Nadella, Jassy, and Pichai burned a third of France’s carbon and called it progress.
It is true that the three companies’ sustainability reports are not technically fraudulent. Microsoft, Amazon, and Google are, in their annual disclosures, quite specific about what they count, and they are now counting more than they used to. In the financial year ending March 2026, the three firms together reported 119 million metric tonnes of CO₂-equivalent emissions — an increase of about 18 percent in a single year, against roughly 101 million the year prior. The combined number is now close to a third of France’s annual emissions, on a par with a country the size of Czechia. The reports are public; the math is not complicated; the math does not require a credentialed interpreter.
The math, in fact, is the point. Microsoft’s 25 percent year-over-year increase to 20 million tonnes was, in the company’s own accounting, “driven primarily by the expansion of our datacentre infrastructure.” Google’s 18 percent rise was attributed to “increases in supply chain activities that supported the rapid expansion of our business,” which is the same statement in different words. Amazon’s 16 percent overall increase and 20 percent supply-chain increase were, in a paragraph of some editorial nerve, framed as “making progress” toward a 2040 net-zero goal. Three chief executives, three reports, three different adjectives around the same physical fact: the AI buildout is, in real-world terms, a very large infrastructure programme that emits a great deal of carbon, and the carbon number is now large enough to require its own accountant.
The marketing materials will tell you the cloud is a magic place where other people’s computers do your math. What it actually is, in physical terms, is a warehouse in Ashburn or Dublin or Norway, packed with silicon, cooled by water and forced air, and plugged into a grid that, outside of a hydro-dam or a nuclear baseload, is still largely burning carbon. When a company like Microsoft says it is expanding its datacentre infrastructure, it means pouring concrete, laying fibre, installing megawatts of switchgear, and running continuous cooling loops that consume power even when the servers are idle. The physical reality of the AI buildout is that the datacentres are driving US clean energy growth but still fundamentally threatening the climate, because the demand is outstripping the supply of new electrons.
The hyperscalers are on track to spend $765 billion this year, mostly on building the physical infrastructure for artificial intelligence. The Uptime Institute estimates that the big projects announced last year will consume 1.3 per cent of the world’s electricity — a near-doubling of current datacentre demand. This is not a rounding error. This is a material re-allocation of the global energy budget to train and run large language models, which strain both climate progress and water supplies. The gap between what the current generation of large language models demonstrably do in well-defined domains and the trillion-dollar capital expenditure poured into them is the exact interval of the bezzle — that gravity-defying stretch, as John Kenneth Galbraith named it, when the embezzler has his gain and the victim feels no loss. And the physical cost of that bezzle is being dumped into the atmosphere.
Google’s sustainability report notes an 18 per cent increase in emissions over the past year, while simultaneously claiming its AI systems have already reduced emissions elsewhere by 41 million tonnes. Shaolei Ren, an electrical engineering professor at UC Riverside, pointed out that Microsoft’s own filing suggests a shrinking supply of carbon credits on global markets, meaning the virtual offset is running out of actual trees to plant. The mechanism here is worth being precise about. A carbon credit is, at root, an accounting fiction designed to let a company in one jurisdiction pay to avoid an emission in another. When the hyperscalers say they are on track for “net zero” by 2030 or 2040, they are not saying they will stop emitting; they are saying they will buy the remaining accounting credits from a market that the financial sector has already over-allocated. It is the same financial engineering that treats an unmeasured future reduction as a present-day permission structure.
But the credit shortage is only half the story — the other half is that the cloud model is designed to make the emissions disappear from every balance sheet except the hyperscalers’ own. Cecilia Rikap at UCL has identified the structural trick at the heart of this: as corporations migrate to the cloud, they outsource their digital carbon footprint to the hyperscalers, obscuring their own environmental impact. The bank running the algorithmic trading, the defence contractor running the logistics model, the media company running the recommendation engine — none of them report the power draw of the GPU cluster doing the work. They just report their software spend. The hyperscalers absorb the carbon, positioning themselves at the ultimate chokepoint between the corporate economy and the physical grid, and they charge the rest of the economy a toll to enter. It is a perfect mechanism for shifting the physical cost of the digital economy into a blind spot.
The political economy that hollowed Canadian heavy industry in the 1990s — when a foreign acquirer bought a rolling mill in Selkirk, Manitoba, and cared more about quarterly capital allocation than the community that built it — is the exact same political economy now running on the global atmosphere. The playbook is always the same: extract the surplus, lock the workforce in by raising the cost of leaving, find the next class of suppliers, and do it to them. The hyperscalers are extracting the surplus of the physical world — the watts, the water, the rare earths — to build a digital enclosure, and they are charging the rest of the economy a toll to enter. The only difference is that in 1995, the byproduct was laid-off steelworkers. In 2026, the byproduct is 119 million tonnes of carbon.
The structural remedy is not to ask the hyperscalers to write better sustainability reports. It is to treat the datacentre not as a cloud, but as what it actually is: a heavy industrial facility. That means subjecting AI datacentres to the same environmental permitting, grid-impact studies, and carbon-accounting standards as an aluminium smelter or a petrochemical refinery. It means ending the accounting fiction of carbon offsets for Scope 3 emissions, and forcing the companies buying the cloud services to carry the carbon on their own balance sheets. You cannot regulate a magic cloud, but you can regulate a warehouse full of servers.
1,200 warehouses full of servers are not a cloud, and the atmosphere will not wait for the marketing department to retire the metaphor.