The cloud may be the most successful piece of branding in modern technology.
It sounds light. Clean. Almost weightless. Your files float into it, an AI model thinks somewhere inside it, and the answer returns a few seconds later.
The physical version is less poetic.
It is concrete, steel, cooling equipment, transformers, transmission lines, backup generators, thousands of servers, and an electricity meter moving at a speed that would make most households feel physically unwell.
I recently read a Financial Times analysis of 60 major data centres planned in the United States by Amazon, Microsoft, Google, and Meta. Based on the current regional power mix, those facilities could produce 101.5 million tonnes of carbon dioxide every year once fully operational.
That is roughly equivalent to:
7% of US power-sector emissions in 2025
27 coal-fired power plants
24 million petrol-powered cars
Those comparisons are large enough to produce a dramatic headline, but the more interesting story sits underneath them.
Big Tech is not simply consuming more electricity. Its demand is arriving so quickly, and at such enormous scale, that utilities are changing what they plan to build. Renewable projects are still being added, but so are new gas plants. Some coal retirements are being delayed. In several locations, fossil-fuel generation is being built specifically to serve hyperscale data centres.
The AI race is quietly becoming an energy infrastructure race.
And servers, inconveniently, do not run on press releases.
First, the 101.5 million tonnes is a scenario
It is important to be precise here.
The FT figure is not a guaranteed forecast of what these data centres will emit. It is an estimate of what their annual emissions could look like if all 60 projects are completed and powered using the latest available snapshot of their regional electricity grids.
The analysis used:
Regional emissions factors from the US Environmental Protection Agency's 2023 eGRID data
Power usage effectiveness figures reported by the companies
A 70% load factor
Carbon dioxide emissions only
This means the final number can move in either direction.
If the grid adds enough genuinely new clean electricity, emissions could be much lower. The FT model shows substantial reductions under scenarios with 25%, 50%, or 70% more clean generation than the existing mix.
If clean-energy projects are delayed, electricity demand exceeds current forecasts, or gas and coal plants run more often at the margin, emissions could remain high—or the average-grid calculation could even understate the short-term impact.
So the honest reading is not, “These facilities will definitely emit exactly 101.5 million tonnes.”
It is this:
With today's electricity system, the planned expansion is large enough to become a new emissions source on the scale of dozens of coal plants.
That is still a serious result.
The data centre is only as clean as the next power plant
Technology companies often say that their electricity use is matched with renewable energy. That can be true in an accounting sense while the physical grid serving a facility still depends on fossil fuels during many hours of the year.
Imagine a data centre consuming electricity continuously in one state while its operator buys renewable energy credits linked to a wind or solar project somewhere else. Across the year, the purchased clean energy may equal the facility's total consumption.
But the server does not pause when the sun goes down or the wind weakens.
At that moment, the local grid must supply whatever generation is available. If the marginal source responding to the extra demand is a gas or coal plant, the additional computing load is still associated with additional emissions—even if the company's annual spreadsheet balances nicely.
This is the difference between annual matching and hour-by-hour carbon-free electricity.
Annual matching has helped finance a huge amount of renewable capacity and should not be dismissed as meaningless. The problem is that it does not prove a data centre is running on clean power every hour, in the same region, or at the exact moment its demand affects the grid.
The faster electricity demand grows, the more visible that gap becomes.
Utilities are reaching for gas
According to the FT analysis, three-quarters of the utilities serving the 60 planned facilities are building or planning additional gas-fired generation. Among utilities that still operate coal plants, roughly a third are delaying retirements.
In 17% of the cases, utilities explicitly told regulators that new gas capacity was being built to satisfy demand from a particular hyperscale data centre.
This is not happening because renewable energy disappeared or because every grid operator suddenly developed an emotional attachment to gas turbines.
It is happening because utilities face three uncomfortable requirements at once:
The new demand is enormous.
Data centres expect power around the clock.
The facilities are often being completed faster than new transmission, storage, and clean generation can be connected.
Gas is dispatchable, familiar, and relatively quick to build. From a utility planner's perspective, it is the obvious short-term answer to a customer requesting hundreds—or even thousands—of megawatts with little tolerance for interruptions.
The climate problem is that a short-term answer can operate for 30 years.
