Digital Realty breaks interconnection records on agentic AI: the bottleneck is no longer the chip, it's power

By Zendoric · October 1, 2026.

Digital Realty Trust (NYSE: DLR), a data center owner and operator present in more than 50 markets, has strung together three consecutive record quarters in its small-deployment business, the 0-to-1 megawatt segment (MW, the unit that measures the electrical power a facility consumes). The driver, as chief financial officer Matt Mercier explained at an RBC session on communications infrastructure, is agentic AI: systems that carry out tasks autonomously by chaining calls to other models, tools and data sources. Mercier described it as a technology still at a very early stage, but one that is already leaving its mark on the numbers.

What is rising is not just raw compute, but interconnection: the cross-connects that link different customers' servers within the same building, the bulk fiber connecting campuses to one another, and ServiceFabric, Digital Realty's own connectivity platform. According to Mercier, deployments that used to hover around 300 kilowatts are now asking for between 500 kW and 1 MW, and requests of 1 to 5 MW are already common among enterprise customers and service providers. It is a clue to how agentic AI actually works in production: less a single giant model training in isolation, and more a set of systems calling each other at low latency that need to be physically close and well connected.

Digital Realty today has 3 gigawatts (GW) of operational capacity and another 1.4 GW under development —a jump that, according to Mercier, could expand the company's portfolio by close to 50% in two years— plus 7 GW of land with potential for the next two to five years or more. Of that land bank, between 3.5 and 4 GW already have signed power supply agreements; the rest is, in the executive's own words, the hard part: 'there has never been a better time to be in this industry, but it has probably never been harder to secure power and the related capacity,' Mercier said. It is the phrase that sums up the sector's real state at the end of 2026: the bottleneck is no longer making more chips, but finding the electricity to switch them on.

That scarcity is supporting Digital Realty's prices rather than hurting them. The company posted renewal spreads —the price increase on renewing a contract relative to the previous one— of more than 60% on contracts above 1 MW, and described those in its smaller segment as 'very healthy'. Mercier expects net income growth of 10% this year, similar to last year's, and expects to keep up that pace for several more years. To fund the expansion without driving up debt, the company has diversified its sources of capital: last year it launched its first closed-end private equity fund, which has already raised more than $3 billion in capital, and placed a 600 million Swiss franc bond with a 2% coupon.

One detail deserves attention: 100% of the $1.4 billion in contracts signed in the first half of 2026 came from investment-grade hyperscale customers, that is, large established clouds with solid credit ratings. Mercier acknowledged that the company has been deliberately cautious with so-called 'neoclouds' —younger, more leveraged AI compute providers, a segment that has accounted for much of the circular financing that has raised doubts in the sector over the past year—, though it has accepted some small, highly interconnected workloads of this kind at specific facilities. It is a discreet way of saying that not all AI compute demand deserves the same level of trust.

Geographically, Digital Realty has tilted its development pipeline toward the United States, drawn by the scale of AI and cloud demand, though it maintains projects in Europe, the Middle East, Africa and Asia-Pacific, with recent expansion in Johor (Malaysia) and new operations in Japan. On regulatory uncertainty and local opposition to data center construction, Mercier was clear: the company has no plans to scale back its pipeline.

This piece is not about a new model or an AI lab: it is about the physical layer that makes everything else possible, and that is where its value as an indicator lies. While the public debate focuses on which model reasons better, institutional capital —private equity funds, Swiss bond buyers— is betting on whoever owns the land, the cable and the power supply contract. It is the classic thesis of selling shovels in a gold rush, with one peculiarity: here the shovel has a bottleneck of its own, and it is the electrical grid.

In the short term, that bottleneck carries a cost that should not be played down. The accelerated construction of data centers competes for land, water and megawatts with communities and other industries, fuels the local opposition that Mercier himself admits to without committing to address, and puts pressure on power grids that in markets such as Northern Virginia, Europe or Singapore have been under capacity strain for years. That cost is not paid by Digital Realty alone: it ends up being felt, in part, by household and industrial electricity users in those same regions if generation does not grow at the same pace as compute demand.

In the long run, however, this same infrastructure is what underpins the thesis we defend at Zendoric: the abundance AI promises does not arrive by decree, it is built megawatt by megawatt and cross-connect by cross-connect. The more interconnection there is between specialized AI systems, the cheaper it will be to run the agentic workflows that today automate administrative tasks and that, over time, could accelerate everything from drug discovery to more efficient management of the power grid itself. Therein lies the paradox worth pointing out: the technology that in the long term could help solve energy scarcity depends, right now, on winning the race for available electricity. Whoever resolves that bottleneck first —with nuclear, renewables or more efficient chips— will set the pace at which the rest of the industry can advance.