The scarcest resource for AI companies is quietly changing its name. Over the past two years, everyone watched chip manufacturing capacity; recently, however, a wave of studies posted densely to arXiv points to another bottleneck: **electricity—and the waiting time to connect to the power grid**. To connect a data center to the transmission grid, operators must first apply to the grid operator, undergo review and approval, and only then begin construction—the so-called “interconnection queue” has already stretched to **over five years** in some parts of the United States. Meanwhile, operational data centers now reach hundreds of megawatts in scale, and announced projects are entering the gigawatt range. To wait for the grid—or not to wait? This question is rewriting the economics of AI infrastructure.
## Pricing "Time" into Electricity Costs
A technical-economic analysis posted to arXiv in November delivers a direct, quantified answer. Researchers used an open-source framework to compare four power supply options for an AI data center at the same location: connecting to the public grid; AC-coupled photovoltaics (PV) plus battery storage; DC-coupled PV plus battery storage; and an islanded natural-gas power plant with diesel backup (i.e., operating entirely off-grid, self-supplying). All four options deliver identical energy services and reliability levels; the model incorporates climate-driven cooling, multi-stage power conversion, and physical battery aging.
If only traditional accounting is used—i.e., average levelized cost of electricity (LCOE) over the system’s lifetime—microgrids combining PV and storage (small-scale, self-contained generation-and-consumption systems capable of independent operation) are almost never the cheapest option; the grid or gas-fired generation appear more economical. But the researchers introduced a "**speed-adjusted energy cost**": computing power idle during construction—unable to generate revenue—is itself a loss, and should be monetized and added to total cost. With this adjustment, the picture flips: **DC coupling** (where both PV and batteries connect via direct current, eliminating inverters and improving conversion efficiency) reduces lifetime energy costs by an average of 18% versus AC coupling—equivalent to about $0.09 per kWh; and because microgrids can be built much faster than waiting for interconnection, factoring in foregone compute yields **DC-coupled PV-plus-storage microgrids as the lowest-total-cost option at 85% of locations across the continental U.S.** Even if gas turbine delivery accelerates dramatically, this share remains at 55%. The paper’s conclusion in one sentence: deployment speed—not unit energy cost—determines the economically optimal path.
## Putting Servers Inside Factories
If building your own power plant isn’t appealing, there’s a second path: simply avoid constructing large-scale data centers altogether. Another study posted to arXiv the same month proposes installing AI servers directly inside **existing factories**. The logic is straightforward: factory grid connections are sized for peak production loads, leaving substantial unused capacity during off-peak hours; moreover, AI load peaks often do not coincide with production load peaks.
Researchers modeled 31 Chinese manufacturing sectors, pairing forecasted 2030 AI demand with empirically measured hourly production loads, and compared four deployment strategies: on-site installation per factory; centralized hosting at a single group facility; task routing across multiple factories within the group; and building new large-scale data centers. Results show: new large-scale data centers require **1,488 MW** of newly added interconnection capacity; placing servers inside factories and scheduling them according to production rhythms requires only 556–576 MW—a ~60% reduction; and if tasks can be routed in real time across multiple factories within the group to sites with available headroom, new interconnection demand drops to just **30 MW**, a 98% reduction. Most individual factory sites need less than 1 MW—by contrast, the median peak demand among the 69 AI data centers and campuses in the Epoch AI database ranges from 21 to 1,135 MW, with a median of 182 MW: **the typical interconnection demand of a newly built AI data center exceeds that of a factory-level site by more than two orders of magnitude**. This hinges on open-weight models (models whose weights are publicly released and deployable locally) achieving sufficient capability and hardware efficiency to support localized deployment; and crucially, it is not merely the label “inside a factory” that matters—modeling shows that if the host factory itself carries no production load to co-locate with, the single-point hosting option actually requires slightly *more* interconnection capacity than building a new data center; what truly matters is co-location *and* intelligent scheduling.
