The hidden cost of AI: how big tech's climate promises are collapsing

Google promised to be carbon neutral. Microsoft pledged to go carbon negative by 2030. Then they discovered AI.

In its 2024 environmental report, Google quietly disclosed that its greenhouse gas emissions had surged 48 per cent since 2019. The cause: data centres for AI. Microsoft made a similar admission soon after, with emissions up 29 per cent since 2020 and every climate commitment it had made now under threat.

These are not startups that never thought about sustainability. They are among the most valuable companies on Earth, with entire departments dedicated to environmental responsibility. They made bold climate pledges in front of shareholders and the public. Then they built AI systems that made those pledges impossible to keep.

The tell

The numbers are not the story. The decision behind the numbers is.

Climate targets do not slip 48 per cent in five years by accident. Someone at Google and someone at Microsoft signed off on AI expansion knowing it would break the public commitment. The bigger of the two carbon footprints in this article belongs to a company whose chief executive has spent a decade telling investors that climate is a board-level priority. The targets did not change. The behaviour did. Sustainability quietly demoted itself from a constraint to a reporting obligation.

That is the actual event. Not the emissions; the trade.

The pattern

Companies missing self-set climate targets is not new. The pattern of missing them because a more lucrative race opened up is not new either.

Through the 2010s, oil majors set net-zero targets and quietly walked them back when the economics of drilling shifted. Through the 2000s, fast fashion announced sustainable cotton programmes and then expanded turnover faster than the supply chains could clean themselves. The shape is consistent: a public pledge made in a calmer commercial moment, followed by an opportunity that the pledge would have cost too much to honour, followed by carefully worded admissions that the targets are now "challenging".

The AI version is faster than the others, and the actors involved are better resourced than the others, but the move is the same. Google's own environmental report concedes that "as we further integrate AI into our products, reducing emissions may be challenging". Microsoft's chief sustainability officer told the public the company has "a lot of work to do" to meet its targets. Translated out of corporate-speak: we have decided to let the emissions rise.

The mechanic behind the rise

The mechanism is straightforward, and that is what makes it durable.

Training a single large AI model can produce as much carbon as five cars over their entire lifetimes. That is just the training phase, before a single user query. Once deployed, every interaction consumes more energy. Millions of users asking chatbots questions, requesting image generations, running AI-assisted searches: each request hits servers that need power and cooling. Global data centre electricity consumption is projected to double by 2026, and AI is the primary driver. Water consumption for cooling has become significant enough that data centres are straining resources in already water-stressed regions.

Layered on top is an accounting fiction that lets the headline numbers look better than the physics. When a hyperscaler says a data centre runs on renewable energy, that often means the company has bought renewable energy credits somewhere else, not that the actual facility is powered by clean electricity. A data centre in Virginia might be drawing from a grid that includes coal and natural gas while the operator holds credits from a wind farm in Texas. The electrons powering the AI query come from the local mix. The clean-energy claim comes from the certificate.

True renewable alignment would require data centres to run on clean power in real time, matched to actual consumption. That is technically harder and more expensive. So companies settle for the accounting version instead.

There is a second silence underneath the first. Properly optimised models can reduce computational requirements by 90 per cent or more for many use cases. Smaller models, trained thoughtfully, can deliver comparable results to massive ones. Techniques like distillation, quantisation, and intelligent caching could slash energy consumption. The research exists. It is just not where the money is going. Bigger models make better demos, generate more impressive benchmarks, and are easier to market. Every company racing to build the biggest model is also racing toward the highest energy consumption.

And the disclosure is selective. Tech companies report emissions in three scopes: direct, purchased electricity, and everything else in the supply chain. The 48 per cent and 29 per cent headline numbers are largely scope 2, the electricity. The full lifecycle impact, including the rare-earth extraction, chemical processing, global shipping, and rapid obsolescence cycle of AI accelerators, lives mostly in scope 3, where the numbers are harder to bound and easier to leave out.

Who loses

Three groups carry the cost of this trade. None of them sat in the room when it was made.

Communities living next to AI data centres. A single large data centre can use as much water daily as a small city. In Arizona, parts of Texas, and parts of Spain and the Middle East, AI expansion now competes directly with agricultural and residential needs. Both Google and Microsoft have faced local opposition to data centre projects on the basis of water consumption, and have responded with replenishment pledges and aquifer-restoration programmes. The pledges may be sincere. The physics is not negotiable. Servers generate heat, and removing that heat requires either massive electricity for air conditioning or massive water for evaporative cooling. The household whose bore runs dry next summer is not consoled by an offset purchased on the other side of the country.

