The power requirements of a GPU compute cluster are not the same as those of a corporate office building. An office can absorb brownouts, momentary undervoltage events, and scheduled load shedding. A cluster running active inference jobs cannot. When an AI model responds to a query, every server in the hot path needs to be drawing power continuously, at the right voltage, at the right frequency, for the duration of that response and every response that follows it. This is not a soft requirement. It is baked into the hardware design.
This creates a specific, hard constraint for data center operators that is different from how most industrial and commercial energy procurement works. Understanding why requires looking at what power intermittency actually means at the hardware level, not just at the grid planning level.
The Capacity Factor Problem
Wind power across the United States operates at a fleet average capacity factor of roughly 25 to 42 percent, depending on region. That number represents the fraction of total possible output the fleet actually produces, averaged over a year. Offshore wind trends toward the higher end. Onshore wind in the central plains varies widely by season and weather pattern.
The issue is not the annual average. It is the distribution. Capacity factor is consistent only in aggregate. For a specific generator on a specific day, output can be close to zero for extended periods. A three-day high-pressure system in July in Texas will reduce wind output across thousands of square miles simultaneously. A data center in that region cannot purchase wind power from a continent away to smooth that event: power travels over transmission lines with finite capacity and real losses.
Solar has the same structural limitation from a different angle. Production is zero from dusk to dawn, reduced by cloud cover, and varies seasonally in ways completely independent of the facility's load needs. A data center running AI inference at 3 AM does not benefit from solar energy produced at noon the previous day unless that energy has been stored, and storing it at data-center scale introduces its own complications.
Why Storage Does Not Bridge the Gap
Battery storage is the standard answer to intermittency concerns, and for smoothing short-duration fluctuations, 2 to 4 hours, that answer is technically correct. A battery buffer of that size can handle most sub-daily variability without data center uptime impact.
The problem appears at longer durations. A facility drawing 20 MW of compute load that faces a 48-hour wind and solar lull needs approximately 960 MWh of storage to maintain operation without grid contribution. At current battery installed costs, that storage capacity alone is a substantial nine-figure capital commitment. The financing structure, land requirements, thermal management systems, and replacement cycle of a battery bank of that size represent a major infrastructure commitment in themselves, and that commitment scales directly with the length of the weather event you are protecting against.
No battery chemistry currently in commercial deployment solves multi-day seasonal intermittency at data center scale economically. Long-duration storage technologies are in development, but commercial availability at the scale and cost needed is not a near-term reality. Anyone who has spent time modeling the actual numbers, which we have, reaches the same conclusion.
What Hardware Actually Tolerates
Modern compute infrastructure has specific power quality tolerances that constrain what an intermittent-plus-storage arrangement can reliably deliver. High-density GPU servers require supply voltage within approximately plus or minus 2 percent of nominal. Frequency deviations beyond a narrow band trigger protective responses in power supply firmware. Sustained voltage sag events, even within nominal IEEE 1100 power quality standards, can cause server power supplies to initiate protective shutdown rather than continue operating in a degraded state.
Uninterruptible power supply systems handle brief transitions effectively. A well-designed UPS carries a facility through a grid transient lasting seconds. What UPS systems are not built for is continuous operation over hours. Their battery capacity is sized to enable graceful shutdown or bridge to a generator start, not to power production workloads through extended generation gaps.
The Transmission Layer
Even for facilities with access to nominally reliable grid-connected generation, the transmission and distribution infrastructure between the generator and the facility introduces reliability dependencies that are separate from generation availability. A substation failure or transmission line fault removes power regardless of whether the generating plant is running normally.
Interconnection access is also not guaranteed. New large industrial loads connecting to transmission networks in most US markets today face study queues measured in years and cost estimates that remain uncertain until the study is complete. Some markets have backlogs of four to five years for new large-load interconnection requests. Operators planning data centers for operation in 2028 or 2029 who have not started interconnection processes are already facing real schedule risk.
What This Means for Applied Atomics
We are not arguing that the grid is poorly run or that renewable procurement has no role in data center energy strategy. Grid-connected renewable PPAs make sense for many applications, and the carbon accounting benefits are real. The specific constraint is narrower: a facility running continuous, latency-sensitive AI inference workloads cannot stake its uptime on variable generation as its primary power source.
A co-located compact fission unit, sized to match the facility's baseload demand and connected directly to campus switchgear, addresses that constraint in a way that intermittent generation cannot. The physics of fission do not depend on external conditions. A reactor producing 8 MW at 11 PM on a cloudy January night produces the same power as it does at noon on a clear July afternoon. That property, continuous dispatchable output without a weather dependency, is the fundamental thing data center operators are buying when they engage with us.
The technology to produce it at compact scale, without requiring a large utility-scale plant infrastructure, and with an AI control layer that tunes reactor output to the facility's actual load curve, is what we are building. Whether we execute on that within the regulatory and technical constraints of nuclear development is the question we are working to answer. We think the answer is yes. The power problem we are solving, however, is not in question.
The Planning Horizon
Compact nuclear infrastructure cannot be procured six months before a facility needs it. Site selection, regulatory pre-application engagement, detailed design, licensing, and construction represent a timeline measured in years, not quarters. Data center operators planning new capacity for the early 2030s need to be evaluating this option and beginning preliminary conversations now. The alternative, continued dependence on grid power with increasingly strained interconnection capacity and growing intermittent generation penetration, works until it does not. Building the infrastructure to guarantee power for the next decade of AI compute means making those decisions well before the demand arrives.
