Who Keeps the Savings?

I feel like pretty much every essay about AI right now asks and answers the same question: is AI here to stay? Will AI take our jobs? Is SaaS dead? etc, etc, etc. Yes. It's boring. To be honest, I think we must move on, because the interesting questions all start one step later. One I think is more controversial and has kept me circling back is this one: when a tool makes work cheaper, who keeps the savings?

For a while I told myself a comfortable answer. The friction of building things is collapsing, so the small player gets stronger: anyone can build now, the local underdog finally has leverage, the village gets the same superpowers as the city. I believed a softer version of the line you hear at every conference: that AI democratizes. I've come to think that's half true in a way that makes it mostly false. So let me try to say the harder thing.

Democratization of capability, concentration of power

Here is a fact: AI democratizes capability but concentrates power, and those two things move in opposite directions.

It's genuinely true that one person can now do what used to take a team. The tool in your hand got dramatically more powerful. But the same is true of the giant's tool, and the data, the compute, the distribution, and the lawyers all flow harder toward whoever already had them. So the small player is more capable in absolute terms and weaker in relative position. Their tool improved; the gap is widening anyway, because capability is cheap now and the things that actually decide outcomes are not.

We confuse the two constantly. "Anyone can build an app" feels like power. It isn't. And as we are all realizing, building was never the bottleneck. The bottleneck has always been distribution, trust, capital, and access to the state. And AI does almost nothing to hand those to the small player while doing a great deal to sharpen them for the large one. Democratization of production is real. Democratization of power is mostly a line in a pitch deck.

The pattern was never only about technology

To see why, stop looking at software and look at the places where capital meets a small community directly (I'll mention regions in Mexico, USA and Europe as examples).

Take Chiapas. The region around San Cristóbal is famous for two things that turn out to be the same thing: extraordinary rates of Coca-Cola consumption, and a bottling operation drawing local water under concessions a local cooperative could never obtain. Or take the palm oil plantations spreading across the southern lowlands, reorganizing land and water around an export crop. Neither of these is a story about superior technology beating inferior technology. It's a story about capital arriving with overwhelming resources and capturing a resource base: water, land, distribution, even ritual and habit, before the locals can respond, and doing it with the state's signature on the paperwork.

The decisive advantage was never a better bottle. It was the concession, the capital, and the relationship with the government. The local player didn't lose because their drink was worse. They lost because the contest was never held on the field they could see.

AI does not break this pattern. It is a new vector for the same old move. And here is the part that I think should worry the optimists: the real danger isn't that AI replaces the local player with a robot. It's that AI lets a distant incumbent understand, localize into, and extract from a small market faster and cheaper than ever; often before the local player even realizes the market is now contested.

Who kept the savings? Not Chiapas.

I have to take back something

Which means I owe a correction to the comfortable version of my own argument.

I used to say the moat moves to context: to language, to trust, to knowing the terrain, and that this favors the local, the rooted, the small. That's true, and it's the more dangerous half that I left out: AI just dropped the price of acquiring context.

The incumbent can now hire the model instead of the anthropologist. It can study your market, localize into Tzotzil or regional Spanish, model the buying habits of a town it has never visited, and arrive with infinite patience and a localization budget you can't match. Then it just has to adjust info that wasn't updated and listo: they started with a massive advantage. Context was a moat only for as long as it was expensive to acquire. AI lowered that cost, which means it lowered the threshold of market size that's worth an incumbent's attention. "Be local, be trusted" is still necessary. But it is no longer sufficient. It buys you time, not safety.

The technology is not neutral

While writing this post I started asking myself whether AI is changing capitalism or capitalism is changing AI, as if it were a clever paradox. But I think it isn't, and the resolution matters.

Here's what I see happening: AI isn't rewriting capitalism's logic. It's pouring fuel on the tendencies that were already there: concentration and commodification first among them. But the sharper truth runs the other way. Because AI is built by capital, it is optimized for capital's goals: cutting labor cost, capturing attention, extracting margin. Worker augmentation and local flourishing were never the brief. They're side effects we hope for, not objectives anyone is funding.

This is the thing to sit with. A technology carries the incentives of whoever paid to build it. The current wave was paid for by the largest concentrations of capital in human history, and we are watching them redirect tens of thousands of jobs and hundreds of billions of dollars toward owning the infrastructure the rest of us will rent. When Meta grows revenue by a third and still cuts thousands of people to fund compute, that isn't a robot taking a job. It's a job sold to pay for the road, so that everyone who drives on it later pays a toll. Value created here, captured there. Once you see that shape you see it everywhere I'd say... in the layoff, in the bottling plant, in the plantation.

