AI Data Centers: What the Numbers Say, and Who Made Them

Synergy Research Group, a market research firm tracking cloud and data center markets, publishes a quarterly count of hyperscale data centers, and this March 2025 dispatch covers where that count stood at the end of 2024.

Hyperscale providers operated 1,136 large data centers at the end of 2024, double the count from five years earlier, while total capacity has taken under four years to double on its own. A bar chart tracks the count climbing from roughly 430 in Q4 2017 to 1,136 in Q4 2024, and Synergy credits the growth more to the increasing average size of new facilities than to raw numbers, naming generative AI workloads as a driver of that larger scale.

A companion pie chart splits Q4 2024 capacity by region: the United States holds 54%, China 16%, Europe 15%, the rest of Asia-Pacific 10%, and the rest of the world 5%. Synergy forecasts 130 to 140 new hyperscale facilities a year going forward, with capacity growth still driven more by scale than by count.

Every figure comes from Synergy's own proprietary dataset, tracked commercially rather than published for independent audit, so treat it as one firm's read on the market rather than an audited count. The dispatch sticks to aggregate counts and capacity share too: it says nothing about which facilities are oldest and names no individual site, so look elsewhere for data center history.

arXiv hosts this preprint, an AI-energy coupling framework that models where AI data centers are being built and what that concentration does to the power grids beneath them. It earns catalog space for pairing corporate, policy and media data with quantitative energy-system modeling rather than repeating a single company's roadmap.

The paper's numbers run from 2025 to 2030, and it presents them as projections, not measurements. North America, Western Europe and the Asia-Pacific are modeled to hold more than 90% of new AI compute capacity, and the six leading firms' aggregate electricity use is projected to climb from roughly 118 terawatt-hours in 2024 to somewhere between 239 and 295 terawatt-hours by 2030, about 1% of global power demand. Those figures come out of a forecasting model, not a count, so treat the range as stated uncertainty rather than a settled outcome.

The more specific finding is regional. The authors define their own Power Stress Index to flag where new data-center load meets a less accommodating grid, and by that measure Oregon, Virginia and Ireland score above 0.25, while Texas and Japan absorb comparable new load more easily. The index is the paper's own construction, not an established grid-reliability standard, so its ranking is only as sound as its inputs.

The honest trade-off: this is an arXiv preprint that has not been through peer review, and the abstract page gives you the headline figures while 32 pages of methodology and 8 figures sit behind the PDF link.

OSTI.GOV is the U.S. Department of Energy's public clearinghouse for federally sponsored research, and this page catalogs the 2025 update to Lawrence Berkeley National Laboratory's data center energy usage report.

The page itself is a bibliographic record, not the report — a title, a short abstract, an author list, and a green "View Technical Report" button that links out to the full PDF elsewhere. The abstract is still worth reading on its own: it estimates data centers could draw 11.8% of total U.S. electricity by 2030, with a scenario range running from 9.5% to 15.3%. For comparison, it notes that the prior 2024 edition of the same report put the 2028 figure at 6.7% to 12.0%, which gives a sense of how quickly the projection has moved between editions.

Below the abstract, a details table names Lawrence Berkeley National Laboratory as the research organization and three DOE offices as sponsors, alongside a DOI, an OSTI ID, and eight listed authors. That is enough to cite the work or check its provenance, but there is no methodology, no charts, and no regional breakdown on this page — those live in the linked technical report, a separate download.

This record suits readers who want to verify a data center electricity statistic against its original government source, or who need clean citation details rather than a full read-through. Anyone after the underlying methodology or the supporting charts will need to follow the button through to the report itself.

Patterns, a Cell Press journal, published this analysis by researcher Alex de Vries-Gao in December 2025, and PMC hosts the full text without a paywall, which is useful because most writing about AI's environmental footprint gets cited far more often than it gets actually read.

The piece starts from an admission: while there are established ways to estimate the global power demand of artificial intelligence, the carbon and water footprints of the data centers that host it are far less well characterized. De Vries-Gao works through why that gap exists. Cooling systems draw water directly, but the water and carbon embedded in the electricity a facility buys depends heavily on the local grid mix, and operators rarely publish facility-level numbers. The paper pulls together existing estimation methods and their assumptions rather than presenting new field measurements from inside a data center.

