AI as the Next Utility

There is a moment in the life of every foundational technology when it stops being a product and starts being a substrate. Electricity made that transition between roughly 1880 and 1920. It began as something a wealthy factory owner generated on-site for their own machines, with a private dynamo, a private boiler, and a private engineer who knew the eccentricities of that particular installation. It ended as something you bought from a wall, indistinguishable across providers, priced by the kilowatt-hour, governed by tariffs and reliability standards, and traded on markets that the people consuming it never had to think about.

The interesting part is not that this happened. The interesting part is how predictable it was — and how thoroughly nobody saw it coming until it was already underway. The first utilities did not call themselves utilities. The first power markets did not call themselves markets. The vocabulary arrived after the infrastructure, not before, and it arrived borrowed: language from gas distribution, from telegraphy, from finance.

This publication exists because I think we are at the equivalent moment for artificial intelligence, and the vocabulary has not yet arrived.

The argument in one paragraph

AI inference and training compute have all the structural properties that, historically, push a technology toward commodification: they are fungible-enough at a given quality tier; demand is bursty and badly correlated with local generation capacity; supply is capital-intensive, geographically constrained by power and cooling, and slow to build; and the marginal product (a token, a FLOP-second, a forward pass) is quantifiable to many decimal places. Whenever those properties co-occur, the historical pattern is the same: vertically-integrated providers give way to a layered market — generators, transmitters, retailers, traders — and the unit of trade gets standardised, graded, and abstracted from the specific machine that produced it. The interesting question is not whether this happens to AI, but what shape it takes.

Why the cloud is not the answer

The natural objection is that we already have this. The hyperscalers sell GPU-hours by the second; spot markets exist; brokers exist; resellers exist. Surely we are already in the commodity era?

I would argue we are in the pre-commodity era — roughly where electricity was when Edison was running DC mains down a single street in lower Manhattan. The cloud sells you access to a specific provider’s specific hardware in a specific region, with a specific software stack, billed in the provider’s units, with switching costs that any honest CTO will tell you are substantial. None of that is a commodity. A commodity is what you get when the unit of trade has been so thoroughly standardised that the identity of the seller is irrelevant to the price.

The electricity analogue is instructive. Long after every major city had electric service, you still bought it from your local generator. The market for electricity — the thing that made power a commodity in the modern sense — required: a standardised unit (the kWh), a transmission layer that could move it between regions, a settlement mechanism that could clear trades faster than physical delivery, and a regulatory regime that made the standardised unit enforceable. Each of those took decades. Each of them was politically contested. None of them existed when people first started selling electricity by the wall socket.

We do not yet have any of those for AI. We have rough proxies — $/M-tok prices on aggregator dashboards, GPU spot indices, occasional contractual standards inside enterprise procurement — but no unit-of-trade with the regulatory and technical scaffolding that would let it function as money for compute. That is precisely the layer this publication is interested in.

Five framings worth taking seriously

Different futures involve different units. The site is organised around five framings of “AI as commodity,” because I think it matters which one turns out to be the right shape.

Compute-backed inference tokens. The unit is “one token of inference output, at quality grade X, deliverable within a settlement window Y.” This is the most direct analogue to electricity-by-the-kWh. It assumes inference demand is the dominant load and that quality tiers are coherent enough to anchor a contract. SF Compute and a handful of inference brokers are already feeling their way toward this.

Capacity futures. The unit is “the right to N GPU-hours of class-Z hardware, in calendar quarter Q.” This is closer to the way natural gas or aluminium trade than to electricity, and it aligns with how hyperscalers actually plan capex. It implies a forward curve, hedging products, and eventually exchanges. It is the most boring future and also the most likely to arrive first. (See: capacity future.)

Training-FLOP commodities. The unit is the FLOP-hour of training compute at a specified precision and interconnect grade. This is the “raw material” framing — useful precisely because training and inference have very different supply-demand dynamics, and pretending they are the same will obscure both.

Crypto-native AI tokens. The unit is a chain-issued token redeemable for compute, where settlement, identity, and metering are all on a ledger. Bittensor is the obvious incumbent direction; whether this framing dominates depends largely on whether the regulatory regime for the previous three framings is built inside or outside the existing financial system.

Output and data rights. The unit is not compute at all but the right to use specified model output (see output rights), or the right to have one’s data excluded from training corpora (see data exclusion). This is the framing closest to how intellectual property already trades, and it is the one most likely to get to a working market first because it sits on familiar legal infrastructure.

These framings overlap. They are not mutually exclusive. They are also not exhaustive — a sixth or seventh will probably matter in ten years that nobody is naming yet. The point of laying them out side by side is to make it easier to ask: which assumptions does each framing rest on, and which assumptions are currently doing the most invisible work?

What this publication is, and is not

This is a research and speculative-design project. It is not a prediction service, a trading desk, a policy advocacy outlet, or a breathless newsletter about the AI Bubble. It is a place to do three things slowly and in public:

1. Frame. Borrow vocabulary from the closest historical analogues — primarily power markets, secondarily commodities exchanges and regulated utilities — and apply it carefully to AI infrastructure.
2. Speculate. Design fictional artifacts: contract specs, grade sheets, ticker mockups, rule books for exchanges that do not exist yet. The point of designing the artifact is to surface the assumptions hiding in the words.
3. Annotate. Track real-world signals — pricing dashboards, capex announcements, regulatory filings, academic work — and note what each one implies about which framing is gaining ground.

It will be wrong about a lot. The value, if there is any, is not in being right early — it is in being legible early, so that when the infrastructure does take a recognisable shape, there is already a vocabulary on the shelf for talking about it.

The next dispatches will start populating the five framings in turn, and laying out the first speculative artifacts. If any of this resonates, stay close.

— Utilitus