I built a course around a segment of this industry I call the missing middle, the space between the hundred-plus megawatt hyperscale campus and the small edge box tucked into a telecom closet. Most conversations skip right over it, because a facility in the tens of megawatts is too big to be simple and too small to make a headline. This week gave me three fresh data points that the segment I have been teaching for over a year is no longer a gap in the market. It is becoming its own tier.
A 1.7 Megawatt Signal From Toronto
DeepInfra, an AI inference cloud that owns and operates its own GPU infrastructure, opened its ninth data center this month, its first outside the United States. The facility sits in Toronto, runs on roughly 1.7 megawatts of capacity, and hosts just over a thousand Nvidia GPUs dedicated entirely to running already trained models rather than building new ones.
Notice what that facility is not. It is not a hyperscale training campus, and it was never meant to be one, because inference workloads want to sit close to the people and businesses using them, and a company built specifically around inference chose a facility sized for that job instead of chasing the scale a training cluster would require.
A Twenty Megawatt Campus Built From Small Pieces
A very different company called Duos Edge AI is proving the same point from the supply side. Its Columbus, Georgia campus is being filled not by one giant tenant but by a series of smaller deployments stacked on top of each other, a 2 megawatt agreement here, another 10 megawatts coming online by the end of August, a further 10 megawatts planned for the fall. Its tenant on the latest deal, a firm called Nistar, exists specifically to package these smaller compute commitments for institutional clients who want AI infrastructure exposure without building a campus of their own.
Add those increments together and the site lands squarely inside the 10 to 40 megawatt range I have spent the better part of a year arguing is the industry’s most overlooked opportunity. Nobody announced this campus with a press conference the size of a hyperscale groundbreaking. It is simply getting built, tenant by tenant, at exactly the scale most communities, most utilities, and most regional developers can actually deliver.
Why Bigger Alone Was Never Going to Be the Answer
A recent survey of more than 500 enterprise technical leaders found that 86 percent of companies running their own GPU infrastructure are using half of it or less. Sit with that for a moment, because the industry has spent two years racing to build the largest possible facilities, while a large share of the compute already installed inside ordinary enterprises sits idle.
That is not an argument against building more capacity.
It is an argument against assuming every unit of new capacity has to arrive in gigawatt form to matter. A right sized facility that a regional developer can actually finance, permit, and fill with real tenants is worth more to the market right now than another headline campus competing for the same handful of hyperscale anchor leases.
Where This Fits Against the Rest of This Week
Everything else I wrote about this week, the institutional capital converging on infrastructure funds, the powered land deals being priced by the megawatt, the entitlement risk now baked into every site selection model, shares one characteristic. All of it is hyperscale flavored, describing a market built for the biggest players making the biggest bets.
The missing middle is the counterweight to all of that. It is where a regional developer, a smaller infrastructure fund, or a utility without room for a five hundred megawatt customer can still participate meaningfully in this buildout, using the same fundamentals of power, land, and demand driving the headline deals, just at a scale that actually clears the market today.
What This Means for Boards and Allocators
If your institution is only looking at the hyperscale end of this market, you are competing for the same small number of anchor tenants as everyone else with a checkbook. The next real deployment window is sitting in the 10 to 40 megawatt tier, where financing is more achievable, entitlement risk is lower, and regional operators like Duos are already proving out the model in real time.
Boards evaluating where to deploy capital next should be asking whether their pipeline includes anything in that range at all, rather than assuming the only rational move is chasing the same gigawatt campuses that seven infrastructure funds are already competing over. The middle of this market is not a consolation prize. It is where the next decade of participants actually gets built.
PLUS: If this sparked an idea or made you think differently about AI infrastructure, here are four ways we can continue the conversation.
1. Grab the guide.
Before advancing a data center site, there are 12 questions I make sure I can answer. I wrote them up here: [The 12 Questions Every Real Estate Professional Should Ask Before Advancing a Data Center Site].
2. Ask me a question.
If you’re looking at a deal, powered land opportunity, site, or AI infrastructure project and something doesn’t quite add up, reply and tell me what you’re seeing. Each week, I choose a handful of questions and share how I would think through them.
3. Clarify your positioning.
Trying to figure out where you fit in AI infrastructure? Whether you come from real estate, energy, construction, finance, telecom, technology, or another industry, reply with “Positioning” and a few lines about your background. I’ll share where I think your experience is most likely to create leverage as the market evolves.
4. Advisory Conversations.
I advise developers, investors, utilities, technology companies, and public sector organizations navigating AI infrastructure strategy. If you’re exploring growth, site selection, market positioning, partnerships, or board-level decisions, reply with “Advisory” and a brief description of what you’re working on.