For an AI data centre, cheap electricity can come with an expensive wait. A preprint posted on 12 November 2026 models why US operators could be better off paying more for solar-plus-battery electricity if it lets their computers start work sooner. The advantage depends on how much that earlier computing is actually worth.
Alex C. Newkirk and colleagues model a training facility with 100,000 H100-equivalent GPUs—the chips doing the AI calculations—at 99% annual availability. Solar plus batteries using direct-current wiring has the lowest cost among the modeled options after accounting for deployment time in 84.9% of their continental-US locations. Those are simulated geographic locations, not a survey of operating data centres or demonstrated savings.
The result is sensitive to the calendar: allowing gas power to arrive in three years instead of roughly four cuts solar’s winning share to 55.6%. The 99% availability target also matters: this is not a promise of uninterrupted service.
The November posting follows an earlier preprint listed in December 2025.
Why the electricity price is only half the question
A microgrid is a local power system; the off-grid versions here operate without the public electricity grid. The authors’ published model compares two solar-and-battery designs, natural gas with diesel backup, and a grid connection over a 27-year project horizon.
One solar design uses alternating current, or AC; the other uses direct current, or DC, reducing conversions between panels, batteries and computing equipment. But more efficient wiring does not automatically make solar the cheapest source of electricity. Before the value of deployment time is included, the grid generally wins the paper’s cost comparison.
The model distinguishes lifetime cost per unit of delivered electricity from a second measure that adds the value of computing forgone during construction. This “speed-adjusted” cost is a way to compare projects, not a tariff an electricity supplier charges. A low price per kilowatt-hour answers what power costs; it does not answer what waiting for it costs.
How a year of waiting becomes a billion-dollar assumption
Solar construction takes two years in the model’s defaults, while the baseline gas schedule is roughly four. Grid timing varies by region. The authors value unavailable computing at $2.40 per GPU-hour, with costs expressed in 2022 dollars. That is a modeling input, not a current rental quote or a demonstrated profit margin.
A simple illustration using those assumptions shows why time can overwhelm the electricity-price comparison:
100,000 GPUs × 8,760 hours × $2.40 per GPU-hour = $2.1024 billion for one full year.
That is the undiscounted value assigned to a full fleet’s unavailable computing—not forecast revenue, cash lost or project savings. This arithmetic omits discounting and operating costs; the model’s calculation code discounts its annual waiting-cost amounts. Turning the illustrative figure into an investment case would also require evidence that the computing could be used or sold.
Power arriving early is valuable only if buildings, chips, networks and workloads are ready too. The International Energy Agency notes that demand generally builds gradually as servers and customers come online. Treating every chip as productive from day one is therefore an assumption to challenge, not a cheque waiting to be cashed.
Faster gas—or a faster grid—can change the answer
The accelerated-gas scenario removes 29.3 percentage points from solar’s preferred-location share. That comparison rewards speed, not a technology label. The paper also retains regions where grid power wins because relatively cheap electricity and shorter connection timelines coincide.
There are ways to change the grid timetable as well. The IEA describes phased and non-firm connections: arrangements that can connect a facility sooner while limiting when or how much power it may draw. Their suitability depends on what interruptions or restrictions the workload can tolerate.
That widens the practical question. A planner should compare credible local options for starting useful work, including staged grid access—not assume the only choice is permanent isolation or years of waiting for an unrestricted connection.
The reliability and construction catches
The headline comparison requires full service in 99% of annual hours. The remaining 1% of a non-leap year is about 88 hours, or 3.65 days, although those hours need not occur together. This is a translation of the model’s availability target, not observed downtime. It cannot be treated as equivalent to continuously available customer service.
The authors’ limitations include sizing solar and batteries for typical weather, assuming mature DC equipment supply chains and treating land acquisition as frictionless. Their grid delays also use power-generation queues as a proxy for data-centre connections. Some grid upgrade costs are omitted. These simplifications can favour different options; there is no single correction that makes every result conservative.
Cheap land and quickly available land are not the same thing. If securing a solar site or its equipment delays the start, the very advantage being priced can shrink. For neighbours assessing a real proposal, what to ask about data-centre construction and local commitments is a separate, more local question than which technology wins a simulation.
Nor does this project-level ranking establish lower household bills. Who pays for network expansion is a different issue, explored in what data-centre power plans mean for electricity bills.
Compare the date, the workload and the service level
For a real decision, replace the model’s schedule with evidence about land, permits, equipment and commissioning. Replace its chip-hour value with a plausible value for the work actually ready to run, including a gradual start. Compare every power option against the same reliability requirement, then check whether the preferred option survives less favourable dates and computing values.
The useful comparison is the cost of usable computing by a credible date, not electricity alone. Solar can win that comparison even with pricier electricity—but only if its head start survives contact with the actual project.
