Data Center Return on Investment

The cheaper path: agents

The panels after this one are a reading. This one is an argument, made by the author of this site, and it is set out with the points against it so you can weigh it.

What an agent is. An AI model answers one thing at a time: a question, a paragraph, a single fix. It does not remember the larger goal or check its own work across a long job. An agent is the part that does. It holds the goal, keeps track of what has been done, checks each result, and decides what to do next, over and over, until the job is finished. The agent needs no training. It is just the loop wrapped around the model, so it is built once and cheaply, not trained at enormous cost like the model itself. The model makes each step good. The agent keeps all the steps pointed in the same direction.

Working in a loop like this, where each pass starts from what the last one produced, is called recursion. It is why an agent can finish a job far too large for a model to do in one answer: it breaks the job into pieces the model handles well and keeps them adding up to one result. This page was written that way: drafted, checked against its sources, corrected and checked again.

What each one costs. Building a new frontier model is the expensive part. Epoch AI finds the cost of the largest training runs has grown about 2.4 times a year since 2016, and projects runs costing more than a billion dollars by 2027. Putting an existing model to work is a different order of cost. The author’s own agent, Rasputin, runs on a commercial model for hundreds of dollars a month.

The argument. If loops built around today’s models deliver most of the value people actually get, then each new, larger training run buys less than it costs, and the case for building the power grid around ever-bigger training runs gets weaker. The better investment would be in using the models we have well.

The points against it.

Each step is still the model’s. The agent can carry a task far beyond what the model could finish alone, but every individual step is only as good as the model taking it. A habit the model has shows up in every pass unless the agent is built to check for it. The author’s agent repeatedly turned carefully hedged statements into flat absolutes until a person caught it, because nothing in its loop was looking for that error. Stop training, and agents keep working and keep improving as their loops improve, on steps no better than the last model built.

Cheap per agent is not cheap in total. An agent makes many passes to finish one task, so it uses far more computing than a single question. Running models for users is already a large and growing share of the industry’s computing. A world where billions of people use agents all day still needs data centers. It would need them for running models, not only for training them.

Nobody stops first. Even if every AI lab agreed agents are the better path, the first company or country to stop training risks falling behind the ones that keep going. That is why the spending continues, and it brings the question back to who pays for the race.

How to use the loop yourself. You do not need to build an agent to work this way. You need the habit. Many people ask an AI one question, take the answer and stop. The loop means never stopping at the first answer. Here is what that looks like on something everyone has, an electric bill:

The first answer is only a starting point. Every question after it comes out of the answer before it, and you stop when you understand the bill, not when the AI stops talking. That is the loop, and it works on a lease, a medical bill or a loan the same way.

Where it lands. Agents do not make the data centers unnecessary. They change what the data centers are for. The return-per-watt chart further down is the test of this argument: if the gain from each new doubling of power keeps shrinking, the argument gets stronger.

The numbers

The first line and the third are the subject of this page. The cloud companies’ spending is roughly twenty times the revenue of the companies that sell only AI. That comparison is rough: the cloud companies also earn AI revenue of their own, and not all of their spending is for AI. It still shows the size of the gap. Investors and lenders are financing it, and households cover part of the electricity build through their power bills.

The basin

Think of a ball resting in the bottom of a bowl. Give it a nudge and it rolls right back to the bottom. It takes a real shove, hard enough to knock it over the rim, to send it somewhere else. A lot of money behaves the same way: give it a small scare and it settles right back where it was. The bottom of the bowl is where it keeps wanting to return.

Here, the bottom of the bowl is a habit. Whenever AI gets more capable, investors treat that as the reason to pour money into the next round of computing power. As long as that habit holds, a bad quarter or a falling stock price is shrugged off and the money flows right back in. It is helped along by ordinary investor instincts: piling into whatever is the hot story, companies spending on data centers to look like they are keeping up, and everyone chasing the same bet at the same time because everyone else is.

All of it rests on one belief: that every extra dollar spent on computing power and electricity buys AI that someone will actually pay for. As long as investors believe that, the money keeps coming back to the bottom of the bowl. If enough of them stop believing it, the money rolls out and goes looking for somewhere else to be. This page does not try to guess when that might happen.

Who holds it up

No single company is running this. The labs need the cloud companies’ computers, the cloud companies need the chip makers and the power companies, and all of them need investors to keep believing the labs will produce something worth paying for. Everyone in the chain is depending on everyone else.

