Intelligence is on its way to becoming a utility: plug into the outlet, get tokens. Where the power, the chips and the memory come from is covered every day. This piece looks the other way, downstream, at the appliances that will turn those tokens into economic output. It is a speculative thought exercise, not a forecast.

AI use : 10%.

Benjamin West’s painting above shows Benjamin Franklin pulling electricity out of the sky with a kite, the experiment of 1752. Ten years earlier, Franklin had designed a better wood stove. He had the electricity and he had the heater. It would be more than a century before the first electric heater warmed a room.

The current and the appliances

The working assumption here is simple: intelligence, measured in tokens, becomes as ubiquitous as electricity. You plug in and you get tokens, priced by the unit, at whatever volume you need. The price of a constant level of capability has been falling by roughly an order of magnitude a year, so “100x cheaper” is a matter of a couple of years, not decades. And as Jevons observed with coal, when a resource gets that much cheaper, we do not buy the same amount for less money. We use far more of it. So picture a company that receives a hundred times more intelligence, at a hundredth of today’s price.

The upstream part of the story (the power plants, the GPUs, the memory, the data centers and the debt that finances them) is covered daily by the press, and it is a race run with very deep pockets. It is not the subject here. Neither is the time it takes companies to reorganize around a new source of power; an earlier essay on the driveshaft covered that. This piece assumes the current is already in the walls, and asks what plugs into it.

Electricity on its own produces nothing. It becomes economic output through appliances: the motor, the lamp, the furnace, the refrigerator, the heater. Each one turns the same undifferentiated current into a specific result for a specific job. Three questions follow:

  • Which intellectual tasks, done today, would benefit from a hundred times more intelligence?
  • Which intellectual tasks are not done today, and would be worth doing?
  • What software turns tokens into those results, and is it a defensible company, or a wrapper that the next model will eat?

As this is my field, I’ll take the example of a factory.

Anatomy of a heater

A heater has three parts. A heating element turns the current into heat. A thermostat measures the room and switches the element on and off. A dial lets a person say what they want. Without the thermostat, you have a fire hazard. Without the dial, nobody can use it.

A token appliance has the same three parts:

  • The element: the model, or the agent, that turns tokens into a draft, a decision or a design. It is the part everyone demonstrates, and the part anyone can buy.
  • The thermostat: the loop that measures what actually happened after the decision (in the books, in the warehouse, on the machine) and feeds it back into the next one.
  • The dial: the place where people set goals and constraints, write instructions, review and override. Call it the hub.

Keep this anatomy in mind. It decides who can build these appliances, and who gets to keep them.

The heaters: today’s work, with 100x the intelligence

A heater does an old job, warming a room, with a new source of power. The first appliances of the token era will do the same: take the intellectual work a factory already does, and do it with a hundred times more intelligence.

The gain is not the same analysis, done faster. It is a different shape of work. Management today is a sampling problem: attention is scarce, so we sample. We review the top twenty suppliers, not the two thousand. We work the top of the scrap Pareto, not every deviation. We re-plan once a month and live with the plan in between. With a hundred times the intelligence, four things move:

  • From sample to census: every supplier, every item, every invoice line, every deviation gets looked at.
  • From periodic to continuous: the monthly review becomes a standing process that re-decides as soon as something changes.
  • From one option to thousands: each decision explores far more alternatives before one is chosen.
  • From days to minutes: the answer arrives while it can still change the outcome.

Applied to the intellectual work of a factory, it looks like this:

