NextFin News - Microsoft, Amazon and Google are discovering that the carbon cost of artificial intelligence is no longer an abstract future risk. In their latest sustainability disclosures, the three companies said they emitted 119 million metric tons of carbon dioxide equivalent in the fiscal year ending March 2026, up from about 101 million metric tons a year earlier. On the French government’s latest provisional estimate, France emitted 369.2 million metric tons in 2024 excluding land-use changes. The tech trio’s combined footprint is therefore now roughly one-third of France’s territorial total. The question is no longer whether AI raises emissions. It is whether the industry can keep promising net zero while its physical buildout is rewriting the emissions curve.
The scale of the jump matters, but so does where it is coming from. Microsoft said its emissions rose 25% to 20 million metric tons in fiscal 2025, and it said the increase was driven primarily by expansion of its datacentre infrastructure. Google said its emissions rose 18% in 2025 as supply-chain activity accelerated to support rapid business expansion. Amazon said its emissions increased 16% in 2025, with supply-chain emissions up 20% and datacentre building and construction explicitly included in that rise. All three companies still say they are aiming at net zero: Google and Microsoft by 2030, Amazon by 2040. But the latest numbers show that the bottleneck is moving from software growth to the industrial footprint needed to support it.
That shift changes the story. It is tempting to read the emissions spike as a simple electricity problem: more AI means more servers, more servers mean more electricity, and more electricity means more carbon unless the grid decarbonises quickly enough. Yet the reports show a broader mechanism. The higher emissions are being driven not only by electricity use but by embodied carbon in construction, hardware and supply chains. Microsoft’s report says the global shift toward AI is increasing demand for energy, water, land and materials. Google says its emissions rose because of supply-chain activities supporting business expansion. Amazon points directly to datacentre construction. The carbon problem is now being loaded into capital expenditure before a facility has delivered a single extra query.
That matters because second-order effects are starting to dominate the first-order story. The obvious first-order effect is higher power demand and higher emissions. The less obvious second-order effect is that climate pressure is migrating into procurement, permitting and capital intensity. The market can model power contracts. It is much harder to value the time and cost of building enough low-carbon infrastructure, the availability of carbon removals, or the political friction that comes with large new datacentres. In that sense, the climate question is becoming a capex question, and the capex question is becoming a margin question. If that capex keeps climbing faster than revenues, the emissions issue becomes inseparable from the economics of AI scale.
The Emissions Surge Looks Structural, Not Cyclical
The first answer to the market’s question is that this does not look like a cyclical spike that will mean-revert on its own. A cyclical emissions burst would usually reflect a temporary burst of build activity followed by flattening as projects mature and capex normalises. But the current AI wave is not a one-off cycle in that sense. It is a structural change in the architecture of digital infrastructure. More compute requires more datacentre floor space, more chips, more cooling, more grid connections, more power equipment and more industrial inputs. The emissions path follows the model, not the quarter.
The strongest evidence for a structural read is that the companies themselves describe the problem in structural language. Microsoft says AI infrastructure is driving demand for energy, water, land and materials. Google says its buildout is accelerating faster than the grid is decarbonising. Amazon says AI is creating new demands for energy, water and infrastructure. Those are not temporary operating glitches. They describe a new production function for cloud and AI: the physical system required to scale the digital product has become much more carbon intensive.
“AI infrastructure is driving demand for energy, water, land, and materials.”
That line from Microsoft is the core mechanism. The emissions story is no longer confined to the server room. It now runs through concrete, steel, semiconductors, transport, land use, cooling and grid upgrades. The more useful question is not whether firms can buy cleaner electricity for today’s load, but whether they can reduce the embodied emissions locked into each new unit of capacity. That is a much harder problem, and it is why the comparison with France lands so sharply. France is a national emissions profile. Microsoft, Amazon and Google together are now close enough to one-third of that number to make their climate footprint look macroeconomic rather than corporate.
This is also where the second-order effect sits. The obvious first-order effect is higher power demand and higher emissions. The less obvious second-order effect is that climate pressure is migrating into procurement, permitting and capital intensity. The market can model power contracts. It is much harder to value the time and cost of building enough low-carbon infrastructure, the availability of carbon removals, or the political friction that comes with large new datacentres. In that sense, the climate issue is becoming a capex question, and the capex question is becoming a margin question. The companies are not just buying electricity; they are financing an industrial transition that may take longer than the AI cycle itself.
That makes the conventional bull case incomplete. The standard argument says today’s emissions are the price of building tomorrow’s efficient AI economy. That may be partly true. But a buildout that expands faster than the grid decarbonises can still raise absolute emissions for years, even if unit emissions improve. The issue is scale. Efficiency gains have to outrun demand growth. So far, the companies’ own disclosures suggest the opposite is happening. And because the emissions come in part from supply chains, a cleaner power contract alone cannot solve it. The carbon has already been embedded in the stuff that builds the server farm.
There is a second layer to this argument. Once emissions become embedded in capex, the story shifts from operational sustainability to strategic resilience. A company can sign more renewable contracts, but it cannot easily accelerate the delivery of transformers, concrete, cooling equipment and chip manufacturing capacity. Those constraints raise the cost of the next increment of compute. So the emissions problem and the infrastructure bottleneck are the same problem seen from two sides. That is why the carbon issue is now a business issue, not just a disclosure issue.
