NextFin News - Microsoft, Amazon and Google emitted 119 million metric tonnes of carbon dioxide equivalent in the year ending March 2026, about a third of France's total annual emissions, and nearly a fifth more than the year before. After years of promising to cut their climate footprint, the world's largest technology companies are now driving it up — and the engine is the artificial intelligence data centre build-out that has turned Big Tech into the fastest-growing source of electricity demand in advanced economies.
The reversal is not a one-off accounting blip. It is the first visible consequence of a structural shift in how the digital economy consumes energy, one that is set to push global data centre carbon emissions higher for the rest of the decade even as the companies insist their net-zero pledges remain intact. The central tension is stark: the same infrastructure spending that is fuelling the AI revolution is also dismantling two decades of efficiency-led emissions discipline in the technology sector.
The Numbers Behind the Reversal
The latest sustainability reports from the three companies document a sharp inflection. Microsoft said its carbon emissions rose 25 per cent over the past year to just over 20 million metric tonnes, "driven primarily by the expansion of our datacentre infrastructure". Its location-based Scope 2 emissions — the measure that reflects the actual carbon intensity of the grids where electricity is consumed, rather than the renewable energy certificates a company buys — jumped 21 per cent, while total electricity consumption climbed 24 per cent to 37 million megawatt-hours. Electricity-related emissions accounted for 13 per cent of Microsoft's total footprint in 2025, up from 2 per cent the year before.
Google reported an 18 per cent rise in overall emissions, attributing it to "increases in supply chain activities that supported the rapid expansion of our business". Its location-based electricity emissions rose 37 per cent from 2024 to 2025, and power consumption has grown 250 per cent since 2019. Amazon's overall emissions increased 16 per cent, with supply chain emissions — which include data centre construction — up 20 per cent, and Scope 2 emissions up 34 per cent. Before this year, Microsoft's emissions had appeared to flatline at about 16 million tonnes in both 2023 and 2024. The break in that trend is the story.
Globally, the picture is consistent. The International Energy Agency estimates data centres consumed 415 terawatt-hours of electricity in 2024, roughly 1.5 per cent of global demand and about 0.5 per cent of global CO2 emissions. That consumption grew about 12 per cent a year from 2017, then accelerated to 17 per cent in 2025 — with AI-focused data centres surging 50 per cent in the same year. The IEA now expects data centre electricity use to more than double to around 945 TWh by 2030, slightly more than Japan's entire electricity consumption today, and to around 1,200 TWh by 2035 in its central scenario.
Capital expenditure tells the same story from the demand side. The largest technology companies spent more than $400 billion in 2025, and that figure is expected to jump a further 75 per cent in 2026. Spending by just five technology companies is now larger than global investment in oil and natural gas production. Goldman Sachs Research forecasts data centre power demand will rise 50 per cent by 2027 and 165 per cent by 2030 compared with 2023, with around 122 gigawatts of capacity online by the end of the decade. Gartner projects AI-optimised servers will account for 44 per cent of data centre power use by 2030, up from 21 per cent in 2025, while the IEA expects accelerated-server electricity use to grow 30 per cent a year against 9 per cent for conventional servers.
Why Efficiency Alone Cannot Fix This
For two decades, the technology sector's environmental case rested on a simple premise: computing would keep getting more efficient, so emissions would fall even as usage grew. That premise is breaking. The efficiency gains are real — the IEA's high-efficiency scenario shows hardware and model improvements could cut data centre electricity demand by 20 per cent by 2035 relative to its central case — but they are being overwhelmed by the volume of new compute being deployed. This is the Jevons paradox in practice: cheaper, more efficient compute does not reduce total energy use; it invites more of it.
"While the decision increases our reported emissions in the near term, it enables us to increase the development of new carbon-free electricity rather than relying on certificates alone," Microsoft said in its environmental report. "We believe this change will create more long-term sustainability benefits. Growth-related emissions pressure was expected."
The mechanism is physical, not financial. A data centre is a 20-to-30-year asset. Once a facility is built, sited and connected to a grid, its electricity demand is locked in for decades. The carbon intensity of that demand depends on one thing: what generates the marginal megawatt-hour on the grid it plugs into. In the United States, data centres already account for nearly half of electricity demand growth between now and 2030, and renewables plus nuclear are not being built fast enough to meet all of it. The IEA expects gas-fired generation for data centres to more than double from 120 TWh in 2024 to 293 TWh by 2035. Morgan Stanley projects the data centre industry will emit 2.5 billion metric tons of CO2 by 2030.
