Carbon capture could absorb more than 90% of US data-centre emissions, study finds

On 13 July 2026, a Cornell University researcher and his co-authors put a number on a long-running question in the AI-energy debate: more than 90% of the carbon-dioxide emissions produced by the United States' data-centre fleet could, in principle, be captured and permanently stored if existing post-combustion capture technology were bolted onto the cooling and flue systems of those facilities. The figure is not a forecast. It is the upper bound of a thermodynamic calculation published the same day, and it sits well above what a commercial plant could realistically deliver today. The gap between those two numbers is now the story.

The paper arrives at a politically awkward moment. US electricity demand from hyperscale computing is on track to grow faster than almost any other end-use category this decade. The largest operators have signed multi-gigawatt power purchase agreements, restarted idled nuclear units, and queued behind utilities for new transmission. None of that solves the carbon problem on its own. The new study argues that point-source capture, applied at the smokestack or the diesel-behind-the-grid, deserves a second look, not as a substitute for clean power but as a complement to it.

What the model actually counts

The lead author, Hon Chung Lau, an adjunct professor in Cornell's Department of Chemical and Biomolecular Engineering, and his collaborators ran a series of process simulations on amine-based and carbonate-based solvent systems attached to representative data-centre exhaust streams. The headline figure, more than 90%, describes the share of CO₂ that the capture unit can pull out of the flue gas at design flow, before accounting for the parasitic load the capture plant itself imposes on the facility.

That caveat matters. Capturing CO₂ takes steam, pumps, and compressors. The energy those auxiliary systems draw either comes from the grid, in which case it lifts the very emissions the unit is trying to cut, or it is siphoned from the data-centre's own output, in which case the effective compute delivered to customers falls. The study quantifies both effects and reports that the net reduction, after the energy penalty, remains substantial, though it does not give a single combined percentage in the abstract.

The economics still bind

Cost is the obvious next question, and the one the abstract does not answer. Commercial post-combustion capture on US coal and gas plants has historically run between $50 and $120 per tonne of CO₂ avoided, depending on concentration, scale, and how the captured gas is transported. A data-centre flue stream is not identical to a power-plant stack; the CO₂ concentration is lower because the exhaust is heavily diluted with air-cooling flow. Lower concentration means larger equipment and more energy per tonne captured. The Cornell group flags this in its methodology section and treats the data-centre case as more demanding than a conventional utility retro.

There is also the question of what happens to the CO₂ once it is separated. Geological sequestration requires pipeline or truck transport to a Class VI well or an enhanced-oil-recovery site. The US permitting pipeline for Class VI storage has lengthened, not shortened, in the last two years. Without a sequestration pathway, capture is just expensive compression into a tank.

Why a frontier-model build-out makes this harder

The arithmetic looks different at frontier scale. A single 1-gigawatt hyperscale campus can draw as much continuous power as a mid-sized US city. Train a next-generation foundation model on that campus for a training run measured in weeks, and the cumulative energy input is in the terawatt-hour range. Apply capture to that exhaust, and the absolute volume of CO₂ handled rises into the millions of tonnes per year per site, an order of magnitude above anything the current carbon-capture industry has ever processed at a single location.

That scale issue is also an opportunity. A multi-million-tonne-per-year anchor offtaker would, in principle, underwrite the kind of trunk-line CO₂ transport infrastructure that today exists only in pockets, mostly tied to oil fields in Texas and North Dakota. The US Inflation Reduction Act's 45Q tax credit, which pays operators per tonne of CO₂ sequestered, was designed for exactly this kind of anchor load. Whether the Treasury guidance and the prevailing wholesale power price together make the unit economics work at hyperscale is a question the study does not attempt to settle.

What the critics say

Environmental groups that have long opposed carbon capture as a fossil-industry fig leaf are already sharpening their response to this paper. The standard critique runs as follows: the cheapest tonne of CO₂ to avoid is the one never produced, and every dollar spent on retrofits is a dollar not spent on new wind, solar, or transmission. Defenders of capture counter that renewables-plus-storage have not yet reached the firm-24/7 level that hyperscalers require for training runs, and that a layered strategy, clean power first, capture second, may be the only path that keeps the AI build-out politically survivable.

There is also a methodological objection specific to this study. Critics note that data-centre emissions are dominated by the indirect, Scope 2 load on the grid, not the direct, Scope 1 exhaust at the site. Capturing the site's own flue addresses only a fraction of total lifecycle emissions, even if that fraction is large in absolute terms. The Cornell team acknowledges this framing in its discussion section and treats capture as one tool among several.

What stays uncertain

The sources do not specify a target cost per tonne at which capture at hyperscale becomes commercially viable, nor do they name any operating hyperscale facility currently piloting the technology at gigawatt scale. The paper is a process-simulation result, not a field demonstration. Whether the engineering firms that supply amine and carbonate systems to power plants can scale their units to data-centre specifications, and whether pipeline and well capacity exist to receive what they capture, are open questions. Monexus will track both as the field develops.

This piece treats capture as one instrument in a portfolio, not a verdict on whether the AI build-out should proceed. Cornell's result narrows the engineering case; it does not close the political one.