A skeptic’s take on AI electricity load growth

Every projection of data center electricity use, stated or not, multiplies two things: how much service people demand, and how much electricity it takes to deliver a unit of it. Koomey’s history is that the second has repeatedly swallowed the first. Data center electricity did double from 2000 to 2005, taking the sector from roughly half a percent of US electricity to about 1% by 2010. Then from 2010 to 2018 compute output rose roughly sixfold while electricity use rose 6% in total, partly from shrinking transistors and new hardware architectures, and substantially from workloads migrating out of inefficient corporate server rooms into hyperscale facilities with better cooling, better power delivery and much higher equipment utilization.

Kann presses the obvious objection: the migration to hyperscale was a one-time move that has already happened, so that lever cannot be pulled twice. Koomey’s answer is that it was never the only lever. Beyond transistor shrink there is hardware architecture, more efficient software, better algorithms, which he notes are domain-specific but often worth orders of magnitude, and special-purpose silicon co-designed with the software it runs. DeepSeek, which he describes as hitting similar accuracy benchmarks on much less compute and at much lower cost, is his live example that the software side is not exhausted. He puts the remaining headroom at orders of magnitude, as a possibility rather than a forecast. What makes him expect it to be used is not technology but constraint: from 2000 to 2005 the industry built as fast as it could until the power bill became a problem and then went after efficiency, and he expects that pattern to repeat, with efficiency neglected so far only because the constraints had not yet bitten. The incentive is money, since energy and capital are both large shares of what a data center costs.

On demand he makes a narrower point than the episode’s framing suggests. The argument he is attacking is not really Jevons paradox, and he says tying it to efficiency is a mistake, since measured rebound effects run about 10% to 20%, far too small to carry the claim. What the AI suppliers actually assert is that demand for their product is effectively infinite, so cheaper compute just means more compute and the electricity gets used regardless. Koomey does not say that is false. He says it is an assertion that should be open to discussion rather than assumed, and gives three reasons to hold it loosely. Accuracy gains have historically come from scaling, at something like 10% to 30% per tenfold increase in compute, and each further tenfold is a far harder construction job than when this was a small industry. Hallucinations may or may not be reducible enough that companies accept the liability for what their agents do, which he says nobody knows. And the revenue to justify the announced capital expenditure has not appeared. Kann grants the revenue point and pushes back on the hallucination one, that the models keep getting better at removing false answers. Neither claims to know where demand lands, and both say so.

That is the actual thesis. After thirty years of energy forecasting, Koomey holds that nobody can predict computing more than a few years out, that the two or three year pipeline of announced projects is the limit of anyone’s visibility, and that a stated figure for 2035 is made up. It is a claim about confidence, not about direction, and when Kann pins him to a number the direction turns out to be up. He allows data centers going from about 4.5% of US electricity to roughly 9% or 10% by 2030, which he works out to about one percentage point of total US demand growth per year, and it is that one point he calls manageable rather than explosive. Electrification of vehicles, industry and heat then adds one or two points on top, for perhaps 3% annual growth against the 0.5% of the past twenty years, which he calls the reasonable range. He does not apply the word manageable to the 3%, and by his own arithmetic it is a sixfold acceleration on the last two decades. His evidence for the softer adjective is that the bottom-up estimates and the aggregate statistics disagree: US generation in 2023 came in below 2022 and roughly level with 2018, so the doubling of the data center share shows up as a reshuffling of a flat pie rather than as visible national growth. He offers that as a tension to be ironed out, not a finding.