Geography concentrates the problem. In Ireland, data centres already use around 21 per cent of national electricity, a share the IEA expects to reach 32 per cent by 2026. In the US state of Virginia, the figure is 26 per cent; in Dublin, it is 79 per cent. At these levels, data centres are no longer a marginal load — they are the grid's dominant new customer, with the bargaining power to sign long-term gas contracts and the urgency to accept fossil power when clean electrons are unavailable.
The Clean-Energy Offset Is Real, But It Is Not Enough
The counter-argument is not weak. The IEA expects renewables and nuclear to supply nearly 60 per cent of data centre electricity by 2030, up from 35 per cent today. In Europe, renewables and nuclear are set to supply 85 per cent of the additional electricity data centres require by 2030. The technology companies point to these figures, and to their renewable energy purchases, as evidence that growth and decarbonisation can coexist.
But the offset argument contains two flaws. First, the baseline is small: data centres currently represent just over 1 per cent of global electricity demand and 0.5 per cent of CO2 emissions, so a rising clean share still accompanies rising absolute emissions. The IEA's own central scenario has data centre electricity emissions growing from 180 million tonnes today to 300 million tonnes by 2035, and up to 500 million tonnes in a faster-growth "Lift-Off" case. Second, the accounting that lets companies claim 100 per cent renewable power is increasingly under scrutiny. Location-based emissions — which ignore renewable energy certificates and reflect the actual grid mix — are the measure that has surged, and it is the measure that matters for the atmosphere.
There is also the question of what AI saves elsewhere. The IEA estimates that broad application of existing AI-led solutions could deliver emissions reductions equivalent to around 5 per cent of energy-related emissions by 2035 — far larger than data centre emissions themselves. But the agency warns that rebound effects, such as shifts away from public transport to autonomous vehicles, could undercut those gains, and that AI is "not a silver bullet". The savings are potential and diffuse; the data centre load is actual and concentrated.
Cyclical or Structural? This Is a Regime Shift
The critical judgment for investors is whether this emissions rise is cyclical — a temporary bulge that will mean-revert — or structural. The evidence points decisively to structural. Three comparisons make the point.
During the cloud migration wave of 2010-2015, data centre electricity use grew but efficiency gains held emissions roughly flat; workload consolidation into hyper-efficient facilities offset the volume increase. During the cryptocurrency mining boom of 2017-2021, demand spiked and then collapsed when crypto prices fell, because mining rigs could be switched off, moved across borders, or sold — the load was footloose and price-sensitive. Even the dot-com data centre build-out of the late 1990s, which left a legacy of dark fibre and empty shell facilities after the bubble burst, did not lock in decades of power demand the way today's AI campuses do.
The AI build-out is different on every count that matters. The assets are long-lived and immobile: a hyperscale campus with its own substations and cooling infrastructure cannot be packed onto a truck. The demand is tied to enterprise and consumer adoption of AI services that, once embedded in workflows, do not get switched off in a downturn. And the capital commitment is unprecedented: more than $400 billion in 2025, rising 75 per cent in 2026, locks the industry into a multi-year construction pipeline that cannot be unwound without writing off tens of billions of dollars.
The driver will not self-correct. Grid connection queues mean new data centres cannot simply wait for clean power; they connect to what is available, which in many regions is gas. In the United States, the Department of Energy launched a "Speed to Power" programme in September 2025 to fast-track generation and transmission projects, and directed coal and natural gas plants previously slated for retirement to keep operating to ensure grid reliability. That is the clearest possible signal that the energy system is treating data centre load as firm, baseload demand — the kind served by fossil generation, not by certificates.
This is a regime change in the energy profile of the technology sector, not a cyclical fluctuation. Cyclical emissions spikes revert when the driver fades; this driver compounds.
The Second-Order Effects the Market Has Not Priced
The first-order effect — higher emissions — is already in headlines. The second-order effect runs through three channels that are less visible but more consequential.
First, regulation is moving from voluntary pledges to hard constraints. Jurisdictions are beginning to mandate power usage effectiveness standards for new data centres, requiring fresh facilities to achieve a PUE of 1.2 or better, and to scrutinise the grid connections they receive. Europe is requiring energy and environmental management systems for large operators, and several markets are mandating 100 per cent renewable energy use for data centres from 2027. These rules raise the cost and complexity of expansion in the cleanest grids, potentially pushing new capacity toward regions with weaker standards and dirtier power.
