This is an opinion, not a market report. There’s no single dataset that settles it, and smart people reading the same announcements land on different sides.
So, are AI data centers getting smaller on average, even while the biggest ones keep getting bigger? My answer is yes, for a specific reason. And no, in the one way that matters most if you’re putting money in. Both are true at once. If you’re looking at a deal, you need to know which one applies to the site in front of you.
The case for smaller
A data center’s size is capped by three things. How much power it can get. How long it takes to get permission to use that power. And how much of a community’s grid it can reasonably eat. All three have been getting harder at the top end and easier at the bottom.
A gigawatt-scale campus needs its own generation, a transmission-level interconnect, and a stack of emissions, water and land-use approvals that can take years. Most regional grids can’t absorb a load that size without new generation built just for it.
A site in the 10 to 15 megawatt range is a different animal. Lots of existing substations and distribution feeders can take that load with far less new infrastructure. The permitting path is shorter, because the environmental and interconnection review is lighter. And many more places, towns and secondary cities that could never host a gigawatt campus, can host something this size.
That asymmetry is the whole “smaller” thesis. If the long pole for a gigawatt site is years, and the long pole for a modest site is a fraction of that, and demand isn’t waiting for the slowest permit in the country, the rational move is to build more, smaller sites and stitch them together over high-capacity fiber.
A dozen 12-megawatt sites linked by fiber can start training workloads in a fraction of the time a single 150-megawatt equivalent campus takes to permit, with similar aggregate capacity. Under this view, distributed-and-linked isn’t a compromise. It’s the faster path to the same capacity, and speed is worth a lot when demand moves this fast.
The case against: hyperscale still wins on the numbers that matter
The counterargument isn’t that the mechanism above is wrong. It’s that one fact dwarfs it. The biggest campuses that do clear every approval are so large that one of them outweighs a very large number of small sites combined. And the companies with the balance sheets to sit through years of permitting have good reasons to keep doing it.
Economies of scale here are real and large. A big campus shares one set of power infrastructure, one cooling plant, one network core and one ops team across a huge pool of capacity. Spread the same megawatts across a dozen smaller sites and each one needs its own version of all that.
The fixed-cost layer (substations, cooling plants, fiber backbone, security, ops staff) gets paid once at a big site and many times over across distributed capacity. Once an approved gigawatt-class campus clears the queue, it can out-produce a lot of small sites on raw capacity and on cost per unit, for as long as demand justifies that scale. The years of waiting don’t vanish from this argument. They’re treated as a cost you pay once, for an asset that pays it back for a decade or more.
Put plainly: the smaller-sites thesis is about where more of the new projects are landing. The hyperscale thesis is about where most of the actual capacity ends up. Both can be true in the same market, because they answer different questions.
What this means if you’re a principal
If you have capital at risk in a specific deal, it doesn’t matter much which thesis wins in aggregate. What matters is which category your site is in, and what that means for diligence.
Interconnect and power contracts. A smaller site’s speed advantage depends on an interconnection agreement and a power contract that are actually signed, dated, and sized to the load. Not assumed. If the site is quietly sitting in the same kind of queue a big campus would be, just at a smaller scale, the whole appeal is gone. Ask for the dated paperwork, not the pitch.
Latency and fabric limits across a distributed “cluster.” A set of smaller sites linked by fiber is not the same as one big site for every workload. Training jobs that need tight, synchronous communication across every participant are sensitive to the latency and bandwidth between sites. On a single campus everything sits on one fabric, so that problem doesn’t exist. Before you underwrite a distributed footprint as one cluster, ask what the inter-site link actually is, what workloads it was designed for, and whether the economics assume a workload mix that can live with that latency.
Resale value. A smaller site built for one tenant’s workload is a less liquid asset than a hyperscale campus with power and land that almost any large operator would want. If the underwriting assumes the site can be re-let or sold at a decent price when the anchor tenant leaves, ask who the realistic buyers are at this scale and location. That pool is smaller for a 12-megawatt site in a secondary market than for a campus with gigawatt-scale power already secured.
Financing. Lenders price these two categories differently, because the risks really are different. A smaller site’s lender is underwriting a shorter, more certain permitting and build timeline, against a thinner pool of refinancing comparables. A hyperscale lender is underwriting a longer, riskier approval timeline, against an asset class with deeper comparables and usually stronger anchor tenancy once built. Neither is automatically the better credit. Ask which one the financing terms in front of you were priced against.
Where I land
I think both halves are true and will stay true for a while. The count of new sites is tilting toward smaller, faster-to-permit footprints, because power and permitting limits at the top end are real. And the share of total capacity sitting in a small number of hyperscale campuses isn’t shrinking, because the economics of scale at the top end are also real.
So don’t pick a side in the debate. Figure out, site by site, which category your deal is in, and diligence it against the risks that apply to that category, not the other one.
Once you know the category, the arithmetic for what a cluster of a given size should cost to run is in what it actually costs to run a GPU cluster. And the questions to ask about whoever stands behind a site, large or small, before money moves, are in how do you vet a GPU cloud provider before you wire the deposit.