The AI Boom Needs Carbon Removal

Somewhere in Virginia, Texas, or Arizona, a data center is being commissioned this month that will draw more power than a small city. The server racks inside will train and run artificial intelligence models for years to come. And the electrons feeding it will, in all likelihood, come partly from natural gas — because that is what can be built fast enough to meet the demand.

AI is driving a major new wave of data center construction, and with it, a surge in demand for power and infrastructure. The International Energy Agency projects that the electricity consumption of global data centers could more than double to around 945 terawatt-hours by 2030, comparable to Japan’s entire electricity demand today.

That matters because much of the new electricity demand from data centers is still likely to be met by power sources where natural gas plays a central role. The backlog for new combined-cycle gas turbines — the more efficient type of gas plant, which generates electricity from both a gas turbine and the heat it produces — already stretches to five years. As a result, some data centers are turning instead to single-cycle gas turbines, which can be deployed more quickly but are even more carbon-intensive. In any case, that means fossil-fuel use for this generation of digital infrastructure is already largely locked in. Some of the emissions that follow can be reduced through efficiency and grid decarbonization, but a significant share will persist for years to come. I believe that closing this gap must be the job of carbon removal.

Carbon removal is the process of physically taking carbon dioxide back out of the atmosphere. At Climeworks, we have spent the past 17 years developing and deploying direct air capture technology that removes CO2 from the air and stores it in the ground for thousands of years. More recently, we launched our Climeworks Solutions business that works with third-party providers of other technology and nature-based carbon removal methods, such as reforestation, to help customers access a broader range of approaches and price points.

According to the United Nations Intergovernmental Panel on Climate Change, carbon removal will be necessary if the world is to come close to meeting its climate goals, even alongside deep emissions cuts. For companies building and using digital infrastructure, the question this raises is simple: What do they do about the emissions they cannot yet eliminate?

The strongest near-term answer is to treat carbon removal as part of the cost of digital infrastructure — not as a substitute for clean energy, but as a complement to it. Trying to pair every data center directly with a direct air capture plant may sound attractive, especially because data centers have power, land and waste heat. But in practice, that kind of integration is still highly site-specific and not yet an easy model to repeat at scale. A more realistic solution is to treat carbon removal as part of the cost of cloud and AI products, where it can be built into existing pricing and contracts. In other words, carbon removal should be built into the cost of the digital product itself, rather than physically attached to every data center site.

The incentive is simple: As companies come under growing pressure to account for the emissions linked to the digital infrastructure they rely on, data center providers that offer a credible lower-emissions product will have an advantage.

One criticism of using carbon removal in this context is that it could prolong the use of fossil fuels. That concern deserves to be taken seriously, but it also needs a nuanced answer. There is an important difference between using carbon removal to justify new fossil infrastructure, and using it to address residual emissions that cannot yet be avoided. The latter is the role that serious climate frameworks assign to carbon removal.

Data center operators are not turning to natural gas because carbon removal exists. They are doing so because natural gas can provide the speed required by the current pace of compute growth. Carbon removal should therefore not be seen as a substitute for decarbonization, but as a way to manage a real constraint in an energy system that cannot decarbonize instantly.

The relevant comparison is not carbon removal versus renewables. It is unabated fossil-powered data center expansion versus expansion in which some of the resulting emissions are credibly and durably addressed. In that sense, the growth of AI infrastructure also creates an opportunity for carbon removal: It can bring larger volumes into the market, support scale-up, and help drive down costs over time.

The economics of integrating carbon removal into AI infrastructure are more feasible than one might assume. In December, Julio Friedmann, one of the best-known experts on carbon management and carbon removal, wrote in a Substack article that a gigawatt of advanced data center capacity can generate around $10 billion to $12 billion in annual revenues. Against that scale of value creation, the cost of addressing residual emissions through carbon removal becomes more manageable.

The emissions associated with that computing power depend heavily on how it is supplied. Based on our own calculations, assuming the current U.S. grid mix and utilization rates of around 85% to 100%, a gigawatt of data center capacity would emit approximately 3 million to 4 million tons of CO2 per year. Behind-the-meter natural gas generation would produce a similar level of emissions. Renewable power can reduce those emissions significantly, while nuclear power could reduce them further.

In practice, not every gigawatt of data center compute will be powered in the same way. But assuming roughly half is supplied by renewable or nuclear power, average residual emissions would still be around 2 million tons of CO2 per year for each gigawatt of compute. That is a substantial volume — and exactly the kind of residual emissions gap that carbon removal can help address.

A portfolio of carbon removal solutions, which can directly mitigate these emissions, only costs a few hundred dollars per ton. While that is a meaningful cost, it is manageable given the economics of AI products. It is affordable enough to make a start, especially for companies that want to offer a credible lower-emissions digital product.

So, who pays? In the near term, the most likely model is that cloud and AI service providers procure carbon removal and build the cost into their products, while customers create the commercial pressure and ultimately support that cost through procurement. Even if companies are speaking more cautiously about net zero than they were a few years ago, the underlying need for credible value-chain emissions data has not disappeared. Organizations still face growing pressure to account for scope 3 emissions through disclosure rules, investor-facing reporting frameworks and supplier requirements. As their use of cloud and AI grows, they will increasingly ask providers a simple question: What emissions come with this compute, and what are you doing about them? Once buyers start routinely asking that question, carbon removal moves from being a climate nice-to-have to a product feature.

Climeworks has reduced the cost of direct air capture significantly since our first plant came online, and that trajectory will continue as the market grows. But cost curves do not come down on their own. They come down when buyers decide that a cleaner product is worth paying for. The cost of solar electricity fell around 90% between 2010 and 2023, driven not just by technology but also by early procurement commitments from the likes of Google, Microsoft, and Amazon that gave manufacturers the confidence to invest at scale.

Carbon removal is approaching a similar inflection point. In April, Climeworks signed an agreement with NTT Data — one of the world’s largest digital and IT service providers — to remove carbon dioxide from the atmosphere, as part of its commitment to net zero.

The business case, then, is simple. The AI boom is creating enormous economic value. But it is also creating residual carbon emissions that cannot be avoided only by clean power and increased efficiency. The solution is not to wait for a perfect zero-carbon grid, and it is not to force a bespoke carbon removal engineering solution onto every data center site. I believe the solution is to integrate carbon removal into the digital infrastructure offer now, and let customers choose it. That’s how lower-emissions compute becomes real and scalable. And that is why carbon removal needs to become an essential part of responsible AI growth.