Explaining the ‘Watt-Bit Spread’

The framework is borrowed from the spark spread, which power traders use to compare the cost of a unit of gas against the value of the electricity that unit can produce. Janous applies the same shape one conversion downstream. The data center industry exists to turn watts into bits, so the relevant spread is between the cost of the electricity going in and the value of the compute coming out. He thinks no energy conversion in the economy returns more. The value side has risen steeply over the previous 18 to 24 months, and not only because compute earns revenue: securing electrons is itself a moat, since whoever holds the power can train larger models and run more inference. The cost side has barely moved, because a standard utility tariff does not reprice when demand surges. The practical consequence of the gap is a price signal that never arrives. If utilities and independent power producers cannot see what the next watt is actually worth, they build less infrastructure and fewer GPUs get plugged in, and the same muted signal fails to reach the people making transformers and switchgear.

To use the framework you need three inputs, and the third is the one that gets lost in summaries. The first is what a watt is worth to a compute buyer, which is high. The second is what the buyer currently pays, which is the prevailing tariff. The third is the year of delivery. Janous is explicit that the spread is about the value of capacity in a given year: a megawatt in 2027 is worth more than a megawatt in 2032, because by 2032 he expects power to be relatively more abundant and the scarcity premium to have decayed. So the spread is not one number but a term structure that narrows as you look further out, and the thing being mispriced is not energy but the timing of the first electron. That reframing is what makes the framework operational: a utility should be hunting for investments that pull delivery forward, and for customers who will pay more for 2027 than for 2032. The premium shows up as higher demand charges recovering the cost of acceleration rather than as a permanently higher price per megawatt-hour, and it is underwritten by the fact that data centers run 8,760 hours a year and do not get switched off, so the asset stays used and useful.

The reason the gap persists is that the two sides of the table are running different playbooks. Downstream players, which is to say cloud providers, chip makers and model developers, think in terms of the theory of constraints, where enterprise value comes from throughput and you therefore overbuild at the bottleneck and pay a premium to clear it. That is rational for them because base infrastructure is cheap relative to the whole stack. Land is about 1% of the fifteen-year total cost of ownership, and the full build including GPUs runs about $25 million per megawatt, so the roughly $1 billion of utility work behind a gigawatt sits in front of some $24 billion that follows it. A utility faces the opposite arithmetic: it sells an electron in 2027 for the same price as one in 2030, which gives it no reason to stockpile 345 kilovolt transformers or to compress a schedule, and a lead time stretching from two years to seven is absorbable inside that model. Janous says that is not ultimately in the utility’s interest, since it should want more customers sooner, but its business model was not designed for this. Kann pushes back that the framing understates the constraints utilities live under, including regulators with their own priorities, a mandate to put reliability first, and a rate-of-return structure that governs what they can spend and recover. Janous’s answer is the instrument he calls an advanced grid tariff, named by Katie Fehrenbacher: the same logic as a green tariff, where a customer who wants a particular quality of service pays for it without harming other ratepayers, except that here the thing being bought is speed. A utility identifies technologies that can be deployed quickly, grid enhancing technologies, dynamic line rating, storage used as transmission, long duration batteries, and recovers their cost from the specific customers who want power in 2027 instead of 2030.

Where it breaks down is on the durability of the spread and on who is good for the money. Kann’s sharpest challenge is that the spread may be an artifact of the cycle, inflated by an arms race and a lot of capital sloshing around, and that an infrastructure owner is being asked to commit to assets lasting 20 to 40 years against willingness to pay that is only visible to about 2030. Janous concedes the possibility and then states a belief rather than evidence: he expects demand for compute to exceed available power through 2030, so the marginal gigawatt stays valuable until equilibrium returns. The second weakness is buyer quality. Utilities cannot easily tell a hyperscaler from what Janous calls two guys with a truck, when a queue position can cost as little as $10,000 and a single utility can be sitting on tens of gigawatts of requests. His filter is capital: anyone holding a gigawatt of queue position should be able to point to billions of dollars they have access to, and if they cannot they are not a serious player. The third constraint is political rather than analytical. Both speakers say directly that they do not want residential rates rising to serve AI, which means the whole design depends on cost allocation being done properly, with benefits flowing to other ratepayers through better reliability or lower cost. Kann makes that condition explicit and immediately adds that he does not know how far it can go.