It cannot make what it needs to survive. The cycle cannot produce its own chips, generate its own electricity, or approve its own power lines. It buys all of that from the outside. Being enormous does not change that. A very big buyer is still just a buyer.

Growing fast is not the same as paying off. Plenty of things grow fast for as long as someone else is footing the bill, and stop the moment the money does.

Where this reading could be wrong. Every business needs outside money, so needing it does not by itself mean the cycle is being propped up. The real question is who is making the calls that keep it going. Right now that is mostly the investors and boards inside the cycle, who can keep choosing to fund it. That changes the day the decision passes to people who can say no from the outside, lenders who stop lending, regulators who set the terms, or ratepayers who refuse to keep paying. Until then, the cycle is steering itself.

Who pays

The answer depends on where you live. A 2026 study using causal methods finds that from 2015 to 2024, data centers lowered average residential electricity rates slightly. For every 10% increase in data center capacity in a state, residential rates fell about 0.4%. A large new customer spreads the grid’s existing fixed costs over more electricity sold, and everyone else’s share of those costs gets smaller.

That only works while the grid has room. Once new power plants and new lines have to be built for the demand, the cost of building them lands in rates. The regional grid serving thirteen states and Washington, DC, from New Jersey to Illinois, reached that point first. The price it pays power plants to be available rose from $28.92 to $269.92 per megawatt-day in one year, and to $329.17 the year after. One analysis attributes 63% of the first increase to data centers, about $9.3 billion recovered from customers.

Many households have no room to absorb that. As of February 2026, about 21.5 million US households, one in six, were behind on their energy bills. The average monthly residential bill rose from about $121 in 2021 to $156 in 2025. Low-income families spend about 8.6% of their income on energy, nearly three times the 3.0% other families spend. Thirty-three states have no rule against shutting off power in summer heat, and about 80% of federal energy assistance goes to heating rather than cooling. Data centers are one cause of rising bills among several: fuel prices, storm and wildfire costs and new transmission lines all push in the same direction.

The larger risk is still ahead. The same study that found falling rates warns that if the expected data center demand never arrives, utilities will have built plants and lines nobody needs, and the effect reverses. This is where the basin above reaches the household. If the belief at its rim fails, the companies can slow their spending in a quarter. A power plant or transmission line already approved stays in the utility’s rates and is paid off over decades by the utility’s customers, households included.

Who carries that risk is being decided now. As of March 2026, 77 special electric rates for very large customers had been approved or proposed across 60 utilities in 36 states. In May 2026 Pennsylvania’s utility commission advanced a proposed model, guidance that each utility still has to turn into its own filing, with terms like these:

Where terms like these are written and enforced, the cost of guessing wrong about AI falls on the companies making the bet. Where they are weak or missing, it falls on everyone else on the bill. Whether the buildout is worth that risk is a judgment for voters and regulators. This page only shows where the risk sits.

Return per watt

The whole buildout rests on one bet: that each new surge of computing power buys enough extra AI capability to be worth it. That bet can be checked against public data, and the check comes down to a single question. Every time the power that goes into training doubles, how much better does the AI actually get? As long as each doubling still delivers a real jump in capability, the bet is holding. Once the jumps start shrinking, each new power plant is buying less and less, and the bet is weakening.

The numbers to answer it already exist. Epoch AI publishes a single capability score for each frontier AI model, drawn from more than fifty benchmark tests, and separately publishes how much power and how much computing each model took to train. Line the models up from the least powerful to the most, and you can watch how much extra capability each extra unit of power actually bought. If the return on each doubling stays high, the power is paying off; if it keeps falling, it is not.

One more thing to watch: the cost can be counted two ways. You can count the electricity a model pulled from the grid, the same kind of power a household pays for. Or you can count the computing it did, the raw amount of work, setting electricity aside.

Those two used to rise together. They don’t anymore, because the machines keep getting more efficient. Chips do the same work on about 34% less power each year since 2008, and training methods need about a third less computing each year to reach the same result. So a model can do far more computing than before while drawing only a little more electricity.

That gap is the part worth watching. It is the extra computing that efficiency lets the machines do without burning more electricity for it, so it never reaches the grid as new demand. Whatever efficiency can’t cover still turns into real power draw, and that is what the grid, and the household, are left to carry.

