Latest / The Fiscal Policy Podcast with Fexingo: Government Budgets, Stimulus, and Public Spending / Why Government Data Centers Are 40 Percent Empty
Transcript
- Lucas: So there's this number that keeps showing up in GAO reports and inspector general audits, and it's almost too absurd to say out loud: the federal government runs roughly eight thousand data centers, and on average, only about sixty percent of the server capacity inside them is actually doing anything. Luna: Sixty percent utilization — so forty percent is just... sitting there, drawing power, generating heat, costing money? Lucas: Exactly. The Government Accountability Office put out a report last year — 2025 — and they estimated that the government spends something like six billion dollars annually just on data center energy and cooling. A huge chunk of that is wasted on idle servers. And this isn't a new problem. They've been trying to consolidate data centers since the Obama administration. Luna: But clearly it hasn't worked. What's the obstacle? Is it technical, cultural, or just bureaucratic inertia? Lucas: It's all three, but I think the core issue is a misalignment of incentives. When an agency owns its own data center, that's 'their' infrastructure. It gives them control, it's part of their budget, their staff run it. If you tell the Department of Agriculture, 'Hey, you need to move your workload to a shared cloud platform,' the people who manage that center might lose their jobs, or at least their relevance. So they push back. Luna: Right, so it's a classic principal-agent problem. The taxpayers want efficiency, but the agency decision-makers want to preserve their turf. Lucas: Exactly. And there's also a procurement angle. A lot of these data centers were built under multi-year contracts with specific vendors. The government is locked in. Even if they wanted to shut a center down, they're still paying for the lease or the maintenance contract. So the center stays open, running at thirty percent utilization, because it's cheaper to keep it running than to break the contract. Luna: Oof. So you're paying for servers you don't need, plus the electricity to run them, plus the cooling to keep them from melting. That's the trifecta of waste. Lucas: And the privacy and security concerns add another layer. Some agencies — like the Department of Defense or the VA — handle sensitive data. They are legitimately worried about putting that data on shared infrastructure. So they build their own private cloud, which then sits underutilized because they over-provision for peak loads. Luna: I remember reading about the VA specifically. They had something like a hundred and twenty data centers a few years ago, and they announced a big consolidation plan. Lucas: Right. In 2016, the VA said they would consolidate from 120 centers down to maybe 12. Fast forward to 2026, and they've closed... fewer than 20. They spent over a billion dollars on the effort and still have more than a hundred centers. The GAO called it 'persistent underperformance.' Luna: A billion dollars to close 12 data centers? That's about eighty million per closure. What is that money actually buying? Lucas: A lot of it is migration costs — moving applications off old servers onto new ones, re-architecting software, testing security. But also, the VA kept changing the plan. They'd start a migration, then a new CIO would come in and say, 'Actually, we're going to do it differently now.' So they'd abandon the work in progress and start over. Luna: That sounds like a classic case of 'throwing good money after bad.' Lucas: It is. And the private sector figured out a solution to this years ago: just stop owning data centers. Amazon, Google, Microsoft — they run hyperscale data centers that achieve utilization rates above eighty, even ninety percent, because they pool demand across thousands of customers. The government could, in theory, just buy compute capacity from them. Luna: And some agencies do. The CIA signed that big deal with Amazon Web Services back in 2013. But adoption has been slow. Lucas: Very slow. A 2024 report from the Office of Management and Budget said that only about 25 percent of federal IT workloads are in the cloud. The rest are still in those eight thousand data centers. And the cloud spend itself is often inefficient — agencies buy reserved instances they don't fully use, or they lift and shift applications without optimizing them for the cloud, so they end up paying more than they would have on-premises. Luna: Speaking of spending — this is the part of the show where I mention that our show itself runs on listener support. If you get value from these deep dives into how your tax dollars are actually spent, and you want to help keep this ad-free, there's a simple way to chip in. Lucas: Yeah, it's a small group of listeners who contribute monthly through buy me a coffee dot com slash fexingo, and that genuinely funds the research and production time for episodes like this. No ads, no sponsors, just people who think this kind of fiscal transparency matters. Luna: Exactly. And we're grateful for every one of them. So, back to the data center problem — what would a real solution look like? Lucas: I think the most promising approach is to change the procurement model. Instead of buying servers and building data centers, agencies should buy 'compute as a service.' You define your workload requirements, and vendors compete to provide the infrastructure. That shifts the risk of underutilization to the vendor, who can pool demand across multiple agencies. Luna: That's essentially what the Federal Risk and Authorization Management Program, FedRAMP, was supposed to enable — a standardized security approval so agencies could buy cloud services off the shelf. But the process is still slow. Lucas: It is. A vendor might wait 18 months to get FedRAMP certified. That's an eternity in tech. By the time they're approved, the product has moved on. So you get this mismatch — the government wants modern infrastructure, but the procurement process forces them to buy yesterday's technology. Luna: Is there any agency that's actually doing this well? A model worth following? Lucas: The U.S. Digital Service — now part of the Technology Transformation Services — has helped a few agencies migrate successfully. One example is the Small Business Administration. They moved their loan-processing system to the cloud and cut costs by about 30 percent while improving uptime. But those are one-off successes. There's no systematic pressure to consolidate. Luna: So what would create that pressure? A mandate from OMB? Congressional action? Lucas: Congress could tie funding to closure milestones. Say, 'You get your IT budget only if you shut down 10 percent of your data centers per year.' But that's politically difficult because it means taking on agency heads and their congressional oversight committees. The other route is to make the cost of ownership transparent. Right now, most agencies don't actually know what their data centers cost, because the bills are buried in different budget lines. Luna: Classic. If you can't measure it, you can't manage it. Lucas: Exactly. The GAO recommended that OMB create a centralized inventory of all federal data centers with real-time utilization data. That was in 2010. They repeated the recommendation in 2016, then again in 2020, then again in 2025. And each time, OMB says 'we're working on it.' Luna: So we've known the fix for fifteen years, and we still haven't done it. That's a pretty damning indictment of the system. Lucas: It is. But I'll end on a slightly optimistic note. The rise of generative AI is forcing agencies to rethink their compute strategy, because AI workloads need massive GPU clusters that no single agency can afford to build. So they're being pushed toward shared infrastructure and cloud services out of necessity. That might finally break the logjam. Luna: So maybe AI will do what fifteen years of GAO reports couldn't — force the government to modernize its data centers. Lucas: Stranger things have happened. But we'll see. For now, those eight thousand data centers are still humming along, forty percent of their servers doing nothing, burning through billions of your tax dollars. Luna: And we'll keep tracking it. Thanks for listening.