The phrase green computing sounds virtuous and means less than it appears, and the gap between the two is becoming harder to ignore. The computing industry has spent a decade making its machines more efficient, and the machines are, by almost any measure, dramatically more efficient than they were. The emissions of the industry have not fallen in proportion. They have, by several credible accounts, risen, and the discrepancy between the efficiency that was achieved and the emissions that followed is the central problem the field has not yet confronted. Efficiency is not the same as cleanliness, and confusing the two has cost the industry a decade of honest accounting.

The mechanism that defeats the efficiency is old enough to have a name: when a thing becomes cheaper to use, people use more of it. Computing that grows more efficient per unit of work tends to be used for more units of work, and the growth in demand has, for years, outpaced the gains in efficiency. The data center that draws half the power per computation draws twice as many computations, and the climate does not care which half of the equation improved. The industry has been winning the battle it chose to measure and losing the one that matters, and the difference is measured in carbon.

The accounting that hid the problem

Part of the reason the discrepancy went unaddressed for so long is that the accounting made it easy to hide. The companies that report their emissions have, for years, reported the emissions they produce directly — the electricity they buy, the fuel they burn — and described those as their footprint. The emissions produced on their behalf, by the manufacturers of their hardware and the construction of their facilities, were categorized elsewhere, and the result was a footprint that looked smaller than it was. The honest accounting, which attributes the full lifecycle of the machines to the industry that uses them, paints a picture considerably less flattering, and it is the one the climate actually experiences.

The shift to renewable electricity was the part the industry got right, and it deserves credit for it. Several of the largest operators now run their facilities, in aggregate, on power they procure from renewable sources, and the emissions that would have followed their electricity have, accordingly, not been produced. But the renewable power is a solution to one part of the problem — the running of the machines — and the manufacturing of the machines, and their disposal, and the networks that connect them, remain stubbornly fossil. The clean data center that runs on dirty hardware is cleaner than it was, and it is not clean.

The rebound that efficiency cannot escape

The deeper problem is the one efficiency cannot solve, because efficiency is the cause of the growth that defeats it. The industry that delivers more computation per watt will, in a market that rewards computation, be asked to deliver more computation, and the question of whether the total emissions rise or fall depends on whether the demand grows faster than the efficiency improves. For most of the last decade, the demand has won, and the artificial intelligence boom has accelerated the trend rather than reversed it. The models being trained are larger, the data centers being built are more numerous, and the efficiency gains, however real, are being outpaced by the scale of what the industry now considers a reasonable amount of computing to do.

The honest path forward is not the one the industry has been telling. It is not to make computing more efficient, though that helps at the margin. It is to confront the question of how much computing should be done, and on whose behalf, and at what cost to a climate that does not distinguish between the computation that was necessary and the computation that was merely possible. That is a question the market will not ask, because the market rewards more. It is a question the regulators are beginning to ask, slowly, and the answer the industry gives will determine whether green computing becomes a description of what it does or a slogan that describes what it would prefer to be remembered for.

The hard math that remains

The technology to compute cleanly, in the sense the climate requires, exists in parts and not yet in whole. The renewable electricity is procured at scale. The efficient hardware is built. The honest accounting is, grudgingly, beginning to be adopted. What is missing is the willingness to treat the growth in demand as a variable to be questioned rather than a fact to be accommodated, and that willingness is the part the industry has never shown. The green computing that the climate needs is not the one the industry is building. It is the one it has not yet decided to build, and the decision, if it comes, will be harder than any efficiency gain the engineers have delivered.