Small Modular Reactors (SMR) for AI DCs, in depth: the designs, connection models, sizing around refuelling, training load swings and realistic timelines

What an SMR is, and which designs matter

The IAEA defines small reactors as up to about 300 MWe per unit; "modular" means the plant is built from standardized, largely factory-fabricated modules. For comparison, a large light-water unit such as an AP1000 is about 1,100 MWe. The attraction is not that small reactors are cheaper per megawatt by physics; large plants have economies of scale. It is that standard modules built in series might shorten construction, reduce financing risk and let capacity grow in increments that match a data center's phases.

The designs differ in coolant, fuel and how they behave operationally, and those differences matter to a data center buyer more than the label does.

Fuel is a hidden schedule risk. The light-water designs use conventional low-enriched uranium. Xe-100, Natrium and Kairos rely on high-assay low-enriched uranium (HALEU), whose commercial supply chain is still being built. Treat any first-power date for a HALEU design as conditional on fuel.

How a reactor campus connects

There are two basic connection models. In a grid-connected arrangement, the reactor sells power into the grid and the data center buys it under a power purchase agreement. The grid absorbs mismatches and outages, but the data center still needs its own interconnection capacity, which is the scarce resource in many regions. In a behind-the-meter or co-located arrangement, the campus takes power directly from the plant switchyard. That can avoid transmission queues and charges, but the campus now depends on the plant's availability, and regulators and grid operators have scrutinized how co-located load shares grid costs. Most realistic designs are hybrids: co-located supply with a grid tie sized for backup.

Siting is shaped by the reactor, not the data center. A nuclear site needs a security perimeter, an emergency planning zone whose size the regulator sets for each design, cooling water or dry cooling, and transport access for large components. Several SMR designs aim for a small emergency planning zone, which is what makes co-location next to a campus plausible. Check it for each design and site rather than assuming it.

Sizing around refuelling and outages

A data center's demand is not its IT load. Facility power equals IT power times PUE, and the supply has to cover facility power through refuelling, maintenance and forced outages. Light-water modules refuel on a cycle of roughly 18 to 24 months, typically taking each module offline for weeks; NuScale's multi-module plants refuel one module at a time. Pebble-bed designs such as the Xe-100 refuel online, which removes the planned outage but not the forced ones. A useful planning rule is to provide N+1 modules, so that with one module out the rest still cover the campus, or to size a grid tie that covers the largest module.

import math

def size_smr_campus(it_mw, pue, module_mwe, house_frac=0.05,
                    spare_modules=1, grid_backup_mw=0.0):
    """Return module count and margins for a co-located campus.
    house_frac: share of gross output the plant itself consumes."""
    facility_mw = it_mw * pue
    net_per_module = module_mwe * (1 - house_frac)
    n_needed = math.ceil(facility_mw / net_per_module)
    # Cover the loss of the largest unit from spares or from the grid.
    spare_from_grid = math.floor(grid_backup_mw / net_per_module)
    n_total = n_needed + max(0, spare_modules - spare_from_grid)
    return {
        "facility_mw": round(facility_mw, 1),
        "modules_for_load": n_needed,
        "modules_total": n_total,
        "net_capacity_mw": round(n_total * net_per_module, 1),
        "margin_one_out_mw": round((n_total - 1) * net_per_module - facility_mw, 1),
    }

print(size_smr_campus(it_mw=300, pue=1.2, module_mwe=77))
print(size_smr_campus(it_mw=300, pue=1.2, module_mwe=77, grid_backup_mw=80))

The house-load fraction above is a placeholder; real values depend on the design and cooling, so take them from the vendor. The structure is what matters: load, PUE, net output per module, and an explicit answer to the question "what happens when one module is down?"

Reactors versus training load swings

GPU training is an unusual electrical load. Thousands of GPUs run in lockstep: compute phases draw near full power, and communication phases, checkpoints and synchronization stalls drop power sharply, all at the same moment across the cluster. At campus scale that becomes swings of tens of percent of IT load over seconds or less, repeating with the step time, and a full drop when a job crashes. Grid operators and GPU vendors have both flagged this pattern; recent rack power systems add energy storage and power smoothing for this reason.