The Strongest Counter-Case Is That This Is Only a Transitional Buildout
The best argument against the structural view is that the current emissions surge is front-loaded. Datacentres are expensive to build, but once they are built, they can run on cleaner electricity, use more efficient chips and cooling systems, and buy carbon removals. On that view, the emissions jump reflects a one-time industrial expansion that should fade as the facilities mature and the power mix gets cleaner. In other words, the carbon spike could be the cost of entry into a more efficient digital era, not evidence of a lasting deterioration.
That counter-thesis deserves respect because there is real evidence that efficiency matters. Google says it matched 100% of its electricity consumption with renewable energy purchases on a global annual basis for the ninth consecutive year, while also reducing operational emissions by 2% year over year. Amazon says it is one of the largest corporate purchasers of renewable energy and one of the most energy-efficient datacentre operators in the world. Microsoft says it is investing in carbon dioxide removal and cleaner infrastructure. These are not empty claims. They show that the companies are already pushing on the right levers.
But the counter-case still does not remove the central problem. The cleanest part of the system is electricity procurement. The hardest part is the embodied carbon in the materials and supply chains needed to expand capacity. Google’s report says its AI infrastructure buildout is currently accelerating faster than the grid is decarbonising. Amazon says AI is creating new demands for energy, water and infrastructure even as it keeps investing toward its 2040 net-zero pledge. That means the current emissions pattern is not just a gap between ambition and reality. It is a consequence of scaling physical infrastructure faster than the low-carbon system around it.
If the transitional view is right, the falsifying signal is concrete: the next reporting cycle should show that scope 3 emissions and construction-related emissions flatten or fall even as datacentre capacity keeps rising. If Microsoft, Google and Amazon expand AI infrastructure again while keeping supply-chain emissions flat or lower year on year, the structural thesis weakens. If supply-chain emissions keep rising with capacity, the structural view wins. A second useful marker would be the ratio of operational emissions to embodied emissions: if operational emissions dominate because new facilities are powered by cleaner grids, the headline problem eases; if embodied emissions keep absorbing a growing share of the total, the problem is becoming more industrial than electrical.
There is also a broader market implication. Investors and policymakers often focus on operational emissions because they are easier to measure and easier to manage. But if the carbon burden is shifting upstream, the economic bottleneck shifts too. That puts more weight on industrial policy, grid buildout, materials science and carbon-market credibility than on simple efficiency slogans. The climate issue becomes less about where servers sit and more about how much carbon is embedded in each new layer of digital infrastructure. It also means the companies’ climate claims will be judged less by the watt-hour and more by the ton of steel, concrete and hardware they bring to market.
One more point is easy to miss. The market often assumes that emissions intensity can improve fast enough to leave the absolute total untouched. That is only true if demand grows modestly. In AI, demand is not modest. It is compounding into a larger system of models, inference, training, storage and networking. Even a 2% operational emissions reduction, like Google’s, can be overwhelmed by a 37% jump in electricity demand. That is the mathematical reason the story has become structural. Intensity can fall while the total rises.
Who Benefits, Who Is Exposed, And What Changes Next
In the short term, the winners are the companies that supply the physical picks and shovels of the AI cycle: datacentre developers, power-equipment makers, electrical infrastructure suppliers, grid operators, cooling specialists and firms that can bring land, permits and capital together quickly. The exposed companies are the hyperscalers themselves, because they must keep funding the buildout while defending climate targets that are becoming harder to reconcile with absolute emissions growth. Energy suppliers can also benefit, especially where new datacentres lock in long-duration demand. The climate risk is that the carbon footprint becomes a feature of growth, not a bug.
In the medium term, the key variable is whether efficiency gains can outrun scale. If AI workloads keep multiplying faster than hardware gets cleaner, absolute emissions can keep rising even if emissions per unit of compute fall. That is the rebound problem in its modern form. The companies can improve the efficiency of every server and still post higher total emissions if demand grows faster than those gains. That is why the headline numbers matter more than the intensity metrics.
In the long term, the problem is structural because the AI economy is physical. It requires more power, more materials, more land and more water. That does not reverse itself. It requires cleaner grids, better chips, lower-carbon construction materials, better cooling, stronger permitting and a carbon market that can actually absorb residual emissions at scale. Until those pieces catch up, the companies’ net-zero targets will depend on a system that is smaller and less certain than the infrastructure boom they are funding. The practical issue is not whether they can announce more climate ambition. It is whether they can make the next hundred datacentres cleaner than the last hundred while demand is still climbing.
The base case is that the companies keep expanding AI capacity and keep looking for ways to slow the emissions growth rate through cleaner electricity, better hardware and carbon removals. The upside case for climate progress is a faster-than-expected fall in embodied carbon per unit of capacity, which would show up in flatter supply-chain emissions even as capacity rises. The downside case is that AI demand keeps outpacing decarbonisation, making absolute emissions growth a durable feature of the sector. If the next reports show higher total emissions again but slower growth than this year, the market may read that as progress; if the supply-chain line keeps climbing, it should read it as evidence that the buildout is still outrunning the fix.
That is why the France comparison matters. It is not a stunt. It is a sign that the carbon footprint of AI infrastructure is becoming large enough to look like a country-level emissions problem. The big technology firms are still talking about net zero. The harder question is whether their physical expansion plan can still live inside that promise.
The carbon problem is moving from the server room to the construction site. Construction is much harder to wish away.
Explore more exclusive insights at nextfin.ai.