  • Finance, and the decisions around it. Today: a monthly close, an annual budget, a cash forecast in a spreadsheet, product costs updated once a year. With 100x: every transaction checked as it happens, a rolling cash forecast with scenarios, the true margin of each order known before saying yes.
  • Strategic access to capital: acquisitions, synergies, selling the company. Today: an event every few years, run with bankers, with a handful of targets screened and a data room assembled in a rush. With 100x: a standing watch on every possible target and buyer, synergies modelled on real operating data, a company ready for due diligence at any time.
  • Supplier relationships. Today: quarterly reviews with the top suppliers, the long tail managed through its invoices, contracts read once, at signature. With 100x: every supplier followed on health, delivery and price, every clause tracked, every renewal prepared, second sources ready.
  • Design of new parts. Today: a few variants iterated from experience; manufacturability checked late, by the few people who know the machines. With 100x: thousands of variants explored, each checked against the plant’s actual machines, materials and costs before anyone commits.
  • Design of new equipment. Today: long custom projects with integrators, few layouts compared, little reuse from one project to the next. With 100x: layouts, retrofits and tooling simulated and compared in depth, each project starting from everything learned on the previous ones.
  • Supply chain decisions, day to day. Today: a monthly S&OP, lead times and safety stocks set once a year, planners correcting the plan in Excel. With 100x: parameters recomputed item by item, the plan re-decided at every event (a breakdown, a late truck, a rush order) with the reasoning written down.
  • Use of resources: material, energy, people. Today: yield and scrap followed on a few indicators, staffing planned weekly on a whiteboard, skills tracked in a spreadsheet. With 100x: material use tuned batch by batch, energy-hungry steps moved to cheap hours, staffing matched to skills, training and preferences, person by person.
  • Logistics orchestration. Today: transport booked by phone and email, fixed routes, dock slots managed day by day. With 100x: every load tendered and consolidated, routes adjusted in real time, customs and claims paperwork done for every shipment.
  • Quality and compliance. Today: deviations analysed for the biggest issues only; audit and regulatory files prepared by hand, against a deadline. With 100x: every deviation traced to a probable cause from its batch history; audit and regulatory files ready at all times.

A clarification, because it matters for what follows. In several of these tasks, the core is already mathematics: a scheduling problem, a lot-sizing problem, a routing problem. Tokens do not replace the solver. They replace the human glue around it: gathering the context, formulating the problem, cleaning the data, handling the exceptions, explaining the answer to the person who signs. The most reliable pattern we have found so far is to have the model write deterministic code, and the code run the operation.

The air conditioners: work nobody does today

The more interesting appliances do a job nobody was doing. In 1902, Willis Carrier designed a system to control the humidity of a Brooklyn printing plant, because damp paper stretched and the colors no longer lined up. It was not a better fan. It set the moisture in the air so that a process would work, which was a new job altogether. Climate control went on to make possible industries that cannot run without it, from semiconductor fabs to the data centers that now produce the tokens.

Why is a useful task not done? Because the attention it needs costs more than the value it brings. Divide the cost of attention by a hundred, and a whole class of tasks crosses the line. For the mathematically inclined: if the value of the tasks a company could take on is spread like a power law, with N(v) ∝ v−α of them worth more than v, the number worth doing at a cost c is N(c) ∝ c−α. Divide c by 100 and that number is multiplied by 100α: tenfold for α = 0.5, a hundredfold for α = 1. Each new task is small. There are a great many of them. In a factory, they fall into four families.

  • The long tail. Everything done today for the top of the Pareto, done for the rest: a root cause for every micro-stoppage, not only the big breakdowns; value engineering on every part in the catalogue, not only the top fifty; an answer to every request for quotation, instead of declining the ones unlikely to convert. Lean asks for every problem to be solved at its root. Most factories can only afford it for the biggest ones.
  • Reading and recovering. The money left on the table because the paperwork costs more than the recovery: late-delivery penalties never claimed, freight overcharges never disputed, warranty claims never filed against suppliers, grants and tax credits never applied for, contract clauses nobody checks. A whole industry, recovery audit, already lives on a share of this money at the largest companies. With reading a hundred times cheaper, every company can afford it, on every line.
  • Counterfactuals and learning. Writing down why each decision was taken, and checking later what would have happened otherwise. A review of every lost deal, every late order, every bad batch. A stress test of the plan against a supplier failure, a new tariff or a doubling of energy prices, every week rather than once a year. Organizations rarely learn from their own decisions, not for lack of will, but because nobody has the time to write them down and come back to them.
  • What is in people’s heads. In our deployments, roughly 40% of the data an agent needs exists in no system at all. It sits with the planner who knows which customer always changes its order on Thursdays, or with the setter who knows which press drifts when the shop is cold, and some of those people are close to retirement. Capturing it, by watching the decisions people override and asking them why, is a task almost nobody does systematically, because it takes time from the people who have the least of it. It may be the best use of tokens of all, because every other appliance runs on that context.

Who makes the appliances?

Now the investor’s question. Each of these tasks calls for a dedicated hub where people and models work together: the data for that task is curated there, the instructions and decision rules are written down there, and decisions are proposed, reviewed and overridden there. A finance hub, a supplier hub, a design hub, a planning hub. Is that a company, or a wrapper the next model will eat?