Where you're actually standing

And don't read "the small player" as only the faraway farmer. The clearest case is probably the person reading this. The mid-level analyst, the in-house designer, the engineer at Oracle who got cut. Or the blue-collar accountant who spends 10+ hours a month on a stupid reconciliation process at months end: they had the most capability handed to them, they're the ones with Claude open all day, and they're watching their position erode fastest. The coffee co-op losing ground to a platform and the analyst losing ground to a model are not two stories. They're the same event at different latitudes.

Here's the twist that's just starting to surface. A lot of the firms that fired the analyst are now quietly spending more, sometimes far more, on tokens to do what the analyst did, and getting something thinner back. The tempting read is "see, AI doesn't work, they'll hire everyone back." Don't take that bait; it's the cope, and it's wrong. Inference gets cheaper every year, and most companies won't reverse the layoffs, they'll just trim the AI bill. The real lesson is sharper: the analyst was never only producing the legible output you were paying for. They were also holding the part that doesn't fit in a context window, knowing which number was wrong before anyone checked, remembering why the weird exception exists, feeling a client about to walk. The market is about to spend a fortune rediscovering that the irreplaceable part of the job was exactly the part nobody had priced. Context, again. The same scarce thing this whole essay keeps circling.

And notice who can afford to run that experiment. A small shop can't survive spending six figures to learn that tokens don't replace its best person. A giant can spend it as a rounding error and keep the lesson. So even the failures concentrate: the capacity to be expensively wrong, survive, and learn from it is itself a luxury good. Which is why "maybe AI is overhyped" is no comfort at all. Being able to afford the overhype is the advantage.

So if you're comfortable and credentialed and a little smug about being safe, I have bad news about where you're standing: from where the capital sits, you're the periphery too.

The infrastructure has a hometown

And here is where the comparison stops being a comparison. I keep holding AI up against the bottling plant and the plantation as if it were a metaphor, but the truth is that it isn't. AI's physical infrastructure does the same thing, in the same way, right now.

A data center is the bottling plant. It lands in a community, draws down a shared resource (here it's electricity and water instead of spring water), and exports the value somewhere else. The numbers aren't subtle. US data centers are on track to nearly double their electricity demand between 2025 and 2028, from roughly 80 to 150 gigawatts; that's like bolting a second Spain onto the grid in three years. They already account for about half of all new electricity demand in the country. And the bill doesn't stay in the boardroom: in the neighborhoods around Virginia's so-called "data center alley," residential electricity bills have climbed more than 250%. An economist at Carnegie Mellon put the hidden health and environmental cost of US data centers at roughly $25 billion a year (the air pollution, the diesel backup generators, the strain on local grids) and it's paid by the people who live nearby, not the people who own the servers.

And, exactly like the plantation, these things tend to land where resistance is cheapest, in communities already carrying pollution and already short on leverage. One watchdog analyst described the projects as extractive and bringing very little to the towns that host them. Seven in ten Americans now say they don't want one built near their home. People can feel the shape of it even when they can't name it: the value is produced in their backyard and kept somewhere else.

This is what "the technology is not neutral" looks like when you can touch it. The cloud has a hometown, and it's usually a town that didn't get a vote.

Who kept the savings? Not the town.

Watch where the savings go

If you want to know who keeps the savings, watch the IPOs.

As I write this, three of the largest public offerings in history are lining up at once: SpaceX, OpenAI, and Anthropic, the company that makes Claude, the very tool I'm partly using to think this through. SpaceX alone is targeting somewhere around $1.75 trillion, which would make it the biggest IPO ever, larger than Saudi Aramco. OpenAI filed targeting $852 billion to over a trillion, while losing more than a dollar for every dollar of revenue it earns. And Anthropic just raised $65 billion in a single round at a $965 billion valuation, more than doubling its price from earlier this year and passing OpenAI for the first time, in what's reported to be its last raise before going public. Together the three represent more than $3 trillion in valuation. It is the most concentrated pile of capital ever assembled around a single technology, and it is about to go liquid.

Look closely at that Anthropic round and you can watch the loop turn. About $15 billion of the $65 billion wasn't new outside money at all; it was capital already committed by the hyperscalers, including $5 billion from Amazon. Amazon funds Anthropic; Anthropic spends much of it back on cloud compute; the same dollars circle inside the same few balance sheets, and each lap the lead gets a little more unassailable. This isn't a market of many players bidding. It's a handful of giants passing the same chips around a very small table, while the rest of us watch the pot grow.