Global Energy Monitor found that US gas-fired capacity in development nearly tripled during 2025 to about 252 gigawatts. More than a third is intended to power data centres directly on site, and completing the full pipeline would expand the existing US gas fleet by almost 50%.
Not every announced project will be built. Many are still at an early stage, and demand forecasts can be wildly optimistic. But once a gas plant is financed, constructed, and connected to a major customer, the pressure to keep using it does not vanish when a cleaner option appears.
That is how an AI infrastructure boom can create a fossil-fuel lock-in.
Big Tech's climate promises are meeting Big Tech's growth
The companies involved are not ignoring the problem.
Amazon, Microsoft, Google, and Meta have invested billions in renewable energy, storage, grid agreements, carbon removal, advanced nuclear power, and other decarbonisation technologies. These efforts have helped create real clean-energy capacity.
The awkward part is that their businesses are growing faster than many of those improvements can compensate for.
Recent company disclosures show the tension:
| Company | Latest reported change | Main pressure mentioned |
|---|---|---|
| Amazon | Total emissions up 16% from 2024 to 2025 | Data centre construction and delivery fuel |
| Microsoft | Total emissions up 25% year over year | Expansion of data centre infrastructure |
| Alphabet | Adjusted “ambition-based” emissions up 18% | Supply-chain activity supporting rapid expansion |
Amazon reports the 16% increase in its 2025 Sustainability Report. Microsoft directly says its infrastructure expansion was the main cause of its 25% rise in its 2026 climate update. Google's 2026 Environmental Report describes the same basic conflict between exceptional growth and environmental responsibility while also reporting major new clean-energy contracts.
This does not mean the climate programmes are fake. In some cases, emissions would have been far higher without them.
It means efficiency and renewable purchases are running up an escalator that keeps accelerating.
A new accelerator may perform far more computations per watt than the previous generation. If the company installs ten times as many accelerators and keeps them busier, total electricity demand can still rise sharply.
That is the rebound problem in a very expensive building.
The grid was not designed for this speed
US electricity demand was relatively flat for years. Grid planning, generation investment, and transmission construction adapted to that world.
Now utilities are receiving requests for individual projects that can consume as much electricity as a small city.
Grid Strategies says utility forecasts point to around 166 GW of additional peak demand by 2030, approximately 20% above estimated 2025 peak load. Data centres account for roughly 55% of that projected growth.
The word “projected” matters.
Some facilities will be delayed. Some will never be completed. Developers may submit requests in several regions while deciding where to build, causing the same future workload to appear more than once in utility plans. Improvements in chips, cooling, and model efficiency may also reduce demand per task.
But utilities cannot wait until every uncertainty disappears. Power plants and transmission lines take years to approve and construct. If planners underestimate demand, they risk reliability problems. If they overestimate it, customers may be left paying for infrastructure that an AI company no longer needs.
That creates a strange situation:
Technology companies want power immediately.
Utilities must plan decades ahead.
AI demand forecasts can change within months.
Gas plants are built to operate for generations.
The timelines do not match.
Clean-energy contracts are necessary, but timing matters
There is a tendency in this debate to choose one of two easy positions.
The first says Big Tech's renewable purchases solve the problem. The second says those purchases are pure greenwashing and solve nothing.
Reality is less satisfying and more useful.
Clean-energy contracts can fund projects that might not otherwise exist. Google says it contracted more than 12 GW of clean energy in 2025, its largest annual total. Microsoft says it matched 100% of its global annual electricity consumption with renewable energy in its 2025 financial year. Amazon says it has matched all electricity consumed by its operations with renewable energy purchased elsewhere.
Those are meaningful investments.
But three details determine whether they keep pace with the new load:
Additionality
Did the contract help build new clean generation, or did it mainly purchase credits from a project that already existed?
Location
Was clean power added to the same grid region in which the data centre created new demand?
Time
Was carbon-free electricity available during the hours the facility consumed it, including nights, low-wind periods, and demand peaks?
The strongest strategy addresses all three. Buying enough renewable certificates to balance a global annual total is easier than supplying a hyperscale facility with additional local carbon-free power every hour.
That harder target is the one the infrastructure boom now requires.
The emissions are not inevitable
The data centre boom does not have to produce the FT's highest-emissions scenario.
There are practical ways to reduce the gap between computing growth and grid readiness.
Build clean power before—or with—the load
Data centre approvals and grid connections can require credible plans for additional generation, transmission, and storage. A facility should not arrive first and leave the utility to find electricity afterwards.