The thesis popular in the labs is that the model is the product, and in the end the company: each generation absorbs the software built around the previous one. There is truth in it. Much of what was sold in 2023 as an AI product (a chat over a PDF, a writing assistant, meeting summaries) is now a feature of the labs’ own assistants or of the tools people already use.

Electricity offers a counterpoint, and a warning. The counterpoint: the utilities sold kilowatt-hours, a commodity with regulated margins, while the appliance makers sold heat, cold, light and motion. The appliances lasted because they carried knowledge the grid did not have: the physics of each use, the safety, the installation, the service. And the largest gains went to neither, but to the companies that reorganized their work around cheap power, Ford’s assembly line being the textbook case. The warning: General Electric made both the generators and the appliances, and became one of the defining companies of the electrical age. If a lab becomes the General Electric of intelligence, the appliance makers are its competitors, not its customers.

So, is the hub defensible? At scale, probably. Below scale, a hub is a prompt, an interface and a few integrations, and the model, or the customer’s own IT team with a good agent, can rebuild it in weeks. At scale, it accumulates what a model cannot download: the data structure that describes the domain, the decision rules of hundreds of companies, the evaluations that say what a good answer is, the integrations, and the habits of the people who use it every day. Scale makes the dial sticky. What makes it durable is being wired to one of two deeper moats.

The thermostat: proprietary data and physical loops. A lab can retrain its model. It cannot train on data it never sees. The deepest moat is a loop the model cannot run without you. The model makes a guess with the current context. The guess is executed on a machine, a line, a truck. The physical world answers: the part is in tolerance or it is not, the truck is on time or it is not. And that answer improves the context for the next guess. Whoever owns the machine owns the loop. In software, the compiler closed this loop, which is why coding raced ahead. In a factory, the machine is the compiler. An appliance that sits on that loop, with the customer’s consent, gets better with every cycle, and none of it can be downloaded.

One subtlety. In “whoever owns the machine,” the owner is usually the factory, not the software vendor. The physical loop is first a moat for the manufacturer. It becomes a moat for a software company when that company aggregates loops across many plants, so that what one line learns improves the appliance on every other line. That aggregation is where an appliance maker earns its place.

The network. Some appliances improve with every participant. A supplier hub where suppliers connect, quote and update their capacity becomes more useful to every buyer with each supplier who joins, and a competitor has to bring them all over at once. It is one reason the marketplace might be the most interesting bet of this era. The same holds inside a company: when every planner, buyer and supervisor works in the same hub, their instructions and corrections accumulate there, and that is precisely the context the model lacks.

What is not a moat: the prompt, the interface on its own, a fine-tuned model on its own (the next base model usually overtakes it, the bitter lesson again), and being first. That gives a simple test for any appliance: if the next model were ten times better tomorrow, would the company be worth more or less? If less, it is a wrapper: the model was doing the work. If more, it is an appliance: a better current makes a good heater more valuable, because the thermostat, the dial and the network were the scarce parts all along.

What would make this wrong

  • The model builds the hub. If agents can generate a bespoke hub for each company on demand, hubs become disposable software. Only the loops and the networks remain, and many companies that look like appliances today turn out to be wrappers.
  • The labs go downstream. They are already funding deployment companies with private equity money. If they run the deployments, they can own the loops, and the General Electric scenario plays out.
  • The factory keeps the loop. Manufacturers may decide the loop is theirs, build their own hubs with the labs’ tools and their own people, and treat software vendors as contractors. The winners of the token era would then be the factories that reorganize around it, which is also, in the end, what happened with electricity.

The scarce part

So, what is the heater? In a factory, it is the finance hub, the supplier hub, the design hub, the planning hub: today’s work, done with a hundred times the intelligence. The air conditioners, which take on the work nobody does yet, may matter even more. Both are built the same way: an element anyone can buy, a thermostat few can own, and a dial people actually use. The current will be everywhere, and it will keep getting cheaper. What will be scarce, as with electricity, are the appliances, and the redesign of the work around them.

Interesting times to build.

How to read this piece. A speculative thought exercise, not a forecast. The author builds and invests in software companies for industrial operations of the kind discussed here, so read the views on defensibility with that in mind. Historical examples are given as commonly told; check them against primary sources before quoting them.