Oct 08

Data Centers’ Hidden 137M-Tonne Carbon Bill

Carbon Before the Servers

That is the CO2e from the ~100 GW of data center capacity planned for 2026–2030 (about 1.1 billion domestic flights’ worth), emitted before a single server switches on. Swapping in 25% GGBS cement and 50% recycled steel could cut about 31 million tonnes.

Everyone is arguing about how many megawatts an AI data center burns once the servers switch on. Almost nobody is talking about what happened before that: thousands of tonnes of CO2e already locked into the steel and concrete, poured and cured months before a single GPU boots up. That carbon shipped already. It is sitting in the foundations. And unlike the electricity bill, nobody is watching it.

Aerial view of a data center building with its cooling plant alongside, car park and fenced grounds
A real data center from above: a plain, shed-like building, with the cooling plant alongside. Photo: Geoffrey Moffett / Unsplash
Figure 1

Where a data center’s embodied carbon actually concentrates, by material.

~65%
Concrete
Illustrative share. Gensler reports concrete can reach up to 80% in data centers.
~30%
Steel
Structural steel, rebar, and framing.
~5%
Everything else
Glazing, finishes, MEP, and other materials.
Illustrative split used throughout this article and the calculator below, not a measured breakdown. It sits inside the “up to 80 percent” concrete range Gensler reports (via Data Center Frontier).

The boom, by the numbers

Data center construction is not a side effect of the AI race, it is the AI race, expressed in rebar. JLL’s 2026 Global Data Center Outlook puts nearly 100 gigawatts of new capacity going up between 2026 and 2030, doubling global capacity to around 200 gigawatts. The Americas hold about half of global capacity and are growing at a projected 17 percent annual clip through 2030, with the US accounting for roughly 90 percent of the region’s capacity. Price tag: about $3 trillion, split between roughly $1.2 trillion of real estate value and $1 to $2 trillion of fit-out for GPUs and networking. Construction costs are climbing too, from $7.7 million per megawatt in 2020 to $10.7 million in 2025, on their way past $11 million next year.

~100 GW
new data center capacity added globally, 2026–2030
$3T
total investment needed to build and equip it, JLL estimate
$11.3M
forecast construction cost per megawatt in 2026
Illustration of a data center under construction, with a finished hall, a steel frame going up, a crane and a concrete truck
The carbon is poured and bolted in long before the servers arrive.

The bill that ships before the electricity does

Embodied carbon: the emissions baked into the building before it does anything at all.

Building materials and construction processes account for roughly 15 percent of global greenhouse gas emissions, according to the 2024 GRESB Real Estate Benchmark, cited by Equinix. Cement alone is responsible for around 8 percent of global CO2 emissions, according to Chatham House. Data centers concentrate that problem: a Gensler report, via Data Center Frontier, found concrete alone can account for up to 80 percent of a data center’s embodied carbon, depending on how much steel is in the mix.

Put an actual number on one building and it stops being abstract. A Purdue University study, “The True Environmental Cost of Data Centers”, measured four real facilities and found an average of about 18,800 tonnes of CO2e baked into construction, before a single server ships. Scale that across just the growth of the US market in a single year, 2025, and the same study projects roughly 1.83 million additional tonnes of CO2e, from structural concrete, steel and site piping alone, based on an expected 25 percent increase in data center construction.

~18,800 t
CO2e embodied in an average data center’s construction, Purdue study
~1.83M t
CO2e added by one year of US data center growth alone, 2025 projection
~14.9M
domestic flights that one year’s growth is equivalent to, at 123kg CO2e per passenger
Before you treat this as gospel

These are averages from a handful of studied buildings and one country’s projected growth, not an audited global total, and the flight comparison uses a single domestic route (Edinburgh to London, per Our World in Data) as its unit. The point is the order of magnitude, not the decimal point: embodied carbon in this buildout is not a rounding error.

Figure 2

Swapping to 75% GGBS cut embodied carbon by 48% on a real project, the Harwell EPAC development.

Standard Portland cement mix2,867 t CO2e
​
75% GGBS mix1,494 t CO2e
​
48% less embodied carbon, same 7,500 m³ of structural concrete.
Source: GGBS Knowledge Share, Harwell EPAC, Southern Construction Framework.

