Off the Run

A Data Center Is a Data Center Is a Data Center

• Not all data centers are the same
• Training, inference and interconnection all carry different risks
• The safest assets may be the ones the digital economy cannot easily leave

It’s all people are talking about these days. Build them. Ban them. How many more do we need? When will we have too many? It’s just like the internet bust. It’s just like canals. It’s just like railroads. It’s just like tulips. It’s not. The data center sits at the heart of the latest tech boom, one moving faster than almost anything I have seen. And like every boom before it, there will be winners and losers. The trick is understanding what exactly you are underwriting and how it fits into the much larger geopolitical, economic, and physical supply chain. Because a data center is not really a single business model. It is a building that can house several completely different ones.

All data centers are not created equal. Yes, they all require land, power, water, chips, and ideally long-term clients. They just don’t all do the same thing. There are training data centers where you create the AI model. Inference data centers that deliver that intelligence. And then there are interconnection data centers that keep the digital economy functioning.

The data centers used for training are most likely the ones you see in the headlines, associated with gigawatts of both power and compute. You need enormous campuses because training a frontier model requires hundreds of thousands of chips working together at the same time, with huge amounts of power, cooling, and network capacity. Scale is the feature. These facilities are often backed by 10- to 15-year contracts with (at least for the moment) investment-grade hyperscalers that see control of compute as the best way to make sure they stay out in front in the AI race. They also don’t need to be anywhere in particular. Wherever the cheapest land and fastest path to power exist is where you want to build, because time to power is money. The risk is not that one model finishes training and everyone goes home. It is that training demand comes in waves, the chips become obsolete quickly, and a facility built around one tenant or generation of compute may not be equally valuable for the next. Who replaces the chips? Who moves in if the tenant leaves? Who pays off the debt?

Inference is when you put the models to work. Every time someone asks Claude or ChatGPT a question, generates an image, or embeds a model into a company workflow, that is inference. This is where proximity starts to matter. Companies care where their data goes (and sometimes the law does too). And some decisions cannot wait for a request to travel to a giant data center and back. A camera on a factory floor may have milliseconds to recognize a worker in the wrong place and stop a machine.

Unlike training, inference will not all live in giant campuses wherever land and power are cheapest. Some will stay there. Some will move into smaller data centers closer to the customer or the company’s data. And the decisions that need to happen fastest, or involve the most sensitive information, may happen on the phone, car or machine itself (some already do). The underwriting risk is that no one really knows how much compute will live in each place, while models and chips keep getting more efficient. How much stays in the mega data center? How much moves closer to the user? And what happens when the device in your hand can do more of it itself?

So, if training facilities manufacture the product and inference facilities deliver it, what’s left? Everything that is probably most important to you. What we call interconnection data centers. The ones supporting the network when you watch Netflix, order something on Prime, host a Zoom call, tap your card at dinner or pay your Amex bill from your phone. These facilities are not valuable simply because they are filled with servers. They are valuable because of who is connected inside them. Networks, cloud providers, banks, retailers and customers all meet there. Rather than sending traffic across long distances and hoping all the handoffs work, companies establish direct physical connections inside the building. Those connections are faster, more reliable, and much harder to move than a rack of servers. It’s kind of like a WeWork, except the other tenants are the reason you signed the lease. You don’t awkwardly avoid eye contact with your neighbor. You connect directly to them to run your business.

And every new customer makes the building harder for everyone else to leave. The underwriting question is where the value actually sits. In the chips? The lease? Or the network of relationships inside the building? The newest chips will become old chips. The next model will need a different amount of compute. Someone will still be paying rent on the room where all of it meets. The building is the easy part to see. From the road, they all look the same.