The public absorbing the climate externality. The carbon budget, the ceiling on how much more humanity can emit before crossing thresholds we have agreed not to cross, is finite and shrinking. Every tonne of CO2 from an AI data centre is a tonne that cannot be emitted elsewhere. When two of the most powerful companies on the planet decide that their growth matters more than their climate commitments, they are not absorbing that cost themselves. They are pushing it into the shared budget that everybody else, including future governments and future industries, will have to ration around. The bill arrives later, and not on their balance sheet.

Honest sustainability programmes inside other organisations. When Google and Microsoft tout renewable energy credit purchases as proof of climate progress while their actual emissions climb, they devalue the entire category. Companies that are doing the harder work, matching real-time clean power to real consumption, accepting slower growth as a cost of meeting their pledges, now have to compete for credibility with the loudest accounting fictions in the room. Sustainability teams inside ordinary firms find their credibility eroded by association. The reporting becomes louder; the meaning becomes thinner.

The steelman

The strongest defence of the current trajectory is the one Google and Microsoft both lean on in public: that AI itself will eventually become a climate solution. Used to optimise grids, accelerate materials science, model fusion reactors, design more efficient buildings, AI could in principle unlock emissions reductions that dwarf its own footprint. On this reading, the present surge in data centre emissions is an investment. We pay the carbon cost now to fund the capability that lets us pay it back many times over later.

Charitably stated, this is not a trivial argument. There are real research programmes pointing in that direction. Some of them are inside the same companies whose emissions are climbing.

The argument fails on the gap between actual and theoretical. The emissions are real now. The benefits are projected. The companies invoking the future-solutions framing are the same ones that, by their own admission, have already decided to miss the targets they set when they expected to do the cleaning up themselves. Worse, the trade is not being acknowledged or debated. There is no public accounting that says: we will emit X tonnes more carbon in service of capability Y, against which we expect to recover Z by year W. There is just a rising emissions line and a rotating set of explanations. A trade-off that is never named is not a trade-off. It is a transfer.

What healthy looks like

It is not that solutions do not exist. It is that implementing them would mean accepting limits.

• Treat efficiency as a first-class metric. Compete on performance per watt, not just on benchmark scores. Optimise models for efficiency before deployment, not after.
• Place data centres where renewable energy is genuinely available, not just where credits can be purchased. Match clean power to consumption in real time.
• Extend hardware lifecycles. Slow the chip-replacement treadmill. Account honestly for the manufacturing and disposal emissions in scope 3.
• Apply proportionality. An AI tool that helps grid operators avoid blackouts and an AI tool that generates novelty images of dogs in hats do not deserve the same environmental subsidy. Some applications justify their cost; some do not.
• Hold leaders to climate targets with the same rigour as financial targets. If a quarterly earnings miss has consequences, a quarterly emissions miss should too.

None of this would stop AI development. It would make it accountable.

The canary

Allbirds was the canary for the speculative side of the AI economy. Google and Microsoft are the canary for the environmental side, and they are a much louder bird.

These two companies have the largest sustainability teams, the most sophisticated reporting infrastructure, the most committed public language, and the deepest balance sheets of any operators in the space. If they cannot keep their climate pledges under AI growth pressure, no smaller actor will. The mid-tier cloud providers, the AI-first startups burning compute on borrowed credit, the enterprises rolling generative tools into every product line: none of them have the carbon-accounting muscle Google and Microsoft have, and none of them are facing softer commercial pressure. They will follow the same path, with less transparency about it.

The Google and Microsoft admissions are not the worst of what the industry is doing to its environmental commitments. They are the most visible. The rest of the sector is watching the two leaders quietly walk back their pledges and learning that the public will accept it.

The environmental pillar

The environmental pillar of responsible AI is the one most easily moved off the table. It does not produce immediate user harm in the way an unsafe model does. It does not generate a regulator's letter the way a compliance breach does. It accumulates, somewhere else, on someone else's budget. That is precisely why it is the test.

Google and Microsoft have not failed at sustainability because the engineering was too hard or the technology too immature. They have failed because, at the moment of choosing between the climate commitment and the AI commitment, they chose AI and let the emissions land where they would. That choice was made by named executives in named meetings, and it has been ratified every quarter since by the absence of any decision to reverse it.

AI which cannot account for its own environmental cost has not been deployed responsibly. It has just been deployed.

---

Sources: Google 2024 Environmental Report; public statements from Microsoft's chief sustainability officer on the company's climate trajectory; reporting on data centre electricity and water consumption referenced in the EU AI Act discussions and broader public reporting through 2024 to 2026.