Make computing demand more flexible
Not every workload is urgent. Training runs, batch processing, backups, media conversion, and some background AI jobs can move to cleaner hours or regions. Data centres are enormous loads, but unlike a hospital or a steel furnace, part of their work can be scheduled in software.
Use more than one clean technology
Solar and wind are fast and increasingly inexpensive, but they need storage, transmission, and firm low-carbon generation to support continuous demand. Geothermal, nuclear, long-duration storage, and better regional interconnection can all contribute. None is a magic button, and several will arrive later than the data centres currently under construction.
Reward efficiency without pretending it cancels growth
Better chips, cooling, model architectures, and software reduce energy per task. Companies should publish enough comparable data to show whether total energy and emissions are falling—not only whether each individual computation is becoming more efficient.
Let hyperscalers carry the infrastructure risk
If a utility builds billions of dollars of generation for one speculative customer, ordinary ratepayers should not automatically inherit the cost when the project shrinks or disappears. Long-term contracts, upfront contributions, and minimum payment commitments can place more of that risk on the company creating the demand.
The common theme is simple: the electricity plan must be part of the data centre plan, not a footnote added after the GPUs have been ordered.
Developers are part of this conversation too
Most developers do not choose where a hyperscaler builds a facility or which power plant serves it. We do choose how much compute our products request.
That does not mean measuring personal guilt for every prompt. It means recognising that waste at software scale becomes infrastructure.
A few choices matter:
Route simple tasks to smaller, more efficient models.
Cache repeated context and results when freshness is not required.
Set limits on agent loops, retries, and runaway tool calls.
Batch work that does not need an immediate response.
Measure accepted tasks rather than celebrating token volume.
Delete AI features that exist only because someone wanted an AI feature.
The last one may save more energy than a surprising number of optimisation meetings.
This is closely connected to the AI price discussion. A wasteful workflow is usually expensive in both money and energy. Cost is not a perfect carbon metric—the same token can be generated by different hardware on very different grids—but efficient software at least avoids consuming electricity for work nobody needed.
The next AI benchmark may be a power connection
For the last few years, the AI industry has competed on model quality, chip supply, funding, and talent.
Power availability is becoming another competitive advantage.
A company can buy accelerators and build a data hall more quickly than a region can approve transmission lines, connect renewable projects, or construct firm clean generation. That means the winning location may not be the one with the cheapest land or the most generous tax incentive. It may be the one where reliable electricity can actually arrive.
This changes the meaning of scale.
An AI company is no longer scaling only software. It is scaling a physical energy system around the software. The larger the facility becomes, the harder it is to separate product strategy from utility planning, environmental policy, and local infrastructure.
The US Energy Information Administration reported that energy-related CO₂ emissions rose by about 2% in 2025, with increased electricity demand contributing to the rise in power-sector emissions. Its Annual Energy Outlook now treats data centre server use as a major driver of future electricity consumption.
The cloud has acquired a very visible footprint.
This is an infrastructure scheduling problem
I do not think the useful conclusion is that AI must stop growing. The technology is not going to disappear, and data centres also run the cloud services, databases, video platforms, business tools, and internet infrastructure we already use every day.
The useful conclusion is that the schedules have to match.
If a data centre takes two years to build while the clean generation and transmission needed to serve it take seven, the missing five years will be filled by something. Today, that something is often natural gas—and sometimes a coal plant that was supposed to retire.
Big Tech's climate commitments will be tested not by how much renewable energy the companies purchase in total, but by whether clean supply grows in the right places, at the right hours, as quickly as their electricity demand.
Efficiency matters. New energy contracts matter. Better grids matter. Flexible computing matters.
But none of them matters at press-release scale. They have to work at data-centre scale.
AI may live in the cloud, but the cloud has a postcode, a power connection, and a carbon footprint.
Sources and further reading
Financial Times: Big Tech's data centre boom poised to drive up carbon emissions
Global Energy Monitor: US gas power development and data centres
ACEEE: Faster and Cheaper—Demand-Side Solutions for Rapid Load Growth
If it annoys you twice, turn it into a tool.
See you in the next build.
— Ballwictb
Originally published on ZyVOP
💡 For more articles like this, subscribe to the ZyVOP newsletter!
Top comments (0)