by Emma Littlewood, Director of Policy and Research at 51toCarbonZero
The recent coalition announcement from Google, Anthropic and other major tech companies committing almost $1 billion to emerging technological carbon removal projects should be welcomed. There’s no doubt that durable carbon removals will have to play a significant role in limiting global heating to anywhere near 1.5C above pre-industrial levels, or even below 2C, given that global temperatures are already at 1.43C.
Carbon removal methods range from nature-based to engineered approaches. To date, nature-based solutions have represented 99.9% of all carbon removals, yet land-based projects already committed through national climate pledges exceed what can realistically be delivered. Investment in engineered carbon removals is therefore essential if they are to play the much larger role needed in future.
But these investments raise a more fundamental question. Are we increasing our reliance on carbon removals to solve a problem that should first be addressed at its source?
As AI adoption accelerates, there is a growing risk that carbon removals become a way of compensating for ever-rising emissions, rather than supporting genuine decarbonisation. That distinction matters, because the challenge facing AI is not simply one of offsetting current emissions – it is one of rapidly increasing demand.
The efficiency trap: Jevons Paradox in action
The environmental impact of AI is still relatively small compared with many other sectors, but it is growing exceptionally quickly. Data centres already consume around 6% of the electricity supply in the UK and US, and demand is expected to rise significantly over the next decade, as AI becomes embedded across almost every part of the economy.
This is a classic example of Jevons paradox. Improvements in efficiency don’t necessarily reduce overall resource consumption; they often increase it. As AI models become faster, cheaper and more capable, they are used by more organisations and relied upon for more everyday tasks. The result is that total energy demand continues to grow, even as individual models become more efficient. As Microsoft CEO Satya Nadella has put it, as AI gets more efficient and accessible, we will see its use skyrocket.
Relevant: Google Now Offering Over $6M In Research Funding For CDR And Superpollutant Solutions
That’s an important point because discussions around AI’s environmental impact often focus on efficiency gains alone. Those improvements are real and should continue. But if demand is increasing faster than efficiency improves, emissions continue to rise.
Carbon removals are not the problem here. In fact, they are likely to become increasingly important for addressing emissions that cannot realistically be eliminated. The challenge is expecting them to compensate for emissions that could, and should, be avoided in the first place.
The removal gap is bigger than it looks
The scale of AI’s growth illustrates why. Even if the industry succeeds in dramatically reducing operational emissions through cleaner electricity, significant embodied emissions from constructing data centres and manufacturing hardware will remain. Morgan Stanley estimates that data-centre emissions could reach 600 million tonnes of CO2 a year by 2030. Even if operational emissions are cut to zero through renewable electricity, the embodied share – roughly 40% of the total, covering construction and hardware – would remain largely untouched, unless there is significant progress in decarbonising cement, steel and mineral mining.
On that basis, the residual emissions left for carbon removal to address could run into the low hundreds of megatonnes a year – around a hundred times today’s entire durable removal output. Meeting those residual emissions through carbon removals alone would require capacity on a scale far beyond the projected capacity.
That is why carbon removals should be viewed as the final step in a decarbonisation strategy, not the first.
The priority should always be to reduce emissions wherever possible. For AI, that means improving the efficiency of models and infrastructure, powering data centres with genuinely renewable electricity, reducing reliance on fossil-fuelled grids and using AI where it delivers meaningful value rather than treating it as the default solution for every task.
Reduction before removal
Transparency also has a critical role to play. Organisations need a much clearer understanding of the environmental impacts associated with AI, including energy consumption, water use, and embodied emissions across infrastructure and hardware. Without consistent measurement and reporting, it becomes difficult to distinguish genuine emissions reductions from accounting improvements.
This isn’t only an emissions story. In the US, a significant number of data centres are being built in areas already under water stress, with much of the country currently experiencing some level of drought. That pressure is part of why some states, including California, have tightened restrictions on new data-centre developments, moving from a moratorium towards stricter controls. Data centres also cause ‘data heat islands’, with air temperatures surrounding them up to 9C higher due to the hot waste air they expel. Genuine decarbonisation cannot be separated from where, and how, this infrastructure gets sited.
The same scrutiny that applies to AI’s energy use should apply to carbon removals themselves. As demand grows, quality, durability and additionality will become increasingly important. Carbon removals should support credible transition plans, not substitute for them.
Relevant: AI Has Become Marketing’s Biggest Sustainability Blind Spot, New Survey Shows
This is particularly relevant as AI becomes embedded in business operations. Organisations understandably want to embrace the productivity benefits, but they should also be asking where it creates genuine value, and what emissions remain once reduction opportunities have been exhausted.
Carbon removals have an essential role to play in limiting global heating. But they cannot become a licence for unconstrained emissions growth.
The conversation should not be about choosing between decarbonisation and carbon removals. It should be about putting them in the right order.
First invest in low-carbon data centres, including green steel, low-carbon cement, self-generated renewable energy and a circular approach to servers. Then invest in CDR. Finally, aligned with the Science Based Targets initiative’s Net Zero Standard, use carbon removals only for genuinely residual emissions.
If we reverse that order, we risk turning carbon removals into a climate IOU – one that no amount of future innovation will be able to repay.
