Big Tech's Climate Pledges Are Breaking Under the Weight of AI Power Demand

The numbers are in. They are not close.

Google's greenhouse gas emissions rose roughly 50% in the first five years after it set its climate commitments. Amazon's climbed 33%. Microsoft's exceeded 23%. Meta's surged past 60%. These are not minor deviations. They represent a systematic reversal of a decade's worth of corporate sustainability positioning, and every percentage point of that reversal runs through data center power infrastructure.

The proximate cause is not complicated: AI inference and training workloads require continuous, uninterruptible power at a scale that renewable generation cannot yet reliably deliver. The grid simply was not built for this. So the hyperscalers are buying what's available. Natural gas now accounts for more than 40% of US data center electricity. Globally, coal still covers roughly 30%.

Patrick Huang, an analyst at Wood Mackenzie, put it plainly: "They are starting to acknowledge that, 'Yeah, we're maybe not on track.'"

That is an understatement. They are not on track.

The Scale of the Problem

Data centers consumed 4.6% of total US electricity in 2024. That figure could nearly triple by 2028. US electricity demand overall could rise 20% over the next decade, with data centers as the primary driver. AI was responsible for a 2.4% uptick in US fossil fuel emissions last year. One sector, one application category, moving national emissions in measurable increments.

The companies themselves are retreating from their own language. Google originally committed to 100% clean energy by 2030. It now describes that target as a "moonshot." Microsoft still holds its 2030 carbon removal goal but frames it as "a marathon, not a sprint." Both companies are burning natural gas at plants that they offset with solar capacity built elsewhere on the grid. The accounting works on paper. The physics does not care about the accounting.

Julie McNamara, a senior energy analyst at the Union of Concerned Scientists, described the dynamic with precision: "Tech companies are allowing implicitly or explicitly an enormous increase in fossil fuel dependence."

Meta's Louisiana Bet Is a Preview

In January, Meta announced plans to build 10 gas-fired power plants to supply its Hyperion campus in Louisiana. The total generation capacity: 7.5 gigawatts. Estimated cost: approximately $11 billion. This is not a transition strategy. This is a long-term fossil fuel commitment made explicitly because the renewable supply chain cannot deliver firm power at this scale on the timeline that AI infrastructure requires.

The federal policy environment is accelerating that commitment. The Trump administration canceled solar and wind grants and moved to eliminate renewable energy tax breaks, with those breaks set to expire in July. The economic math that made renewable procurement attractive to hyperscalers for the past decade is being rewritten. Gas is cheaper, faster to permit in many jurisdictions, and delivers the dispatchable capacity that AI workloads demand around the clock.

The result: the largest technology companies in the world are becoming, structurally, natural gas utilities. That is not hyperbole. That is what the capital expenditures describe.

The Thermal Infrastructure Cost Nobody Is Pricing In

Here is what gets lost in the emissions coverage. Every megawatt of gas-fired generation carries a thermodynamic penalty that compounds the cooling problem, not just at the power plant but inside the data center itself.

Natural gas combined-cycle plants operate at roughly 40 to 60% thermal efficiency. That means for every unit of chemical energy burned, 40 to 60% becomes electricity and the remainder becomes waste heat rejected to the atmosphere through cooling towers, once-through water systems, or air-cooled condensers at the generation site. The electricity that survives that conversion then travels to the data center, where cooling systems consume up to 40% of facility power to manage the heat generated by compute loads. Two separate heat rejection loops, both scaling simultaneously.

AI workloads are driving rack densities from the traditional 10 to 15 kilowatts per rack into the 30 to 100 kilowatt range for GPU clusters. The cooling systems designed for the previous generation of compute density are not adequate. Facilities that are now committing to gas-fired power for 10 to 20 year contract terms will spend those years managing a heat rejection problem that grows with every hardware refresh cycle.

The operators building new campuses today on gas power are locking in two costs at once: the fuel cost and the cooling infrastructure cost. Mechanical plants designed for air-cooled, moderate-density IT loads will require complete redesigns as GPU rack densities climb. The capital spent on that retrofit will not appear in the carbon accounting. It will appear in the operational expenditure line, year after year.

What the Cooling Industry Should Watch

The pivot to fossil fuels is not the end of the clean energy story for data centers. It is a detour with a defined duration. The economics of utility-scale solar and battery storage are still improving. Geothermal is attracting serious investment. Small modular nuclear reactors remain years away but are moving from concept to permitting in several US states.

The window that matters for cooling infrastructure professionals is the next five to eight years: the period during which AI data centers are being built at maximum velocity using fossil-backed power, while the hardware inside those facilities continues to increase in thermal density. The facilities going in the ground right now will need cooling systems capable of handling the thermal loads of 2030 and 2035 hardware generations, not just today's.

Gas-fired power does not change the physics inside the rack. It changes the policy optics and the grid supply equation. The heat is still there. In fact, there is more of it. Every additional megawatt of gas generation means more total energy entering the system, more conversion losses, and more waste heat arriving at the facility boundary for mechanical systems to reject.

The companies that built the last generation of data center cooling infrastructure on the assumption that clean energy commitments would hold, and that rack densities would grow modestly, are already facing the consequences of those assumptions. The next set of decisions, about cooling topology, fluid infrastructure, and heat rejection capacity, will be made against a backdrop of aggressive fossil fuel expansion and GPU cluster densities that were not in any 2022 planning model.

The climate commitments are breaking. The thermal loads are not going with them.