Building Sustainable AI Infrastructure: Lessons from Infra/STRUCTURE 2025

Infra/STRUCTURE is the annual Structure Research event focused on the infrastructure part of the Internet – the people, companies and tech behind the carriers, data centers, servers and service providers that make the Internet work. 

I’ve been to a few of these events over the years. Each one has been more focused, more enlightening, and more useful than the last. So, how good was infra/STRUCTURE 2025?

Infra/STRUCTURE 2025 was great.

It was an excellent conference. Infra/STRUCTURE is really focused on industry direction, industry partnerships, and especially this year… what the heck is happening in AI and AI infrastructure. 

So, what IS happening in AI infrastructure? 

AI was what EVERYONE I spoke to wanted to talk about. Carriers. Hardware vendors. Data center operators. Wannabe data center operators. Wannabe GPU owners. I planned to join many of the keynote sessions and just didn’t have time. Everyone wanted to talk about how to enable AI. I was on a panel called “Managed Infrastructure at a Crossroads” and even that turned into a discussion about AI. 

For anyone in the infrastructure business, this won’t be a surprise, but what these conversations did reveal to me was a quite surprising delta between the hype about AI, and what the real infrastructure players actually understand and are actually doing. Here’s my quick summary and take-aways from the event.

1.      Half of everyone wants to join the AI gravy train       

If we look at the Venn diagram of the conversations I had, maybe 50% of people were looking for a way to get in on the action. I spoke with people who sold their businesses and invested everything in GPU. I heard stories about people who re-mortgaged their houses to invest in GPU. There are major finance companies now looking at GPU purchases the same way they’d look at buying futures. 

There are people with old-school data centers, asking who can help them find power for next-gen GPU servers. There are people with power and land looking for partners to help build a data center. 

And, perhaps most worrying of all, people who have already invested in the land, power and GPU data center facilities, asking who can help them find customers for what they’ve just invested seven- or eight-figure sums in. 

2.                  Everyone else is talking about the bubble bursting

What about the other 50%? They were talking about how crazy this is. How the AI bubble is going to pop - serious cynicism about a) the likely real demand (which nobody knows) and b) the crazy levels of investment in high-end data center infrastructure. 

In the classic infrastructure space, there are many hyperscalers investing literal billions in “next generation” multi-gigawatt data centers, which should be ready in, oh, two or three years from now.  Those companies are probably too big to fail, and of course their investments are driving the hype-train. 

The naysayers (who are really just realists) are seeing an imminent future a bit like the dot.com bubble: when all these mega-projects come on line, will there really be sufficient demand? If not, what happens? 

3.                  Smaller players aren’t too worried about it, though

In the real world outside the stock markets, where business just has to get done – and this is across people I spoke with in networking, storage, compute, GPU, real estate, energy, everything - things must come back down to earth. The bubble may burst, but then the real work begins.

There’s a kind of infrastructure under-class that is actually about a third of the industry, that everyone ignores - at least in the mainstream media and analyst reports. It’s the smaller players in the infrastructure space who are not working on the multi-billion-dollar, multi-gigawatt level. They’re focusing on smaller projects, maybe a modest 10MW or 50MW at a time. They’re building their infrastructure and their business incrementally, and they are building in a smart and sustainable way.

One of my colleagues with me at infra/STRUCTURE calls them “collegiate” players, and that’s exactly right – the same capability and skills as NFL, just without the hype and price-tag.

They are also building for AI, but the key difference is, they actually have customers. This is the approach we’re taking at Patmos, too – finding smart ways to build AI / GPU infrastructure that isn’t about who has the biggest cluster, or the most VC investment – it’s about building in a right-scale way to deliver the capacity that your actual customers need right now. Doing it fast, in months not years, and then doing it again when the demand is there. 

The bubble might burst, but not everyone will fail

For me, this was the key take-away from infra/STRUCTURE 2025. That there is a real understanding among the grass-roots infrastructure people and companies, about what’s happening, what is a little crazy about it, and why sticking to what we know will see us through. 

After all, the Internet wasn’t built by four or five companies with billionaire owners and multi-billion-dollar market caps. It was built by thousands of smart, smaller businesses who had a vision, and the same smart and smaller businesses are now building the AI infrastructure that is going to stay afloat even if or when this bubble bursts.