Now, the eye-watering "returns" floating around get garbled, so let me be precise, because the precision is the point. There are two different multiples people mix up. One is how expensive the stock is: OpenAI priced at around 35 times its revenue, which only tells you how much hope is baked in. The other is what the early backers stand to make, the seed and Series A investors sitting on gains one analyst could only call almost incomprehensible. That second number is the one that matters here, because an IPO is, mechanically, the moment that early advantage cashes out.

And here's the tell. These offerings are being structured to hand a larger-than-usual slice to retail buyers; Anthropic is reportedly targeting about 30% retail, roughly three times the norm. That sounds like democratization, the little guy finally gets in. Look closer. The little guy gets to buy, at the top, from the people who got in early and cheap. We have seen this movie: Figma went from $33 to $143 to $22 in eight months. The machinery is designed, entirely within the law, to move wealth from late buyers to early holders. The small player isn't being invited to the table. He's being invited to be the exit.

Then the loop closes. The money raised in these offerings flows straight back into building more data centers, more power pulled from more towns that didn't vote, to widen the very same lead. Capital concentrates, buys the infrastructure, pushes the cost down onto a community, and sells the story back to the public at the peak. That's the whole machine in one sentence. When I ask who keeps the savings, this is the answer, written in numbers with a great many zeros.

Who kept the savings? Not retail.

A complication I can't leave out, though: the middle class isn't only the chips. Through your pension, your index fund, that "generous" 30% retail slice of the IPO, you own a sliver of the table too. You're the raw material and a tiny shareholder in the machine that processes you. That isn't hypocrisy, it's the trap: it's hard to fight a system that cuts you a check small enough to keep you quiet and large enough to keep you invested.

So what actually resists capture?

My take? Less than the empowerment story claims, and the things that do work are uncomfortable for a tech narrative, because almost none of them are technological.

The first is legitimacy that can't be purchased: being of a place rather than a visitor to it. An incumbent can model a small population in the mountains of Switzerland. But it cannot be that place. Trust that is relational and historical does not transfer with a dataset, and that's not nothing.

The second is staying valuable while staying boring: living in the giant's blind spot, in markets too small, too messy, or too relationship-bound to be worth the cost of entry. A lot of good businesses survive precisely by not being worth conquering. The trouble, again, is that AI keeps lowering the cost of conquest, so the blind spot keeps shrinking.

The third is the one the optimists skip, and it's the only structural answer: collective ownership. The historical response to concentrated capital was never individual cleverness. It was pooling; cooperatives, mutuals, unions, commons, shared infrastructure. If data and compute are the moats, then collectively owned data and shared tooling are the only counter that scales. A coffee cooperative that owns its own market data sits in a fundamentally different position than one renting intelligence from the very buyer it's negotiating against. The question of who owns the model is going to matter more than the question of who can use one.

And the fourth is the one that makes business people uncomfortable, so let me be precise about it. What protected those communities, where anything did, was never a cleverer app. It was rules: zoning, water rights, the moratoriums that have already stalled tens of billions of dollars in data centers. The towns that won didn't out-build the giant. They changed the field.

Here's the part worth sitting with. The field was always the deciding factor. Coca-Cola didn't win on a better bottle; it won on a concession, which is to say, on a rule. The market was never a neutral arena that capital happened to dominate. It was structured by power from the start. So the honest question isn't "market or regulation", there is no unregulated market, there never was. The question is only ever who the rules favor.

Which means regulation is the small player's most underrated weapon, but only one kind of it. There are two, and they point in opposite directions. Pro-competition rules, antitrust, the right to move your own data, interoperability that pries open the gatekeepers, repair rights, the cooperative and tax law that makes shared ownership legal, these lower the wall. Compliance-heavy rules, thick books of process you need a legal department to survive, quietly raise it, which is why the biggest players so often welcome them. When an incumbent asks to be regulated, watch which kind they're asking for. Usually it's the moat, dressed up as responsibility.

And none of this gets written by the periphery alone. It never has. The rural producer can't demand it into existence; neither can the laid-off engineer. But together, and that's the whole point of pulling the comfortable middle class into this story earlier, they're the coalition that has actually won these fights before: antitrust, the eight-hour day, deposit insurance. Individual hustle does not beat structural advantage. Only structure beats structure. And structure only changes when enough people who thought they were safe realize they aren't, and ask for different rules.

