Green IT: The Environmental Cost of the Digital Economy

Every time I open a search engine, stream a video, or run a query through a large language model, something physical is happening. Somewhere, a data center is drawing power, cooling systems are running, and servers are doing work. That physical process has an environmental cost. The cost is real. It is just invisible from where I sit, which is part of what makes it easy to ignore.

Data centers consume enormous amounts of electricity. Estimates of global data center energy consumption vary, and the numbers are genuinely hard to pin down because the industry is not fully transparent about consumption figures. What is clear, based on industry reports and academic estimates from recent years, is that the scale is significant, somewhere in the range of one to two percent of global electricity consumption, though that number is probably rising as AI workloads increase. Major cloud providers including Google, Microsoft, and Amazon have made prominent public commitments to achieve carbon neutrality or shift to renewable energy. These are real, publicly announced corporate commitments, widely reported in business and technology press. Whether they will be met, and on what timeline, is a harder question. I'm treating them here as what they are, stated intentions, not verified environmental outcomes.

The AI training question is where the energy conversation has sharpened recently. Training a large language model requires significant GPU compute over weeks or months. A paper by Strubell and colleagues published in 2019 estimated the carbon cost of training certain NLP models and drew attention to the energy intensity of large-scale model training. The specific numbers from that paper are hedged in various directions depending on the hardware, the energy mix of the data center, and what is counted, but the directional point is well-established: training large-scale models requires orders of magnitude more energy than running them for inference. Once a model is trained, each individual query is relatively cheap to run. The training cost is front-loaded and concentrated. The companies training the largest frontier models in recent years have, as far as I can tell from public reporting, not disclosed detailed energy consumption figures for specific training runs.

There is a carbon accounting issue here that I find genuinely underappreciated. When Google or Microsoft reports on its carbon footprint and its progress toward carbon neutrality, that reporting typically covers what are called Scope 1 and Scope 2 emissions in the GHG Protocol framework. Scope 1 is direct emissions from owned sources. Scope 2 is indirect emissions from purchased electricity. What is generally not included, or is much harder to account for, is Scope 3: the indirect emissions in the supply chain. For a data center operator, Scope 3 includes the emissions involved in manufacturing the servers, networking hardware, storage systems, and cooling infrastructure. Chip manufacturing is energy-intensive and materials-intensive. A modern data center requires thousands of servers with sophisticated components that took significant energy and resources to produce. If a company reports carbon-neutral operations but excludes the manufacturing footprint of its hardware, the reported number understates the total environmental cost.

E-waste is another piece of this that gets treated as a footnote. When servers are retired, which happens regularly because hardware generations turn over and more efficient equipment reduces operating costs, the resulting waste is a real environmental problem. Consumer electronics get more attention in the e-waste conversation, but enterprise hardware retired from large data centers represents significant volume. Proper disposal and recycling requires specialized facilities, not all hardware reaches them, and some components contain materials that are harmful if they end up in landfills.

The cryptocurrency side of this is more contested but worth noting. Proof-of-work cryptocurrencies like Bitcoin are designed so that mining requires computational work, and computational work requires energy. Estimates of Bitcoin's annual energy consumption have been compared, by various researchers and journalists, to the electricity consumption of medium-sized countries. The specific comparisons depend heavily on assumptions about the energy mix of mining operations and the hardware efficiency of the mining fleet. I am hedging these figures because they vary significantly across sources. The directional point, that proof-of-work mining at scale consumes energy in the same order of magnitude as national electricity grids, is widely discussed, though the precise figures are contested.

What I think the IS field hasn't fully absorbed is that information systems research is partly responsible for this situation. Every system that scales, every platform that grows its user base, every analytical pipeline that processes more data has an energy cost that grows with it. The field has been very good at studying adoption, use, and organizational impact of technology. It has been slower to study the environmental externalities of the systems it designs and promotes. Green IT as a research area exists, but it is not central to the field's mainstream concerns. The journals that publish IT value research don't usually ask about the environmental cost of producing that value.

I'm not arguing that the digital economy should stop growing. The question is whether it can grow with more attention to the environmental costs that are currently externalized. That requires transparency about energy consumption from data center operators, honest carbon accounting that includes hardware supply chains, and maybe some rethinking of which workloads are worth the energy they require. Training a frontier model to improve ad targeting is a different environmental calculus than training one to accelerate drug discovery.