Carbon capture plus bigger fans: the math that could take 90% off AI data-center emissions

A peer-reviewed study published this week argues that the emissions cost of running the AI boom does not have to follow the boom's curve. The paper, co-authored by Hon Chung Lau, an adjunct professor in the University of Notre Dame's Department of Chemical and Biomolecular Engineering, models a retrofittable package, industrial carbon capture bolted onto existing cooling loops, paired with heat-pump upgrades, and finds it could remove more than 90% of operating emissions from a representative hyperscale facility. The work lands as data-center electricity demand continues to outrun the grid's ability to clean itself, and as operators increasingly frame the problem as one of on-site engineering rather than new generation.

The result is not a free pass. It is, however, a more sober answer than either the apocalyptic or the dismissive read of AI's energy trajectory. If the technology the paper models survives contact with permitting, capex and, frankly, with actual carbon-capture plants operating at scale outside the oil and gas industry, then the question shifts from whether AI can be decarbonised to who pays to decarbonise it.

What the study actually models

The paper does not propose a greenfield rebuild of an AI campus. Instead, Lau and co-authors simulate a retrofit on what the report calls a representative hyperscale data centre, layering three interventions on top of an existing facility: amine-based or comparable post-combustion carbon capture on the diesel- or gas-fired back-up generation, optimisation of the cooling plant, and heat-pump substitution for the electric resistive heating that currently warms humidification and service water. Combined, the modelled stack removes more than 90% of the operating emissions the baseline facility would otherwise emit, the authors report.

The framing matters. Most published decarbonisation roadmaps for data centres foreground two levers, buying renewable power via power-purchase agreements and contracting for grid-supplied clean energy. The Notre Dame paper reaches for a third, more industrial-tradition lever: treat the diesel generator, the chiller and the boiler as emissions sources that can be engineered the way a refinery or a cement kiln is engineered, with end-of-pipe capture and electrified heat.

Where the technique sits in the wider toolkit

Carbon capture for power generation has a mixed commercial record. Large-scale post-combustion units on coal and gas have run, often with state support, and have often struggled with cost per tonne captured. The data-center paper inherits that track record, both its promise and its bruises. What it adds is a context where the captured stream is concentrated: a single back-up generation unit feeding a single stack, rather than a 500-megawatt power block.

Heat-pump retrofits are the quieter story. Replacing electric-resistance boilers with industrial heat pumps is a known, deployable move across district heating, food processing and chemicals. Applied to data centres, the same hardware class can supply humidification and space heating without the resistive load. Outside engineers contacted by trade press routinely estimate payback periods of three to seven years on such swaps in commercial settings; the paper does not commit to a single number for data centres but flags capital cost as the binding constraint.

The retrofit framing puts the proposal in dialogue with two industry moves already underway. Hyperscalers have made unprecedented renewable power-purchase agreements in the past three years, contracting gigawatts of new wind and solar that are physically remote from the data centres they nominally power. Operators have also signed behind-the-meter gas and nuclear deals aimed at matching load to clean firm supply. Carbon capture plus heat-pump retrofits do not replace any of that. They sit on top of it, addressing emissions the clean-power contracts cannot reach: the diesel back-up, the on-site combustion-based heating loads, and the residual grid mix that survives even an aggressive renewable build-out.

The counter-read: why retrofits may not scale fast enough

Sceptics do not dispute the physics. They dispute the timeline. Carbon-capture projects in power and industry have historically taken longer to permit, finance and commission than their proponents promised; retrofit projects on operating data centres face the additional friction of working around live workloads, customer SLAs and tight outage windows. Heat-pump installations are more routine, but a single campus can require dozens of units, each a long-lead procurement.

The cost question cuts harder. Per-tonne capture costs in the literature span a wide range and depend heavily on plant utilisation; back-up generators in normal operation run rarely, which means the marginal cost of capturing a tonne from a stack that is mostly idle can rise sharply. Lau's paper, according to a Notre Dame release accompanying the study, frames the economics as favourable in aggregate but acknowledges the capital intensity of the capture unit itself.

There is also the question of what the captured CO2 becomes. Enhanced oil recovery, the historic sink for most captured carbon, is a hard sell for an industry whose customers are tightening ESG mandates. Geological sequestration, saline aquifers or mineralisation pathways exist but require permitting, transport and a registry of stored volumes, capabilities still being built out across most US states.

What would have to change for the 90% figure to mean something

A reading of the paper as a near-term policy roadmap is premature. Read as an engineering option set, it is more useful. Three things would have to move for the modelled result to translate into real-world abatement at the pace AI capacity is growing.

First, capital. The capture unit is the dominant capex item in most retrofit scenarios; structured finance or utility-style cost recovery would be needed to spread that across the operating life of the data centre. Second, an honest accounting framework. Today's voluntary reporting frameworks allow renewable PPAs and carbon offsets to paper over on-site combustion loads that the Notre Dame model makes visible. If the goal is the 90% the paper describes, the accounting has to follow. Third, supply chain. Industrial heat pumps at the scale the retrofit implies are not yet produced at the volumes the wider electrification agenda would require; amine solvents and contactor hardware for carbon capture face similar scale-up questions.

The honest uncertainty

The study is a model. It uses representative load profiles and equipment assumptions rather than measured operating data from a retrofitted campus. Outside reviewers will want to see sensitivity analyses around capture rate under partial-load operation of the back-up generators, around the auxiliary energy penalty of running the capture unit itself, and around the long-term solvent degradation that has dogged earlier commercial installations. The paper's authors and independent engineers the trade press has canvassed concur on the broad direction; they disagree on cost, on the realism of high-utilisation capture on rarely-run assets, and on how cleanly the modelled retrofits slot into existing emissions-reporting regimes.

There is one more thing the sources do not say. The 90% headline is operating emissions only; it does not address embodied emissions in the chips, the steel, the concrete or the transmission line. Those will be the next paper.

Desk note: Monexus framed the study as one option in a wider decarbonisation toolkit rather than as a fix, and foregrounded capital cost and permitting risk, issues the paper's own authors flagged, alongside the headline result.