Second, the cost of clean power rises when a single new customer class accounts for nearly half of demand growth in a market. The marginal renewable megawatt-hour becomes more expensive for everyone else — industrial users, municipalities, and households all bid against the hyperscalers for the same wind and solar output. Data centres can also become anchors for new low-emissions power projects, which is the bullish version of this channel; but where grid queues are long, the near-term marginal unit is gas.
Third, the technology companies' net-zero credibility is now on the line. Google and Microsoft have pledged carbon neutrality by 2030, Amazon by 2040, and the gap between those pledges and the reported trajectory is widening, not narrowing. A company that is growing emissions 20 per cent a year while promising net zero within four years is either going to miss the target or buy its way there with instruments that regulators are increasingly sceptical of. The market has not yet repriced these pledges as at serious risk of failure.
Who Benefits, Who Is Exposed
The emissions story has a mirror image on the supply side, and it is where the investment implications become concrete. The beneficiaries are the owners of dispatchable power and the builders of the energy infrastructure that data centres require. Natural gas producers and generators gain a new, large, relatively inelastic customer. Nuclear operators and developers gain a premium buyer that values 24-hour carbon-free power and is willing to sign long-term contracts. Utilities with data-centre-heavy load growth see demand rising for the first time in two decades, ending the era of flat or declining electricity sales in advanced economies. Grid equipment makers — transformers, switchgear, transmission builders — gain a multi-year order book.
The exposed parties are the hyperscalers themselves, on two fronts. Their margins face pressure from capital expenditure that must be justified by AI revenues that remain uncertain: one analysis suggests the industry would need roughly $2 trillion in annual revenue to justify sustained $500 billion-a-year infrastructure spending, a target it is on track to miss by hundreds of billions of dollars. And their climate credibility — a recruiting, regulatory and reputational asset built over 15 years — is being spent down. For investors holding both the technology giants and their net-zero narratives, the data centre boom is forcing a choice between the two.
There is also a decarbonisation market opening up. Morgan Stanley argues that the emissions growth will create a large market for decarbonisation solutions: energy-efficient equipment, green building materials, on-site power, and carbon removal. The same forces driving the problem are financing the fix. Whether the fix arrives fast enough is the question that separates the base case from the upside case.
What Could Prove This Wrong
The strongest counter-thesis is that efficiency and clean-power procurement will outpace demand growth, bending the emissions curve down before 2030. Proponents point to the 30 per cent annual efficiency gains in AI compute and to the IEA's projection that the clean-energy share of data centre power will reach 60 per cent by 2030. If that happens, absolute emissions could peak well below the 300 million tonne central-case estimate.
The falsifying signal is specific and observable. If data centre electricity demand growth slows to below 5 per cent annually for two consecutive years, and the clean-energy share of newly connected data centre capacity exceeds 80 per cent by 2028, the structural emissions-rise thesis is materially weakened. Conversely, if hyperscaler capital expenditure remains above $500 billion a year through 2027 while gas remains the marginal supplier in the US and European grids where new capacity connects, the structural call is confirmed.
What to Watch
In the short term, the signal to watch is the quarterly sustainability disclosures from the hyperscalers — specifically location-based Scope 2 emissions, not the market-based figures smoothed by certificate purchases. In the medium term, watch grid connection data in Virginia, Ireland and Germany, and the share of new data centre capacity signing gas-backed power purchase agreements. In the long term, the decisive variable is whether the clean-energy build-out can stay ahead of a demand curve that is compounding at double-digit rates.
The base case is continued emissions growth through 2030, with the clean share rising but absolute tonnes rising faster. The upside case is that AI-driven efficiency gains in the wider economy — grid optimisation, industrial process improvements, logistics — offset more than the data centres emit, making the sector a net climate positive by the early 2030s. The downside case is that grid constraints force an even larger reliance on gas and coal, pushing data centre emissions toward the IEA's 500 million tonne Lift-Off scenario.
The AI revolution was sold on the promise of dematerialisation — intelligence without atoms. The bill, it turns out, arrives in megawatt-hours, and it is due now.
Explore more exclusive insights at nextfin.ai.
