Drive through the outskirts of any major cloud region and you can spot them: long, low buildings with fresh concrete, generator yards fenced off, and not a single light on inside. They look like data centers. They were financed like data centers. For now, they are the physical embodiment of a bet that hasn’t paid off yet. Business Insider recently flagged this as the AI data center boom’s ghost problem, and the phrase sticks because it captures something the industry is only starting to reckon with. We are building capacity faster than we can fill it, and the gap between the two is starting to show up on balance sheets.
The mechanics are easy to understand once you strip away the marketing. The past two years have been defined by a scramble for compute. Hyperscalers, neoclouds, and a wave of well-funded startups signed leases and broke ground on facilities years before they had tenants lined up, on the assumption that demand would keep climbing. Some of that was rational. AI training workloads genuinely exploded. But not every project was anchored to a signed contract, and not every operator had the discipline to wait.
Now the bill is coming due in installments. Several operators have quietly paused construction or slowed fit-outs, and a handful of planned campuses have been shelved entirely. Others are running facilities at occupancy levels well below what the original pro formas assumed. The hardware is the expensive part. A single rack of high-end accelerators can cost more than the building that houses it, and those chips depreciate on a brutal schedule. An idle GPU is not a neutral asset. It is a depreciating liability with a power bill attached.
There is a second ghost problem, and it is less visible but arguably more corrosive: the infrastructure that was built for one generation of hardware and is now awkward for the next. Cooling designs, power delivery, and rack densities were sized for chips that are already a generation or two old. Retrofitting a facility to handle liquid cooling or the power densities that current accelerators demand is not a simple upgrade. In some cases it is cheaper to walk away and build fresh, which leaves the original shell stranded. The industry has a word for this when it happens to software. Technical debt. In real estate, it just looks like an empty building.
The comparison to the broader tech debt conversation is hard to avoid. The Japan Times recently ran a piece on an AI-found flaw signaling chaos ahead, and the underlying theme is the same one haunting data center operators: systems built fast, on assumptions, accumulate liabilities that compound quietly until someone trips over them. AI is very good at finding those liabilities in code. It is less good at finding them in a signed lease or a cooling contract.
None of this means the boom is a bubble about to pop in a single dramatic moment. Demand for inference is real and growing, and the largest operators have the balance sheets to absorb a few empty halls while they wait for workloads to catch up. The hyperscalers can treat a stranded facility as a rounding error. The danger sits with the second and third tier: the neoclouds that raised debt against future capacity, the regional operators who bet on a single anchor tenant, the developers who assumed financing would stay cheap. When capital gets more expensive and tenants get pickier, those are the projects that go dark.
The more interesting question is what happens to all that stranded capacity. Some of it will get repurposed for inference at lower margins. Some will be sold at a discount to operators with better demand. Some will sit empty long enough to become a cautionary tale. The industry has been here before with fiber in the early 2000s, and the lesson from that cycle is that the infrastructure eventually gets used. It just takes longer than the people who financed it expected, and the people who financed it are not always the people who survive.
What makes this cycle different is the speed. Fiber sat dark for years because the applications that would consume it hadn’t been invented yet. AI capacity is being built against demand that already exists but is concentrated in a handful of customers who are themselves still figuring out their own economics. That concentration is the real ghost in the machine. If a few large tenants slow their spending, the facilities built to serve them do not gracefully downsize. They just go quiet.