Why does nobody know how much energy AI will consume?

The premise is that AI news has become energy news, and he opens with a day on which an independent power producer lost a fifth of its market value on news about a Chinese model. The forecasts that respectable shops publish for data center demand nonetheless diverge widely, and the divergence itself is shown in a chart this note cannot read, sourced in its caption to McKinsey, SemiAnalysis and EPRI. His explanation is structural rather than a complaint about anyone’s method. The answer depends on four variables that are each genuinely open and that interact: AI demand and profitability, algorithmic evolution, computing hardware efficiency, and power supply constraints.

On the demand side the two halves point in different directions. Models are measurably improving, and he finds the newer deep research tools markedly better than their predecessors, but adoption has not followed at the same rate, and the business of building frontier models looks financially precarious because none of the leaders has been able to pull ahead for long. He cites the investor Gavin Baker’s 2023 line about foundation models being the fastest depreciating assets in history, notes that the theoretical promise of general intelligence has kept investors hooked anyway, and asks whether they will stay hooked, answering: who knows. The algorithmic question is sharper. A 2020 OpenAI paper found that model performance improved smoothly with model size, dataset size and training compute, and for four years the industry scaled all three together, to the point that the relationship came to be treated as a law of nature rather than an observation. Around November 2024 reports surfaced that it was no longer holding, partly because the internet contains a finite amount of text but mainly through ordinary diminishing returns. The industry’s answer was to spend compute elsewhere, during inference, letting a model work through steps as it answers. That matters twice over: the most extreme demand forecasts assume training-style scaling stays the prime mover, and training and inference want physically different data centers. Training wants one enormous facility running for months, and because delay does not matter it can sit wherever power is cheap. Inference is spread across many smaller instances, some of which must cluster near population centers because they cannot tolerate delay, while others, such as research-style project work, could run whenever and wherever is convenient.

Efficiency is the variable with the most impressive record and, he argues, the least room left. Through the last decade’s cloud boom, computing demand soared while data center energy use barely moved. He quotes his own earlier writing to explain why that will not repeat: the gain came largely from moving workloads out of inefficient server closets into hyperscale facilities, that migration is essentially finished, and further efficiency inside those facilities has hit hard diminishing returns. Chips are still improving, but packing transistors closer together now generates enough heat to overwhelm the efficiency that density used to deliver, which is why the latest generation doubles the previous one only by switching to liquid cooling. He has come to see advanced cooling as an enabler of denser racks rather than an independent lever on consumption. To matter against order-of-magnitude growth a hardware lever has to be order-of-magnitude, and he sees only two candidates: chip architecture, which he says is not his area and which happens mostly behind closed doors at a handful of companies, and photonics, moving data between chips as light rather than electricity. The photonics turn is the one to keep. It is pursued to break a memory bottleneck rather than to save energy, so if it works it may mainly unlock a scale of AI that was otherwise unreachable, raising total consumption rather than lowering it. He names Jevons paradox for that outcome.

The fourth variable bounds the other three, because access to power is shaping up as the strictest constraint on data center development in most regions through at least the end of the decade. He walks the options in turn. Gas is the American default for firm capacity and its vendors are reportedly sold out for years, with manufacturers deliberately cautious about expanding after over-investing in the early 2000s. Batteries are deploying fast but are duration limited, so each further tranche needs more hours of storage attached and probably costs more. Renewables and storage he still sees great potential in, with unusually big error bars, because the easy locations are scarcer, costs have risen, cuts to longstanding tax credits in the then-current version of the Republican party’s bill would raise them further, and the supply chains run through China. Nuclear and geothermal he is increasingly bullish on, but he puts the timeline to multiple gigawatts at roughly ten years, which makes them contributors from the mid-2030s rather than now, and then only, he says, if investment in their long-term growth continues. Efficiency and load flexibility sit downstream of essentially all of the transmission and distribution bottlenecks, which is why he says utilities are already beginning to take them more seriously. Adding it up in what he labels extremely rough math, he reaches roughly a doubling of American dispatchable capacity over twenty years, then subtracts the share that must replace retiring coal and serve electric vehicles, heat pumps and ordinary economic growth, and lands on a hard cap of several hundred gigawatts for AI. Whether that is enough is precisely the thing nobody knows, himself included.