The two decisions that actually move this number

Illustration of a concrete pour over rebar using a lower-carbon GGBS mix
Concrete: swap part of the cement for GGBS.
Illustration of recycled steel beams next to an electric arc furnace
Steel: specify recycled content.

Not the PPA. Not the chip. The spec sheet.

Two material choices do almost all the work here, and both are boring, unglamorous, and entirely within a project team’s control before ground is even broken.

Aisle between rows of equipment cabinets under a steel roof truss in an industrial hall
Equipment cabinets under a steel roof truss, on a concrete floor. Photo: İsmail Enes Ayhan / Unsplash

Steel. Primary, virgin steel made in a blast furnace can carry an embodied carbon footprint up to five times greater than high-recycled-content steel, per the Carbon Leadership Forum’s Carbon Smart Materials Palette. Electric arc furnaces, which run on scrap, emit less than half the CO2 of the older basic oxygen furnace process, and use an average of 93 percent recycled content versus a roughly 30 percent scrap ceiling for BOF steel. Specify recycled-content structural steel and you are not making a marginal improvement, you are picking a different emissions category entirely.

Concrete. Swapping a share of Portland cement for ground granulated blast-furnace slag, GGBS, is not theoretical. On the Harwell EPAC project in the UK, 7,500 cubic meters of structural concrete specified at 75 percent GGBS cut embodied carbon by 48 percent against a standard Portland cement mix, 1,494 tonnes of CO2e instead of 2,867. Microsoft has set a goal of mix designs that cut concrete’s embodied carbon by more than 50 percent, and Meta has developed an AI-optimized mix with 40 percent lower carbon emissions, both via Data Center Frontier.

The catch nobody mentions

Recycled steel is not infinite. The recycled steel available today is only enough to supply around 30 percent of global steel demand, so virgin production is not going away soon. And GGBS, the same material that makes low-carbon concrete possible, is itself a byproduct of primary blast-furnace ironmaking. Push the industry too hard toward recycled steel and electric arc furnaces, and you shrink the supply of the slag that makes low-carbon concrete possible in the first place. These two fixes are quietly tangled together, not independent levers.

Figure 3

Virgin, blast-furnace steel can carry up to five times the embodied carbon of high-recycled-content steel.

Virgin steel (blast furnace, BOF)up to 5×
​
High-recycled-content steel (EAF)1× baseline
​
EAF steel averages 93% recycled content, versus a ~30% ceiling for BOF steel.
Source: Steel, Carbon Smart Materials Palette, Carbon Leadership Forum.
Inside a data hall with overhead cable trays, black server cabinets and a raised floor
Inside a data hall. The servers get the attention, but the building around them holds the embodied carbon. Photo: Taylor Vick / Unsplash

What if the whole buildout used better specs?

Not one building. All ~100 gigawatts of it.

Run the same two decisions, GGBS in the concrete and recycled content in the steel, across the entire pipeline JLL is projecting rather than a single building, and the number stops being a rounding error and starts being a headline. We could not find a published per-megawatt embodied carbon benchmark for data centers, so this is a modeled estimate, not an audited total: it takes Purdue’s own per-facility average, divides by a real average data center size, 13.75 megawatts, the mean of 2,132 US facilities in an arXiv analysis of US data center capacity, to get an implied ~1,370 tonnes of CO2e per megawatt, then applies that rate to JLL’s ~100 GW of new global capacity through 2030.

~137M t
estimated embodied CO2e for the full 2026–2030 buildout, business as usual
~1.1B
equivalent domestic flights, at the same 123kg CO2e per passenger
How this estimate is built, and where it can break

This combines two real but separately sourced numbers: Purdue’s ~18,800 tonne average building, and a 13.75 MW mean facility size from a sample of 2,132 US data centers (median 4.5 MW, so a few very large sites pull the mean up). Purdue’s four buildings are US projects, and the JLL buildout is global, so the scaling is rough. Dividing by the 4.5 MW median instead would give a total several times higher. Treat it as an order of magnitude, not a forecast.