Its limits, stated plainly. Four things this number does not capture:

• The power figure counts only a model’s final training run. That is a fraction of the computing a developer actually burns to build it.
• It leaves out the power spent running the finished model for customers, day in and day out. That share is large and still growing.
• Capability here is a benchmark score. Doing well on tests is only a rough stand-in for being genuinely useful in the real world.
• There is no money anywhere in this data. It can show how much AI each unit of power buys. It cannot show how much money each dollar buys.

That last point is the one to hold onto. This page is about return on investment, which is a question about money. The number here is about power. It is the closest honest stand-in the public data allows, but it is a stand-in, not the real thing.

How you would know if this page is wrong

This page argues the money keeps flowing back in. That claim can be checked, and here is the simplest way to check it. The five biggest spenders, Amazon, Alphabet, Microsoft, Meta and Oracle, each tell their investors how much they plan to spend on things like data centers, and those numbers are public.

The test: by the end of September 2027, add up what those five say they plan to spend in 2027, and compare it to what the same five actually spent in 2026.

• If the 2027 total is lower than 2026, that is the first real sign the money is starting to pull out, and this page’s reading is weakening.
• If it is the same or higher, the money is still flowing and the reading is holding for now.

Two honest cautions. A drop does not prove the spending was ever wise, and a single down year could be a brief pause just as easily as a real retreat. One year is a signal, not a verdict.

What the little guy can do

Everything above is written as a reading, from the outside, with dates on it. This part is not. It is an opinion, stated plainly, and you are free to reject it. The rest of the page shows you where the risk sits. This panel says what a person with no fortune and no seat at the table can actually do about it.

Start from one fact the page already made. This buildout cannot run on itself. By the panel above, it cannot make its own chips, generate its own electricity, or approve its own power lines, and it cannot manufacture the belief that keeps investors paying. It buys all of that from outside. It looks unstoppable, a tide you can only surf or drown in. It is not a tide. It is a machine, and it runs on things that ordinary people supply: money, attention, power bills, and consent. The one thing worth seeing is that a machine which runs on what you give it can be affected by what you withhold.

So let’s be clear about what this panel is not. It is not a list of ways to tighten your own belt. The whole point of the page above is that a cost created by someone else’s bet is being quietly slid onto your bill. The answer to that is not to use less so you can absorb it more gracefully. The answer is to refuse to absorb it quietly, to make the people making the bet carry the risk of the bet, and to be awake enough to see it happening instead of finding out from a higher bill you assumed you had to pay. Awareness and pushback, not austerity. Five things a person can actually do.

1. Don’t look away. The single most useful thing on this whole page is the fact that the cost of this buildout is being moved onto ordinary households, in dollars, with the numbers shown above. Most people will only ever experience that as a bill that went up, with no idea why, and they will assume it is just how things are. It is not just how things are. It is a decision, made by specific people, in specific rooms, that could have gone another way. Knowing that a $21-a-month increase traces back to a data-center gamble, and not to some law of nature, is the whole difference between a citizen and a mark. Refusing to bury your head is step one, and it is not passive; understanding who did this is what makes everything after it possible.

2. Show up where it is actually decided. The prices in the tables above are not set by the weather. They are set in utility-commission proceedings, and those proceedings are open to the public. Earlier on this page: dozens of special rate deals for very large customers are being written right now, across dozens of states, mostly with no ordinary person in the room. When nobody shows up, the terms are written for the big customer and the cost slides to the household. When people do show up, terms like the ones in the Pennsylvania box above appear, the five-year contracts, the exit fees, the fund for customers who can’t pay. Those protections exist only because someone was in the room demanding them. You can be in that room. Public comment periods, commission hearings, and ratepayer-advocate offices exist precisely so you can.

3. Vote the offices that most people skip. Somebody decides whether your power company gets to pass this cost through to you. In many states that somebody is a public utility commission, and in a number of states its members are elected, or appointed by a governor you elect. These are the lowest-turnout, least-watched elections there are, which is exactly why they are where a small number of awake voters swings the most. The buildout is very good at showing up to these races. Ordinary people mostly are not. Learn who regulates your utility, learn where they stand on who pays, and vote it, every time, all the way down the ballot where the real decisions hide.

