12 to 1.08? That last stretch seems like it would be the hardest. Lucas: It is. And that's where the AI comes in. They deployed a machine learning model that predicts server workload based on historical patterns — think retail traffic spikes on Black Friday, or batch processing jobs that run at 2 AM. The model then adjusts the cooling in advance, zone by zone, instead of reacting. Luna: So instead of the thermostat sensing a hot spot and then cranking the AC, the system knows 'in 15 minutes, rack 12 is going to get hammered' and pre-cools that aisle. Lucas: Exactly. It's predictive rather than reactive. The model also optimizes the chiller plant — deciding when to run the chillers versus the economizers, and at what capacity. Over a year, that optimization shaved off another 0.04 from the PUE. Luna: And what about the financials? A project like this — containment panels, economizers, filters, AI software — that's not free. What's the payback period? Lucas: The facility's annual energy bill was about $12 million. A 40 percent reduction saves roughly $4.8 million a year. The total retrofit cost was around $3 million, so payback was under nine months. After that, it's pure savings. Luna: That's a no-brainer for any CFO. But I wonder why more data centers don't do this. Lucas: A few reasons. One is the perceived risk — no one wants to be responsible for an outage. Another is organizational inertia: the facilities team and the IT team often don't talk to each other. IT doesn't want to change the temperature because they think it will break things, and facilities doesn't have the data to prove otherwise. Luna: So it's a communication problem as much as a technical one. Lucas: Absolutely. And that's where operations as a discipline comes in — it's not just about the engineering. It's about creating the feedback loops and the trust. The data center that did this had a weekly meeting where the facilities manager and the IT ops manager reviewed the PUE data together. That simple. Luna: It reminds me of the Toyota Andon cord — the idea that any worker can stop the line if they see a problem. Here, any team member can flag a temperature anomaly without fear. Lucas: Exactly. And look, we talk about these kinds of operational wins on this show all the time — and we try to make them as concrete as possible because that's how real improvement happens. No ads, no fluff, just the specifics. That's a deliberate choice. Luna: Yeah, and it's one we can only keep making with listener support. If these episodes have helped you think differently about a problem at work, consider throwing a few dollars our way. Lucas: The link is buy me a coffee dot com slash fexingo. It keeps the show ad-free and independent. No pressure at all — but every bit helps. Luna: Alright, back to the data center. So they hit 1.08 — are they done, or is there more to squeeze? Lucas: They're actually targeting 1.04 next. The next frontier is liquid cooling — direct to chip or immersion. For high-density racks running AI workloads, air cooling just isn't efficient enough anymore. But that's a bigger capital investment. Luna: So the lesson here is really about the low-hanging fruit. Containment, set point, economizers — and then use data to fine-tune. Lucas: Right. And the order matters. You don't jump straight to AI before you've sealed the hot aisles. Each step builds on the previous one. That's classic operations thinking: solve the simple, high-impact problems first, then layer on complexity. Luna: And the savings are real. Four point eight million a year — that's not theoretical. Lucas: Not at all. And with data center energy demand projected to grow 15 percent annually through 2030, these kinds of operational improvements aren't just nice to have. They're becoming a competitive necessity. Luna: Especially for companies that are serious about their carbon commitments. Lucas: Exactly. Reducing energy waste directly reduces Scope 2 emissions. So it's a win for the bottom line and for the planet. And it all starts with measuring what you're doing and being willing to challenge the default settings. Luna: Speaking of defaults — I think the default for most operators is 'if it ain't broke, don't fix it.' But this case shows there's huge value in questioning whether 'ain't broke' is actually optimal. Lucas: That might be the most important takeaway. The data center wasn't broken. It was running fine. But 'fine' left $4.8 million on the table every year. Operations is about finding that gap between fine and great. Luna: And then closing it, one degree at a time. Lucas: One degree at a time. That's the episode. Thanks for listening.