The industry has spent this year celebrating a number that should be keeping it awake.
CBRE puts vacancy across North America's primary data center markets at 1.4 percent, a level this market has never seen before. Northern Virginia came in at 0.2 percent after absorbing another 467.7 megawatts of net demand. Hillsboro sits in roughly the same territory. Across every primary market in the country, CBRE counts less than 1,500 megawatts still available for pre-leasing, and more than 80 percent of the record 7,481 megawatts under construction has already been committed to a tenant who has no intention of handing it back.
Read those figures as a demand story and you are right. Demand is extraordinary. Capital is abundant. The build is moving faster than almost anyone predicted three years ago.
There is a second story inside the same numbers, and it has very little to do with leasing velocity. It has to do with what happens to a machine that is never allowed to stop.
Full capacity is a statement about maintenance
A market with 1.4 percent of its capacity available to lease is a market with no slack in it. Slack is the raw material that maintenance has always been built out of.
The entire discipline of preventive maintenance rests on a single assumption, which is that you can eventually take the thing offline. It assumes a window. A Sunday. A shoulder season. Some stretch of low demand when you can valve off a cooling loop, pull a chiller for a compressor rebuild, drain a system, inspect a bearing, or replace a seal in the weeks before it lets go on its own. That assumption is how the industrial world has kept rotating equipment alive for the better part of a century.
Across most of the American market, that window has closed.
When every megawatt in the building is leased and the tenant's training run has been scheduled for weeks, no operator on earth volunteers to take the mechanical plant down for elective service. The request moves to next quarter. Next quarter the site is just as full. Eventually the deferral hardens into a policy that nobody ever consciously wrote down.
What that produces, quietly and across thousands of megawatts, is an industry that has converted itself to run-to-failure without ever deciding to. No operator here is negligent. No engineering team has abandoned its standards. The outcome falls straight out of the arithmetic of being full, which is a far more uncomfortable thing to sit with, because arithmetic does not respond to good intentions.
The same arithmetic works on the equipment from the other direction. A full site runs its chillers, pumps, compressors, and cooling loops at maximum duty cycle more or less continuously, and continuous duty at high load is the exact condition under which bearings wear faster, lubricant degrades sooner, and seals give up earlier than the design tables promised. The machines are aging faster at the precise moment the industry lost its ability to service them on a schedule.
The replacement part now arrives in six months
There was a time when a failed chiller was an inconvenience. You carried a spare, or your vendor carried one, and the outage lasted an afternoon.
That same CBRE report describes mechanical equipment shortages pushing ready-for-service dates from a matter of weeks out to several months, and past six months for liquid cooling conversions. Those lead times change the character of a mechanical failure completely. The clock now runs on where the component is. On a boat. In a queue behind eleven other hyperscale orders. Waiting on a compressor that a manufacturer in another country has not built yet.
A capacity outage measured in quarters is a different business event than one measured in hours. A single unplanned chiller failure affecting 5 megawatts of critical load can trigger SLA penalties in the range of hundreds of thousands to several million dollars per month, depending on contractual terms. That figure does not include replacement equipment costs, emergency engineering, or the tenant who does not renew. It reprices contracts. It moves tenants. In a market with 1.4 percent vacancy, it sends that tenant hunting for alternative space that does not exist.
Scarcity has already priced all of this. Asking rates for large deployments rose 14.5 percent in Atlanta and 19 percent across New York and New Jersey in the first half of 2026 alone, and the price rose in every tier CBRE tracks. Downtime used to be priced against the cost of a slow website. It is now priced against the scarcest industrial commodity in the country.
Redundancy has stopped being a substitute for asset reliability
I have written before about redundancy as the tax the industry pays for unreliability, and about how buying more machines is the only way to make unreliable machines look reliable. Full capacity is where that strategy runs out of road.
Every layer of redundancy assumes four things are available at once: spare equipment, spare space to install it, spare power to energize it, and spare time to commission it. A market at 1.4 percent vacancy, with six-month equipment lead times, interconnection queues stretching well past the point of usefulness, and local community opposition that CBRE now rates on par with power procurement, has run short of all four simultaneously. You cannot purchase reliability from a market with nothing left to sell.
One lever still moves. It is the asset reliability of the equipment already bolted to the floor.
This is the shape the whole physical economy is taking
The pattern is considerably larger than this one asset class. Every automated physical system going in right now is designed and financed on the assumption of continuous operation. Lights-out manufacturing. Automated distribution. Robotic fulfillment. The coming wave of physical AI. The capital only pays back if the asset almost never stops, which means the maintenance window is quietly disappearing across the entire physical economy, for the same reasons it has already disappeared here.
We are building a world of machines too valuable to turn off. We are still maintaining them with a model that requires turning them off.
There is a symmetry in that I find genuinely striking. The buildout the world calls artificial intelligence is going to depend on industrial AI to keep itself alive. The models being trained inside these buildings are only ever as available as the chillers cooling them, and the discipline that keeps a chiller available is the same asset reliability work that has kept refineries, mills, and mines running for decades.
What we believe
Asset reliability has been the whole of Tractian's work for years. In practice that has meant pointing industrial AI at a narrow and deeply unglamorous question: what is this particular machine about to do, and how much time is there left to act on it. We built that capability inside heavy industry, on the same motors, pumps, compressors, and chillers that now sit at the center of every gigawatt campus being energized this year.
Why that matters more in a full market than a loose one is simple. A machine has no need to be offline for its condition to be known. The current a motor draws carries that information. So does the way a bearing vibrates. Both carry it continuously, whether or not anything in the building is trained to interpret it. Interpretation is the day job of industrial AI on a plant floor, and it turns maintenance from something scheduled around demand into something aimed at the single asset that needs attention, inside whatever narrow window an operator can still find. Continuous monitoring is what stands in for the outage window scarcity has taken away.
History will measure this build on utilization rather than installation, and utilization at these densities becomes an asset reliability problem long before it becomes a compute problem. Full capacity has been reported all year as the industry's great achievement, and by every commercial measure it is one. It is also a warning about the price of the next unplanned mechanical failure, in a market that has left itself nowhere at all to put the load.
Every one of those machines is already telling us what is wrong with it. The only question left is whether anything in the building is listening, and whether that signal reaches an operator with enough lead time to act.