Hunter Newby, owner of Newby Ventures discusses whether the industry is ready for the impact of AI

---

Author & Originally Published: Hunter Newby @ CapacityMedia.com

During the process of learning how to ride a bike, the rider often falls off the bike. To mitigate this initial risk, training wheels were invented. Eventually, the training wheels must come off. The first few moments of riding solo are exciting, but even the most experienced riders fall.

The data centre industry is going through a similar experience. There has not been this much excitement since the dot-com bubble.

Looking back, it is not that there wasn’t demand. It was that the demand required infrastructure to be built to support it and that all takes time and money.

High-yield debt to an IPO was the favoured path to an exit, but public companies have this thing called ‛Quarterly Earnings’. Since everything always takes longer and costs more, many dot-com companies were financially restructured.

That falling domino led to the fall of the telecom companies that were building long-haul fibre routes across the US and ‛web hosting facilities’, now known as data centres. This was known as the ‛telecom bubble’.

Today, things are different, yet they are the same. What is the same is that AI and GPUs are like the web and servers. They need data centres to live in and fibre to route over. But the scale is different. The power required for large-language-model (LLM) and machine-learning (ML) AI is greater than what a web hosting facility required.

There are numerous headwinds in the AI data centre industry today. These include supply-chain lead times and the high cost and structure of the capital required to build. The combination of all of these factors leads to delays, and increased costs lead to missed interest payments and restructuring.

Beyond all that, there is this new issue of power. Previously, a 20-megawatt (MW) data centre was considered large. Today, the mid-range sized data centre for LLM or ML AI is 100 to 250 MW. These power demands have exceeded the utility companies’ ability to deliver.

The credit risk today is not with the off-taker. The hyperscaler turned AI company is an AAA credit. The risk is within the structure of the deal. The hyperscalers do not pay anything until the data centre is complete. The data centre provider takes the risk.

For the data centre to be deemed complete, all the parts and labour must be delivered on time, but the power must also be delivered from the utility. Otherwise, there are financial penalties for missing dates and at a certain point, the hyperscale or AI off-taker can walk away, leaving the project unfinished and without a buyer. All the while, the interest payments must be made.

So to narrow it down, what happens when a data centre provider gets a credit facility based upon a creditworthy off-taker agreement and a letter from a utility that says power can be delivered and, based on that, land is secured and capital is deployed to build — but months later, the utility determines that it cannot deliver the power?

‛Behind the meter’ (BTM) is an industry term for self-generated power. BTM was very much a ‛natural gas as a fuel’ concept, also known as a ‛micro-grid’, but the concept is certainly not limited to any fuel type. It is all about autonomy and control.

The latest and greatest BTM fuel source is nuclear power. Buying or securing the rights to use existing nuclear power plants had its moment, but there aren’t many of those deals to do, so now it is on to small nuclear reactors (SNRs). SNRs come with regulatory permits and approvals, which will cause more time delay than what the hyperscalers wish to tolerate.

Maybe this is why the Department of Defense took the hyperscalers and AI companies to the White House in September 2024 to discuss fast-tracking data centres.

AI is moving fast, like an inexperienced rider on a bike. Just like in the dot-com bubble, AI requires infrastructure, which costs a lot and takes years to build.

So whether it’s quarterly earnings or monthly interest payments, infrastructure takes as long as it takes — and usually that is longer than planned, which costs more. The AI training wheels are now off. We shall see who makes it down the road and who falls.