Regulators have spent eighteen months deciding who pays for grid upgrades tied to AI load growth. Almost no one has decided who is qualified to build the physical connection between a data center and the grid — and hyperscale delivery leaders like Sweta Dutta say that gap, not equipment or permitting, is now the real constraint on when new capacity comes online.
For most of the last decade, a hyperscale data center's timeline was set by construction: design a shell, procure long-lead electrical equipment, build, commission. That sequence is well understood industry-wide and has historically been the thing owners managed to. It is no longer the constraint. In most of the markets carrying the current wave of AI-driven demand, the date a utility can actually energize a site is fixed years before construction starts, and no improvement in build speed moves it. Interconnection, not construction, sets the outer date on a hyperscale project — a shift the industry itself has been documenting for several years, through queue data, utility filings and a wave of new state and federal proceedings.
That shift is well covered. Less examined is a narrower, more technical claim: that within interconnection, the specific piece of scope now deciding project timelines is the physical interface between a utility's infrastructure and a data center's own — and that responsibility for building it has moved to organizations that are largely unstaffed to do it. That is the argument made by Sweta Dutta, a construction delivery manager who has spent fifteen years building hyperscale data center capacity on both the contractor and owner sides.
Dutta's career in construction delivery dates to January 2011, when she took a regional leadership role at PSA Design Studio in India, overseeing multi-project delivery for roughly five years. After earning a master's degree in Construction Engineering and Technology from Arizona State University, she moved to the contractor side of the U.S. industry, working in program-level project management at DPR Construction on a greenfield hyperscale campus in Mesa, Arizona, built for Meta at a scale exceeding $1.5 billion, inside Salt River Project territory.
Since January 2024, Dutta has worked on the owner side, as a Construction Delivery Manager at Amazon Web Services, where she leads full campus and regional program delivery across five builds and roughly $1.4 billion in active capital expenditure, spanning five Availability Zones in two North American regions — Columbus, Ohio, under the mid-Atlantic grid operator PJM, and Canton, Mississippi, a materially different utility and regulatory environment. She leads a matrixed program team of more than seventy people across nine functions, including engineering, procurement, construction, commissioning, operations, controls and security. Across her career, Dutta has worked across roughly $3.2 billion in data center and industrial construction.
That combination — contractor and owner, and three of the country's most power-constrained markets, two of them run concurrently — is what put Dutta in a position to see a pattern that queue statistics alone do not capture: which specific piece of the interconnection process actually governs a project's schedule, and who is being asked to own it without the engineering background to do so.
Dutta spoke with this publication in a personal capacity, drawing on fifteen years of industry experience rather than on the practices of any current or former employer.
Between a utility's point of demarcation and a data center's first electrical panel sits a band of work that includes the interconnection facility, the customer substation, medium-voltage distribution and protection, and metering and telemetry back to the grid operator. Utilities historically built this themselves or sold it to owners as a turnkey package. To compress schedules, developers are increasingly building it themselves — a shift that, according to Dutta, hands owners a scope requiring protection and coordination engineering, not construction management, without necessarily giving them the staff to match.
“General contractors are not being replaced by power specialists. That framing sets up a competition that is not happening,” Dutta says. “What is happening is that the owner has absorbed a scope that neither the GC nor the utility previously owned. GCs still build the buildings, and they build them well. The utility still runs the transmission system. The new capability sits in the owner’s delivery organization — and for organizations scaling into owner-side delivery quickly, it is often being staffed out of construction rather than out of power engineering.”
Dutta is careful to draw a distinction. Established hyperscale operators, she says, have generally built deep power engineering benches over a decade or more of owner-side delivery. The capability gap she describes is concentrated among newer entrants — independent developers, converted mining operators and colocation firms scaling into hyperscale for the first time — where the interface scope has arrived faster than the in-house expertise to engineer it.
Dutta stops short of declaring construction a solved problem. “I would resist the phrasing that construction ‘is not’ the bottleneck,” she says. “Once power is secured, construction is emphatically the bottleneck for the last eighteen months, and treating it as a solved problem is how teams miss dates they already paid for. The honest version is that the binding constraint moved upstream — not that it disappeared.”