4. Push the cost back where the bet was made. This is the heart of it. The companies making the bet can slow their spending in a single quarter if it goes wrong. A power plant or a transmission line already approved gets paid off over decades, by you, whether the bet paid off or not. That asymmetry is the whole scandal of the last panel above, and it is not fixed by you using less electricity. It is fixed by rules that make the bet-maker carry the risk of the bet: large customers paying for the lines built to serve them, collateral held against projects that get cancelled, funds for households that fall behind. Demand those rules, by name, from the people in steps 2 and 3. The goal is not to make yourself smaller. The goal is to make the risk sit with the party that chose to take it.

5. Do it with other people. Here is the part there is no way around: one person alone against something this size is easy to ignore. One person is a complaint. A hundred people at the same hearing, a union, a tenants’ association, a ratepayer coalition, a neighborhood that votes together, is a decision the officials in steps 2 and 3 cannot pretend they did not hear. Every protection that has ever been won for ordinary customers was won by people acting together, not by individuals quietly economizing. If you take nothing else from this panel: the way the little guy stops being little is by not being alone.

6. The 80 percenters. The simplest version, and the one that hits where investors feel it. It is a pledge with a trigger. You go to a local website and sign your name. Nothing happens yet. If enough people in the area sign to cross a set number, the trigger, then everyone who signed pays 80% of their next bill and holds back the other 20%, in the same month. If the trigger is not reached, nobody does anything and you pay your bill in full. That is all. Nobody sticks their neck out alone, because the withholding only starts once the crowd is already there: enough of us, or none of us. A whole area 20% short in one month is not a billing error a utility can shrug off, and the message travels straight back to the investors funding this. Stop putting the cost of your bet on our bills. Get the mayor behind it and the local news covering it, and it is unmistakable. Call them the 80 percenters.

This is not a new idea. In 2022, the United Kingdom’s Don’t Pay UK campaign used the same pledge-with-a-trigger approach against soaring energy bills: people signed on in advance, and the plan was to withhold payment together only once enough others had committed to do the same. The 80 percenters is that idea, aimed at bills driven up to pay for someone else’s data-center bet.

Where this is weakest, said as plainly as the rest of the page. Showing up and voting does not guarantee the outcome; the money on the other side is enormous and patient, and it is at the same hearings you are. It is also true that time to attend a hearing is itself a luxury the households most crushed by these bills often do not have, which is why the collective forms, where a few people carry the load for many, matter more than any solo act of civic virtue. And you could reasonably believe the buildout will produce enough value to be worth its cost, in which case the fight is over who pays for it rather than whether it should happen at all, which is still worth having. This panel does not claim to know the future. It claims only that the cost landing on your bill is a choice other people are making, that you are allowed in the rooms where that choice is made, and that the worst thing you can do is assume there is nothing to be done and look away.

What this is, in one plain statement

Strip away the details and one shape is left. This is a machine that pays for itself by convincing people to keep paying for it. A better model brings in money; the money buys chips and power; the chips build a better model; the better model brings in more money. Each part exists because the others do, and the whole thing keeps running as long as the belief at the center holds: that the next round will be worth funding.

There is a name for this shape. The biologist Stuart Kauffman described how a set of parts, none of which can make itself, can still form a loop where each one helps produce the next, until the group as a whole sustains itself. No single member is alive on its own; the circle is what persists. The AI buildout is that circle, drawn in money: labs, cloud companies, chip makers and power companies, each feeding the next, no one of them in charge, all of them depending on the loop staying closed.

And here is the part that matters for anyone deciding whether to put money in, or bracing for what it does to a power bill. A loop like this does not feed itself. It feeds on something from outside, and here that something is funding. It cannot mine its own chips, generate its own electricity, or print its own investors’ confidence. It buys all of that with money that keeps arriving only while the belief holds. So the question is not how big or how fast it is; big and fast is exactly what these loops look like right up until the outside supply stops. The question is whether the food, the funding, keeps coming, and who is left holding the cost if it doesn’t.

And that is the trap in the timing. The money can leave in a quarter, but the costs it leaves behind cannot. A funder can stop funding with a single decision; the power plants and the transmission lines built for the boom are already in the ground, and they get paid off over decades by whoever is on the electric bill. So the loop’s food is quick to disappear and the bill it ran up is slow to clear, and those are not paid by the same people. The ones who placed the bet can walk; the ones who never made it are left paying for the buildout long after the belief that drove it is gone.