Small modular reactor cogeneration for a hyperscale data center

First author

200 MWe data center, ERCOT
BWRX-300 reactor
8,760 hours
109 scenario runs
Hourly MILP co-optimization

Research question

Under which reactor-capital, market and policy conditions can a small modular reactor colocated with a hyperscale data center, and using its steam to drive absorption cooling, supply the campus more cheaply than the grid?

Key insight

An output-based credit makes cooling with reactor steam more expensive

The Section 45Y credit is paid per megawatt-hour of electricity. Steam sent to the absorption chiller is steam the low-pressure turbine does not convert, so each unit of cooling heat gives up both power and the credit on it. The incentive that narrows the reactor's capital gap also raises the cost of recovering its heat.

Why does the credit discourage steam cooling?

Reactor steam can either drive the turbine or feed an absorption chiller that makes chilled water. The production tax credit pays only for electricity. Every unit of steam sent to the chiller lowers turbine output and forfeits the credit on that output. So a larger credit raises the price at which steam cooling pays, leaving more hours to the electric chiller.

Illustrative model with stylized numbers: it shows the mechanism described in the paper, not the study's data or results.

My contribution

First author. I did most of the work: I co-conceived the study, co-developed the methodology, wrote the model software, curated the data, ran the analysis, made the figures and wrote the original draft.

How I approached it

ChallengeThe value of steam cooling depends on hourly physics: wet-bulb derating, crystallization risk, part-load chiller efficiency and the turbine work lost to extraction.
What I didI modeled absorption performance against hourly wet-bulb temperature with a crystallization cutoff, the electric chiller's part-load curve with SOS2 constraints, and extraction with a Willans-line penalty from a steam-table heat balance.
Why it matteredPricing diverted steam matters: treating it as a free by-product can by itself decide the sign of a cogeneration result.
ChallengeOne-at-a-time sensitivities cannot show whether a project falls inside the region where colocation beats the grid.
What I didI designed 109 scenario runs, a baseline and nine sensitivity groups, including a five-by-five grid of reactor and absorption-chiller capital costs and a grid of reactor capital by market year.
Why it matteredThe joint grid placed grid parity near $5,000/kWe under 2023 prices and showed that chiller capital moves the margin by at most 13 points.
ChallengeOne BWRX-300 produces about 2.9 times the campus's average IT load, so the plant is partly a merchant generator.
What I didI swept campus size from 0.5 to 3.0 times the IT load, with the reactor fixed and the grid connection enlarged only for campuses above the baseline size.
Why it matteredSize matching explained part of the gap, not most: the cogeneration margin improved from −62% to −29% at three times the load.
ChallengeMargins measured against grid cost are distorted across market years, because the grid cost itself falls in low-price years.
What I didI reported cross-year results also in absolute annualized cost, next to the grid-cost margin.
Why it matteredIn absolute terms the market year moved the cogeneration cost by about $48 million a year, a fifth of the reactor-capital range.

Design choices

Extract steam at the high- to low-pressure turbine crossover: Alternatives considered: a parallel tap on main steamCrossover steam has already expanded through the high-pressure turbine, so it costs about half as much turbine work; a sweep of the penalty tests this assumption.
Keep a full-size electric chiller and let both chillers share dispatch: Alternatives considered: absorption as the primary chiller, with electric backup for emergencies onlyThe absorber has no steam during refueling and runs mainly in high-price hours, so the optimizer chooses between the chillers hour by hour.
Levelize the ten-year 45Y credit over the 20-year comparison window: Alternatives considered: levelizing it over the reactor's 40-year lifeThis matches the credit to the annualized-cost comparison; the 40-year variant moves the thresholds but not the mechanism.

Method

Mixed-integer linear program that co-optimizes, for every hour of a year, reactor output, steam extraction at the turbine crossover, absorption and electric cooling, and grid exchange for a 200 MWe data center in Houston (ERCOT) colocated with a BWRX-300 reactor; grid supply and an on-site gas combined cycle are evaluated on the same hourly basis. 109 scenario runs vary reactor and chiller capital cost, market year (2022–2024), carbon price, cost of capital, cooling efficiency, battery storage and campus size, with the Section 45Y production tax credit as the baseline policy.

Key findings

At 2023 ERCOT prices and mid-range reactor capital cost (ATB 2024), the nuclear configurations cost 49–62% more than grid supply, even with the Section 45Y credit.
Grid parity at zero carbon price requires reactor capital near $5,000/kWe under 2023 prices. At nth-of-a-kind capital the nuclear options are 77–89% cheaper than grid supply in 2023, but between parity and 34% more expensive in the low-price 2024 market.
A carbon price of $53–64 per tonne of CO₂ closes the mid-range gap when exports are credited at the hourly average grid emission intensity.
Absorption cooling follows hourly electricity prices and supplies 38% of annual cooling, but adds $9.2 million per year relative to the reactor-only plant; it breaks even only at an installed chiller cost near $60/kWc ($570/kWc on a legacy-efficiency campus), against surveyed prices of $450–1,200/kWc.
Capital cost and policy set the boundary; operating choices move it far less. Across the tested ranges, reactor capital spans 321 percentage points of the grid-cost margin and the cost of capital 119, against 12, 13 and 1 for cooling efficiency, chiller capital and battery storage.

Boundary conditions

One colocation-style load profile at one Houston site; hourly average rather than marginal emission factors, with the carbon-price crossovers conditional on crediting exports; demand charges and ancillary-service revenues excluded; reliability not modeled on a common basis across options; the reactor is one fixed commercial unit. The numerical thresholds depend on these market, financing and crediting conventions; the mechanism does not.

What it led to

The same question also matters at the national scale, which is the subject of my next study. Data-center demand and the commercialization of small modular reactors

Publications

Energy Convers. Manag.

Techno-Economic Boundary Analysis of Small Modular Reactor Cogeneration for Hyperscale Data Center IT and Cooling Loads

H. Li, B. She, J. Zhang

Energy Conversion and Management, 2026. Accepted

arXiv Code

@article{li2026smrcogeneration,
  author        = {Li, Honglin and She, Buxin and Zhang, Jie},
  title         = {{Techno-Economic Boundary Analysis of Small Modular Reactor Cogeneration for Hyperscale Data Center IT and Cooling Loads}},
  journal       = {Energy Conversion and Management},
  year          = {2026},
  eprint        = {2608.10999},
  archivePrefix = {arXiv},
  primaryClass  = {eess.SY},
  note          = {Accepted}
}