Two assumptions common among developers, in Dutta's account, do not hold up. The first is that equipment lead times are the primary shortage. Long-lead transformers and switchgear are a documented constraint, but equipment, she argues, is a queue that capital and lead time can move through. “The scarcer input is people who can sit in a utility interconnection study review, read a relay coordination study, and also hold a construction schedule — and there is no lead time you can pay to shorten on that,” Dutta says. “We are short of protection engineers and commissioning authorities with grid-side literacy far more acutely than we are short of switchgear.”
The second is that prefabrication resolves schedule risk broadly. Dutta has direct experience with the technique: she led a pilot that moved a repeatable electrical-room scope from job sites into a controlled shop environment, a change that affected not just assembly speed but where labor was deployed and how much rework was caught before commissioning rather than after. But she draws a sharp distinction between what prefabrication fixes and what it does not. “Prefabrication fixes labor availability, quality consistency and site congestion. Those are worth having. It does not move the energization date by a single day,” she says. “If power is your critical path, prefab makes you finish earlier and wait longer.”
The commercial side of this problem has moved through regulators quickly. The Federal Energy Regulatory Commission directed PJM Interconnection, in a December 2025 order, to establish rules for data centers co-located with generation, then issued a partial acceptance of PJM's compliance filing in April 2026. In June 2026, FERC opened Section 206 show-cause proceedings on large-load interconnection and behind-the-meter generation covering every organized grid operator in the country. The Southwest Power Pool approved its own high-impact large-load review process in January 2026. In Texas, the Public Utility Commission approved a new ERCOT interconnection framework in June 2026 that studies large loads in batches rather than individually, an explicit response to a per-project study model that had been overwhelmed by demand.
The engineering side of the problem has moved more slowly. The North American Electric Reliability Corporation issued voluntary guidance in May 2026 on ride-through performance, protection coordination and model validation for large computational loads. That guidance followed a Level 3 “Essential Actions” Alert — NERC's highest-urgency category of notification — issued after data centers repeatedly dropped more than a thousand megawatts off the bulk power system during grid disturbances the equipment was expected to withstand without tripping. FERC has since directed NERC to file mandatory reliability standards for computational loads by the end of 2026, with a second phase due in early 2027 — meaning owners will not be operating under mandatory physical-interface standards until at least 2027, even as the commercial rules governing cost and access have already been substantially rewritten.
“FERC has now told us who pays, and NERC is being told to work out who is responsible,” Dutta says. “Neither answer helps the team standing in front of a switchboard next Tuesday, and that is where this actually gets decided.”
Dutta's assessment of the interface problem is informed by ongoing research alongside her delivery work. She is currently developing techno-economic modeling of behind-the-meter generation sized for large, near-flat data center loads — a load profile she says behaves differently from the assumptions embedded in most standard generation-sizing methods, because a flat load offers little of the cost arbitrage those methods are built around. A second line of her research addresses structural floor loading for AI-density rack deployments, where design assumptions calibrated to racks in the 8-to-15-kilowatt range are being applied to deployments several times that density, arriving with liquid cooling infrastructure that original loading calculations did not anticipate.
Asked who should be building the power interface for AI data centers going forward, Dutta's position is direct: “The owner should build it, and the owner should staff for it. A lot of developers have taken on the power interface as a scope without taking on the engineering capability that scope requires — they have bought the risk from the utility without buying the expertise that used to come with it.”
Her recommendation follows from the pattern she describes across the industry: owner-operators, she says, should be hiring power systems engineers directly into the construction delivery organization, reporting through delivery rather than retained as design-phase consultants. “The interface problems surface during construction and commissioning, which is exactly when the consultant's scope has ended,” Dutta says.
Manufacturing capacity for long-lead equipment is expected to expand, and the tariff and cost-allocation disputes now before regulators will eventually be resolved. The harder constraint, in Dutta's assessment, is workforce: the pool of engineers able to hold a construction schedule while also reading a relay coordination study is not expanding at the pace new capacity is being added, and the mandatory reliability standards due to take effect from 2027 will add demand for that same skill set.
“We have spent three years solving for how to get power,” Dutta says. “We have spent almost none solving for who connects it, and that is the bill coming due.”