Nuclear power plants at commercial scale are baseload machines. They are designed to operate at or near full rated capacity continuously, and any departure from that operating mode introduces complexity. Load-following on a gigawatt-class light water reactor means managing xenon oscillations, moderating thermal gradients across a large core, and coordinating control rod movements at a scale that requires highly trained operators working from detailed procedures. The engineering cost of doing that frequently is real, and most utility operators simply choose not to.
A compact fission unit at the scale we are targeting, in the range of 4 to 12 MW, has physical properties that change the calculus. But the physics alone do not make the reactor dispatchable in the way a data center needs. That requires a control architecture designed from the ground up for load-following, not just for steady-state baseload operation.
What Dispatchability Means for a Data Center
A data center's power demand is not flat. It varies with workload, time of day, cooling system state, and the mix of compute tasks running at any moment. GPU clusters in inference service draw power proportional to request volume. Batch processing jobs start and stop. Cooling systems cycle. The aggregate facility load can swing meaningfully across a single day.
For a co-located power source to match that profile, it needs to do more than generate a fixed output and let a campus switching arrangement handle the rest. A truly integrated power source tracks the facility's load signal and adjusts output accordingly, within the response time window the facility's UPS and switchgear can tolerate. That capability is what we mean by "dispatchable" in our design context.
The Xenon Problem and Why Compact Scale Helps
The primary physics constraint on nuclear load-following is xenon-135. Xenon is produced in the fission process and also produced from decay of iodine-135, another fission product. It absorbs neutrons with unusually high cross-section, which means high xenon concentration suppresses reactor power. When power is reduced, iodine-135 continues to decay into xenon for several hours, meaning xenon concentration initially rises after a power reduction and then falls as the xenon decays. This creates a transient that makes returning to full power difficult for a window of hours following a load reduction.
In a large power reactor with a long neutron diffusion length and a massive fissile inventory, xenon transients produce spatial oscillations across the core that require careful monitoring and control rod choreography to manage. The core is large enough that different regions can be at different xenon states simultaneously, creating non-uniform flux distributions.
A compact core at the scale we are targeting is physically small. The neutron diffusion length is comparable to the core dimension, which means the core behaves more uniformly. Spatial xenon oscillations that would be a major concern in a large reactor are substantially reduced. The xenon problem does not disappear, but it becomes manageable with the right control approach rather than an operational constraint that limits load-following entirely.
Where the AI Layer Enters
The control architecture we are building is not a rules-based system with lookup tables. It is a model-predictive control framework that combines a physics model of the reactor's neutronics and thermal hydraulics with a machine learning layer trained on simulated operational data. The physics model provides the constraint space: what states are reachable, what transitions are safe, what boundaries cannot be crossed. The ML layer provides the optimization: given the facility's projected load curve and the current reactor state, what control actions produce the best output match while staying within physics constraints.
The approach we are taking treats the xenon inventory and core temperature distribution as state variables that the controller actively tracks, not passive consequences that operators monitor. The system builds a horizon of the expected xenon trajectory over the coming hours and plans control rod position and coolant flow adjustments to pre-condition the reactor for anticipated load changes, rather than reacting to them after the fact.
We are not claiming this works perfectly in all scenarios. This is simulation and design work. The validation pathway for a nuclear control system goes through NRC review and site-specific testing. What we can say is that the architecture is sound, the physics basis is real, and the compact core geometry substantially improves the tractability of the problem compared to what utility operators face on large reactors.
Response Time Targets
Our design targets a load-following response on the order of several minutes for step changes in demand, not seconds. This means the co-located reactor is not responding to individual server-level power fluctuations. Those are handled by the facility's UPS and power distribution system, which are designed for millisecond-level response. The reactor responds to the sustained shift in facility-level load demand, adjusting over a few minutes to match the new setpoint.
This is consistent with how gas turbines behave in combined-cycle arrangements and is entirely within the capability envelope of a UPS-backed facility. The data center's power infrastructure already handles the fast transients. The reactor handles the slowly varying baseload.
What This Is Not
We want to be precise about the boundary of this claim. We are not saying that the reactor will respond to every instantaneous power fluctuation in the facility. We are not saying this control system will be approved by NRC without modification. The licensing process for a novel reactor control architecture is iterative, and the system we deploy will reflect review and feedback from that process. We are saying that the design architecture for load-following operation in a compact fission unit is a tractable engineering problem, and that building a data center co-located power source without addressing it would be building the wrong product.
A nuclear unit that locks in at a fixed output and leaves the data center to manage the mismatch is only marginally better than grid power from a campus integration standpoint. The work we are doing on dispatchability is what makes the value proposition real.
The Integration Picture
The control system's load signal input comes from the data center's DCIM infrastructure through a defined API. The facility shares its current draw, its projected load curve for the next several hours based on scheduled workloads, and its available UPS headroom. The reactor control system uses that data as its load forecast input. This tight integration between the power source and the facility's operational systems is something conventional grid-connected facilities do not have. The campus effectively becomes a single coordinated power system rather than a facility that buys power from an external entity and handles all variations internally.
Getting that integration right is a significant engineering project. The control API, the safety interlocks, the handoff protocols between reactor control and facility UPS, all of it needs to be designed and tested carefully. We are early in that work, and we are doing it because it is what makes a co-located fission unit genuinely useful to the operators we are building for.
