Retail colo’s half-rack tenants are the demand AI ignores

Down a service road at Datapoint Business Park in East London, the building Netwise uses as a data center sits among warehouse containers, distinguishable mainly by the security hardware around it. It is a long way from the futuristic AI campuses in current development announcements, and that distance is precisely what makes it worth underwriting. The tenants inside want a quarter or half rack of reliable, low-latency space to run the ordinary business applications their operations depend on.

Data Center Dynamics toured the facility with Netwise co-founder Matt Seaton, who put the client profile plainly: “There’s a certain client profile that we work with that just want a quarter or a half rack.” For those businesses, as DCD reported, edge facilities exist to house critical infrastructure off-premises, and proximity to the end user is the decisive requirement.

The wider market has gone the other way: hyperscale campuses now often exceed one million square feet, and the industry is developing racks that draw 1MW of power. Synergy Research Group counted 1,360 large hyperscale data centers at the end of Q4 2025 and said hyperscale facilities accounted for 48 percent of all data center capacity worldwide, a share it projects will reach 67 percent by 2031. The geographic concentration behind those numbers is familiar: Northern Virginia alone accounts for 11 percent of global capacity, and power availability is redrawing the list of viable sites.

That arithmetic makes the non-hyperscale colo slice — 20 percent of capacity today, per Synergy — look like a shrinking residual, but DCD’s tour coverage argues the opposite. Most businesses and everyday data demands do not require AI-training capacity; they need smaller, simpler computing close to users at low latency, and retail colo is the product built for that requirement. The requirement is broader than the current slate of million-square-foot announcements suggests.

The mistake would be to read hyperscale growth as evidence that small facilities disappear; the more accurate reading is that the market is splitting. At one end, giant campuses anchor themselves to a handful of enormous tenants and compete for grid capacity; at the other, retail colo sells space in small increments, where each signed half-rack is a revenue line on its own. The second model is less conspicuous, but its cash flow is not waiting on a future AI product cycle.

The capital hierarchy in digital infrastructure, as this publication has argued, increasingly separates contracted assets, which command infrastructure pricing, from merchant shells that must prove a tenant before they can borrow. Retail colo sits between those categories only because its revenue is granular: it lacks a single hyperscale anchor but is rarely an empty shell, and its cash flow is spread across many small contracts. That structure is a different risk from the AI-buildout binary — less dependent on the fortunes of one tenant or one project.

The power angle cuts the same way, now that large projects live or die by grid capacity and the interconnection queue has effectively become the lease in this sector. A facility like Netwise’s East London site needs only a modest grid connection because its load is small and predictable, and it can open without waiting on a transmission project. At a moment when power access is increasingly priced as its own asset, the low-power-density site carries an advantage the headline campuses lack: its financing is not hostage to a utility’s construction timeline.

The temptation in coverage of this industry is to frame retail colo as a junior version of the AI buildout, but that framing mistakes the customer. A company taking half a rack in East London is there to house applications that need to sit close to its users; the product is unglamorous precisely because its demand does not rise and fall with the next round of AI capex. That separation, more than the architecture, is the investment case for low-density retail colocation.

None of this argues for slowing the AI buildout; it argues for pricing humility within it. If hyperscale reaches 67 percent of capacity by 2031, the non-hyperscale colo business will still be selling floor space the same way Netwise sells it today, and the customer will still need half a rack, close to its users. Investors should not try to make that asset look like a small AI campus, because the customer needs a half rack in East London, and that need is not extinguished by the AI cycle.