Reactors prefer constant output. Light-water plants can load-follow, but ramp limits are typically a few percent of rated power per minute, and frequent cycling has fuel and thermal-stress costs. A turbine on a small, islanded system also has little inertia, so a sudden 30 percent load step is a frequency event. The practical design is layered: the reactors run near flat baseload; a battery system or UPS in front of the GPU halls absorbs the seconds-scale swings; the grid tie, if present, absorbs residual mismatch; and software power smoothing in the training stack (holding a minimum power draw during communication, or staggering job starts) shrinks the swing at the source. Natrium's molten-salt storage is one design that decouples reactor heat from electrical output, though it acts on minutes to hours, not on the sub-second swings a training step produces.

Islanded operation, a campus with no grid tie at all, is the hardest case. It needs storage sized for the largest credible load step, protection coordination designed for low fault current, and a plan for a reactor trip, which removes a whole module in an instant. Very few teams should start there.

Worked example: a 300 MW campus

A team plans a 300 MW IT campus at PUE 1.2, so facility demand is 360 MW. Using 77 MWe modules and assuming 5 percent house load, each module nets about 73 MW. Five modules are needed for load (365 MW), and N+1 means six, which is a full six-module US460 plant netting about 439 MW. With one module out, 366 MW remains, leaving a thin 6 MW margin. If the site also has an 80 MW grid tie, the calculator drops the spare module, but then every refuelling outage leans on the grid, and the grid tie must be firm, not interruptible.

With the BWRX-300, two units net about 570 MW, plenty for 360 MW, but losing one leaves about 285 MW, so a single-reactor outage cuts the campus by a fifth. A third unit fixes that at the cost of hundreds of megawatts of surplus that must be sold to the grid. Large units make spare capacity expensive; small modules make it cheap. That granularity, more than cost per megawatt, is the operational argument for small modules.

On the dynamic side, assume training swings of 25 percent of IT load, 75 MW, within a second. A battery system that can deliver 75 MW for a minute is about 1.25 MWh of energy, small in energy terms but demanding in power and cycling, which steers the choice toward chemistries and inverters rated for high cycle counts.

Timelines and the bridge

Be precise about dates. The first SMR units in North America are targeted around the end of this decade: OPG's first BWRX-300 at Darlington, and several US projects aiming for about 2030 or later. Natrium only received its construction permit in 2026. History argues for caution. NuScale's first planned plant, the UAMPS Carbon Free Power Project, was cancelled in November 2023 after its projected power price rose and subscribers left. First-of-a-kind nuclear projects worldwide have run late and over budget.

So an AI build planned for 2027 or 2028 cannot run on SMRs. Bridging supply comes from the grid, on-site gas turbines, contracts with existing nuclear plants, or restarts of retired large reactors. None of those are SMRs, and they should not be counted as such in a plan. An SMR agreement is best treated as a 2030s option that the site design should leave room for: land, interconnection, a substation that can accept a co-located feed, and a power architecture with storage already in front of the halls. See grid connection for AI sites for the interconnection side.

Failure modes

• Licensing slip. A one-year regulatory delay moves first power and leaves a campus stranded. Keep a grid or gas path that can carry the site alone.
• Fuel supply. HALEU designs depend on enrichment capacity that is still being built.
• Cost escalation. Power purchase prices can rise before construction, as the UAMPS project showed. Contract for price reopeners and exit terms.
• Common-mode outage. Identical modules share design flaws; a regulator-ordered inspection could take several offline at once. N+1 does not cover that; a grid tie does.
• Load-step trips. A crashed training job dropping 100 MW at once can trip protection on a weak or islanded system. Model it and size storage for it.
• Planning on the wrong label. Counting a restart or a large-plant agreement as SMR capacity hides the real schedule risk.

Trade-offs

For the wider power picture see GPU data center power requirements, power density in AI data centers and power backup for AI data centers.

What to do next

1. Write the campus load model: IT MW by phase, PUE, and the worst-case load step from training.
2. Run the sizing calculator for two or three candidate designs and record the margin with one module out.
3. Decide grid-connected, co-located or hybrid, and size the grid tie for the largest module.
4. Specify storage in front of the GPU halls for seconds-scale swings, and ask the training team about power smoothing.
5. Reserve land, substation capacity and an interconnection path now, even if the SMR is a 2030s option.
6. Track each candidate design's licensing and fuel milestones against your dates every quarter.