## Making the Load Itself “Softer”
The third path neither builds power plants nor relocates infrastructure: instead, it makes the AI load itself “softer,” enabling the existing grid to absorb more. A November arXiv study turned this idea into a quantifiable planning tool: simulating the full year (8,760 hourly snapshots) on a modified 118-node standard test power grid, allowing interactive requests to be routed across sites and batch-processing tasks to be delayed in time, then asking how many additional data centers the grid could accommodate—termed **hosting capacity**. The answer: temporal flexibility alone boosts hosting capacity by 21.26%, spatial flexibility by 5.37%, and their combination lifts capacity from 4,635 MW to 6,232 MW—a 34.44% increase—**without any grid reinforcement or capacity expansion**; furthermore, the synergistic gain exceeds the sum of the two individual gains by approximately 362 MW—temporal and spatial flexibility are complementary.
But flexibility is not free, nor is it universal. Also posted to arXiv in November, another study splits “flexibility” into two halves: how much users are *economically willing* to shift (modeled via penalty parameters), and how much they are *physically able* to shift (constrained by time windows and maximum transfer volumes). In simulations of the Texas grid, operational flexibility imposes a **physical lower bound on unmet demand—one that no economic incentive can overcome**; longer time windows and larger transfer limits reduce cost, but affect grid utilization differently; time-based rescheduling flattens the net system-wide load peak and, in simulated scenarios involving coal-plant retirements, partially alleviates local transmission bottlenecks.
## To the Grid, AI Is a Strange New Kind of Load
Finally, there is an often-overlooked reality: AI data centers represent a **structurally novel** type of load for the power grid. A systems study posted to arXiv in October notes that they are highly spatially concentrated, transmission-level “point loads,” deployed extremely rapidly, and dominated by power-electronic devices (inverters/converters)—unlike traditional large motors, they possess no inertia. Empirical power trajectory analysis of a publicly available generative AI workload dataset reveals: GPU power spectra are not smooth broadband signals, but exhibit “tonal” features—**comb harmonics** (a series of equally spaced spectral spikes, like teeth on a comb) appear at training step frequencies; inference exhibits sub-second periodicity; and dominant fluctuation frequency bands shift with workload type (training and offline inference occur on second-scale, online inference on timescales of tens of seconds or longer). The researchers cautiously state this is merely a characterization of measured loads—not a universal signature—but argue such loads **deserve spectral description, not just peak-power characterization**, because spectral properties directly impact grid reserve configuration, ramping products (regulation services prepared for rapid load changes), voltage stability, and harmonic interactions. This observation resonates with findings in earlier technical-economic analyses: training loads appear flat on the hour scale, yet individual nodes can fluctuate by thousands of watts on millisecond timescales—and fluctuations across nodes are temporally correlated.
## Queuing Changes Everything
Putting these five studies together yields several conclusions. First, the AI infrastructure bottleneck has shifted: from “Can we buy chips?” to “Can we get connected to the grid—and how fast can we build?” Once speed is monetized, even long-standing decision criteria based solely on unit energy cost flip. Second, the three pathways—building on-site generation, borrowing factory capacity headroom, and softening the load—are unified by a shared logic: **don’t wait for the grid**; they are not mutually exclusive options, but rather three distinct solutions to the same problem. Third, flexibility has physical limits: incentives cannot purchase beyond physical envelopes, yet temporal and spatial flexibility exhibit super-additive synergy—so flexibility should be treated as a quantifiable **planning resource**, not merely an ad-hoc dispatch tool. Fourth, for grid operators, AI loads must be characterized spectrally—not just by peak power—prompting re-evaluation of the entire toolkit for reserves, voltage control, and harmonics. Nearly all these studies remain preprints; several rely on standard test grids and forecasted demand, so definitive industry consensus remains distant—but the direction is already unmistakably clear. Questions left for readers: Should the interconnection queue create fast-track lanes for flexible, “softer” loads? Could the idle electrical headroom under factory rooftops become a new asset class for manufacturing?