Back to the only question that mattered

So: does AI democratize anything?

It democratizes capability, which is real, and which I don't want to wave away: a clinic that can finally dictate its records in regional Spanish, a producer who can read a contract in a language the contract wasn't written to be read in, an agency that didn't have the capacity to deliver enough digital content... these are genuine goods. But it concentrates power at the same time and faster, and capability without power is a more comfortable form of dependence, not freedom.

When a tool makes work cheaper, the cheapness is real. The only question is who keeps it.

So: who keeps the savings? Not Chiapas. Not the town. Not retail. Not you. Whoever already had the data, the compute, and the concession. That's the whole answer; everything above is just the proof. It does not have to stay that way. But it will not change because the tools got better, and it will certainly not change by accident. It changes only if we decide, deliberately, collectively, and probably politically, to build and to govern toward a different answer.

AI is here to stay. Fine. The fight worth having was never about the machines. It's about who they stay for.

A quick note:

I wrote part of this with Claude. Anthropic just raised sixty-five billion dollars. Which means that somewhere in the price of this essay's production is a rounding error that helped pay for another data center, in another town that didn't get a vote. I am, fairly literally, the product. So are you. Anyway... thanks for reading, and welcome to the table. You're the chips.

Sources

Data centers, energy, and local impact

• Consumer Reports, AI Data Centers: Big Tech's Impact on Electric Bills, Water, and More — projected demand growth (≈80→150 GW by 2028), the "extractive" characterization, and the wave of state-level pushback. https://www.consumerreports.org/data-centers/ai-data-centers-impact-on-electric-bills-water-and-more-a1040338678/
• Fortune, Data centers cost the U.S. economy $25 billion a year in hidden health and environmental damage — the ≈$25B/year figure (Nicholas Muller, Carnegie Mellon) and the 250%+ rise in bills near Virginia's "data center alley." https://fortune.com/2026/04/21/data-centers-environmental-health-costs-25-billion/
• Fortune, Data centers now account for half of all new U.S. electricity use — the share of new U.S. electricity demand. https://fortune.com/2026/04/20/us-data-center-electricity-demand-public-opinion/
• Fortune, Americans' AI hate wave might just be gathering steam — the Gallup figure that ~7 in 10 oppose a nearby data center. https://fortune.com/2026/05/19/data-centers-electricity-costs-us-public-opinion/
• Lincoln Institute of Land Policy, Data Drain: The Land and Water Impacts of the AI Boom — water/land burden falling on host communities. https://www.lincolninst.edu/publications/land-lines-magazine/articles/land-water-impacts-data-centers/

AI valuations and IPOs

• Anthropic, Series H (official announcement) — the $65B raise, $965B post-money valuation, and ~$47B run-rate revenue. https://www.anthropic.com/news/series-h
• Bloomberg, Anthropic's Valuation Nears $1 Trillion After Raising $65 Billion — the raise and the first-time eclipse of OpenAI. https://www.bloomberg.com/news/articles/2026-05-28/anthropic-raises-at-965-billion-valuation-eclipsing-openai
• TechCrunch, Anthropic raises $65 billion, nears $1T valuation ahead of IPO — the $15B of previously committed hyperscaler money, including $5B from Amazon, and the "last private raise" framing. https://techcrunch.com/2026/05/28/anthropic-raises-65-billion-nears-1t-valuation-ahead-of-ipo/
• Axios, Anthropic tops OpenAI as most valuable AI startup — valuation context and revenue run-rate. https://www.axios.com/2026/05/28/anthropic-ai-fundraising-openai
• MarketWise, SpaceX IPO 2026: The AI IPO Trap... — the "wealth transfer from late buyers to early holders" framing and the Figma price arc ($33→$143→$22). https://marketwise.com/investing/spacex-ipo-2026-anthropic-openai-figma-trap/
• IndMoney, SpaceX, OpenAI & Anthropic IPOs 2026 — the >$3T combined pipeline and early-investor returns. https://www.indmoney.com/blog/us-stocks/spacex-openai-anthropic-ipo-explained
• AI Funding Tracker, AI IPO Tracker 2026 — SpaceX's April 1 confidential S-1, the $1.75–2T target, and Anthropic's ~30% retail allocation. https://aifundingtracker.com/ai-ipo-tracker/
• Memeburn, OpenAI IPO 2026 Speculation Grows... — OpenAI priced at ~35x annualized revenue. https://memeburn.com/openai-ipo-2026-speculation-grows-after-6-6b-employee-sale/