That framing matters for who should read it. This is a literature synthesis and modeling exercise from a single author, not a measurement campaign, so its figures are informed estimates with real uncertainty rather than audited facts. It suits readers who want to understand how footprint numbers for AI infrastructure get produced in the first place, including the assumptions and the gaps where estimates diverge, more than readers hunting for one headline statistic. The trade-off is density: this is a peer-reviewed research article rather than a blog explainer, and it expects some patience with methodology before it gets to a bottom line.

Koomey Analytics is the site of Jonathan Koomey, an independent energy researcher who writes about data center efficiency and the environmental effects of computing; the site also lists a roster of corporate clients, which is worth knowing going in.

This particular post is an announcement rather than the study itself: it introduces a new white paper on data center onsite water use, and the operator-by-operator numbers live in that linked paper, not on this page. What the post delivers directly is a clear definition of Water Usage Effectiveness (WUE), the metric used to compare cooling water intensity across facilities. WUE is expressed as water consumption or withdrawals in liters divided by computing electricity consumption in kilowatt-hours, and Koomey notes it can be reported on either a consumption basis or a withdrawal basis — a distinction that changes what a given number actually means.

The post also names its data source: company-reported fleet-average figures for the top twenty global data center operators, ranked by Data Centre Magazine in 2025, with numbered references backing the claims rather than bare assertions. That sourcing discipline is useful on its own, even before a reader follows the link to the underlying white paper.

Come to this page for the definition and the framing around WUE, not for a ranking table of operators — that is the honest description of what you will actually find here, and it suits readers who want the metric explained before they go looking for the numbers.

This arXiv preprint presents what its authors call the first published cradle-to-grave life-cycle assessment of greenhouse gas emissions for an AI accelerator, covering five generations of Tensor Processing Units. It has not been through peer review. The seven authors, including well-known hardware researchers David Patterson and Parthasarathy Ranganathan, work at Google and are assessing Google's own TPU hardware using first-party manufacturing data. That dual role is exactly what makes numbers like these available at all, since chip manufacturing emissions are rarely disclosed by the companies that build the hardware, and it is also a conflict of interest worth weighing alongside the results.

The paper tracks emissions from raw material extraction and manufacturing through disposal, deliberately leaving aside the energy and water a chip draws once it is running in a datacenter. It introduces a "compute carbon intensity" metric meant to let engineers compare hardware generations on manufacturing footprint alone, and reports that this metric improves threefold between the TPU v4i and TPU v6e generations.

The appeal here is methodological transparency: the authors describe their life-cycle assessment process in enough detail to work as a rough template for engineers wanting to run a similar study on other hardware. It is a narrow slice of AI's environmental footprint, manufacturing and disposal only, drawn from a single vendor's internal figures, so treat it as one detailed data point rather than an industry-wide benchmark.

JLARC — the Joint Legislative Audit and Review Commission — is the Virginia General Assembly's own oversight agency, and this page is the landing page for its study of the data center industry that has grown so heavily in the state. Lawmakers directed the review in 2023, so it answers to the legislature rather than to either side of the public argument, and it reports effects that cut both ways.

The page itself is framing, not findings. JLARC labels this an archived report and links out to five PDFs — a summary, the full report, recommendations, a presentation, and a consultant review of electric infrastructure and rate impacts — so the analysis sits behind those downloads, not in the page text.

The findings JLARC highlights run in both directions. The industry's footprint in Virginia comes out to roughly 74,000 jobs and a $9.1 billion contribution to state GDP. Against that, the study projects growing data center electricity demand could add $14 to $37 to a typical residential bill each month by 2040. The underlying work also covers siting decisions, effects on residents and natural resources, and the tax incentives the state offers the industry.

Treat the numbers as Virginia-specific. This is one state legislature auditing its own state's industry, grid and tax code, so the job counts, GDP figure and rate projections describe Virginia and won't transfer to another state or country.