Figure 4

Two levers, applied alone and together, across the entire 2026–2030 buildout, not just one building.

Business as usual: all-virgin steel, all-Portland cement~137M t CO2e
​
25% GGBS concrete, alone~123M t CO2e
​
~14.2 million tonnes of CO2e saved, roughly 116 million equivalent domestic flights.
50% recycled steel, alone~121M t CO2e
​
~16.4 million tonnes of CO2e saved, roughly 134 million equivalent domestic flights.
Both specs, combined~106M t CO2e
​
~31 million tonnes of CO2e saved across the buildout, roughly 250 million equivalent domestic flights. The two levers stack: this is close to the sum of the two savings above.
Modeled estimate: Purdue University (per-building embodied carbon), JLL 2026 Global Data Center Outlook (buildout scale), and an analysis of US data center capacity by facility, arXiv (average facility size).

Try it: what would greener specs actually save?

Two material decisions, at either scale

Modeled on the Purdue study’s ~18,800 tonne CO2e average data center, split an illustrative 65% concrete / 30% steel / 5% other (Gensler reports concrete can be up to 80% in data centers). A different split moves the result: roughly 26 to 36 million tonnes saved at the default settings. Toggle the scope, then move the sliders.

GGBS replacing Portland cement25%

​

Calibrated to Harwell EPAC’s real result: 75% GGBS cut concrete’s carbon by 48%.
Recycled content in structural steel50%

​

Calibrated to the Carbon Smart Materials Palette: full recycled content cuts steel’s carbon by up to 80%.
~31M t
CO2e saved across the buildout
22%
reduction vs. an all-virgin, all-Portland spec
~249.5M
equivalent domestic flights, across the buildout
Illustrative model, not a substitute for a real bill of materials or an EPD. It assumes a 65/30/5 concrete/steel/other split of embodied carbon and linear scaling between 0% and each case study’s calibration point, applied either to one representative data center or across the modeled ~137M tonne pipeline baseline above. Run the actual mix design for a real project on the 2050 Materials platform.

Worth reading, or worth ignoring

Worth reading

  • A concrete mix design with a published Environmental Product Declaration, not just a marketing claim
  • Steel specified by recycled content percentage, not just the word “sustainable”
  • A developer who can tell you the embodied carbon number for the structure, not only the operational PUE

Worth ignoring

  • “This data center is carbon neutral”, based only on a renewable energy contract, with the concrete underneath never mentioned
  • “Sustainable materials” with no percentage, no product, and no way to check it
  • Scope 3 embodied carbon left out of a sustainability report entirely, because operational emissions are easier to measure
The bottom line

AI did not invent embodied carbon, it just supersized the construction schedule fast enough that the industry has to make these material decisions at a pace it has never faced before. The fix is not a breakthrough. It is a spec sheet: how much recycled content, how much GGBS, checked against real product data instead of a press release. That is a decision every one of these hundred gigawatts still gets to make.

Want to talk through what this means for your own numbers?Connect with our team

Before you spec a product for the next build, check its actual embodied carbon rather than take a brochure’s word for it. That is what the 2050 Materials platform is for.

Sources referenced

Data center construction pipeline

Previous Built to Beat the Heat

Related articles

Featured image for
Data & Research Sep 28

Built to Beat the Heat

This summer's drought wasn't a fluke, it was a preview. Here's the material data that actually decides whether a building shrugs off a heatwave or quietly cooks behind a great winter U value: decrement delay, solar reflectance, SHGC, and the real products behind each one.

Read more
Bar chart of cost per tonne of CO2e removed for three embodied carbon swaps: cement replacement with 75% GGBS at £10/t, facade brick buildup at £130/t, and timber vs steel or concrete frame at £190/t
Data & Research Sep 23

The Cost per Tonne Test for Embodied Carbon Swaps

Not every embodied carbon cut is worth paying for. Test any swap on cost per tonne of CO2e, see which band it falls in, and try the free calculator.

Read more
PFAS thumbnail
Data & Research Sep 02

PFAS Compliance Meets Countdown

Sealants, membranes and insulation are next in line for PFAS restrictions. The window to get ahead of it, rather than scramble to catch up, is closing.

Read more