Steel For Fuel N° 029 of 56 20 Nov 2024
For AI, energy is nothing, and energy is everything
by Andy Lubershane, Partner and Head of Research, Energy Impact Partners
In this note
The question
If energy is only a small part of what computing costs, why is the AI industry suddenly worried about power, and what should the power sector do with that?
The answer
Energy is a trivial share of the cost of computing and at the same time a serious constraint on it, and Lubershane reads that combination as an opportunity: AI developers who expect to be very profitable ought to be willing to pay a large premium for power, especially low-carbon power, and the power sector should enlist them to help build it. He offers that as a possibility, not a forecast.
03The argument
The post sets two facts against each other. Data centers are the fastest growing source of electricity demand worldwide, and on the forecast he reproduces, global capacity and consumption are both on track to roughly triple within four years. Yet energy stays a small line in the cost of computing. Data center rent in top American markets far exceeds the cost of power generation capacity; that rent excludes the chips, which can cost more than everything else combined; and an outside estimate of frontier model training puts energy at a modest fraction of the chip bill, roughly the cost of breakroom smoothies. Hence the title: energy is nothing as a share of cost and everything as a constraint, because the power sector is running what he had called earlier that year a gauntlet, meaning surging demand against bottlenecks on building infrastructure.
The opportunity falls out of that gap. If power is cheap relative to computing and leading developers anticipate extremely profitable products, they ought to be willing to pay a substantial premium for it, and because those companies are staffed with people who want to do something about climate change, he judges they appear willing to pay an especially high premium for low-carbon power. The hyperscalers’ nuclear deals are his evidence this is under way. What he wants is more than selling them electricity: enlist those investors to advance emerging technology categories by paying a premium and possibly taking a share of the risk on first projects, which he is encouraged to see utilities beginning to do programmatically.
04What you need to know first
- Data center rent
- What a tenant pays for capacity: space, fiber connectivity, cooling, security, backup power and a grid connection, and explicitly not the computing hardware inside.
- Offtake agreement
- A long-term contract to buy a plant’s output, which is what lets a generator be financed and built.
05Details worth keeping
- The forecast is SemiAnalysis’s from March 2024, which he singles out as rigorous against a field he calls pretty iffy. It covers every class of facility, hyperscale and colocation, cloud and AI, not AI alone.
- His cost comparison of data center rent against generation capacity is made only in a chart; the text gives the direction and no numbers.
- The programmatic route he points to is Duke Energy’s proposed “Accelerating Clean Energy” tariff framework, which he says all three named hyperscalers support.
06Claims worth citing
All figures as stated on 2024-11-20; cost shares and deployment dates move quickly.
- Data center capacity is on track to grow more than threefold within four years, with consumption roughly tripling to about 4.5% of global electricity demand. SemiAnalysis, March 2024, cited by Lubershane
- Energy appears to be about 10-20% of the cost of the Nvidia chips in training frontier models. (Lubershane’s reading of an Epoch AI cost breakdown from June 2024, which sits in a chart the note cannot see)
- Microsoft has a 20-year offtake agreement with Constellation to restart Three Mile Island Unit 1 by 2028. Lubershane
- Google has an agreement with Kairos to deploy 500 MW by 2035. Lubershane
- Amazon has an agreement with Energy Northwest to deploy X-energy small modular reactors, and invested in X-energy’s $500m fundraise. Lubershane
- On 18 November, Duke Energy’s market capitalization was $87 billion against Google’s $2.15 trillion. Lubershane
07Where it’s contested
Nothing is contested; there is no second voice. The confidence gradient is the useful part, and it sits on the load-bearing step.
- Stated as possible, not likely. He calls the demand signal and the policy attention undoubtedly a source of pressure, then says only that it is possible these forces can also be harnessed as accelerants.
- The premium is asserted. The argument needs developers to actually pay one; the support is three announced deals plus a claim about what the staff of those companies want.
- The cheap-energy premise is present tense, and the post does not ask what becomes of the argument if energy’s share of computing cost rises.
- The ask is sized, not designed. The case for a hyperscaler shouldering part of the burden is the market capitalization gap, which a footnote dates to two days before the post.