In 2026, a company whose entire business was knowing things lost roughly 60% of its value over the past year.

Gartner sells research. For decades, executives paid subscription fees for access to analysts who had spent careers accumulating industry knowledge, and the arrangement worked because that knowledge was genuinely hard to assemble on your own. Then it stopped working. Gartner’s contract value, the metric that measures the health of the subscription base, grew 1% year-over-year in a quarter where management had promised acceleration. The stock fell more than 20% in a single session, and months later it fell more than 27% in another. A Goldman Sachs analyst wrote that the company faced structurally higher AI risk (Gartner noted that excluding US federal government business, its contract value grew by 3.5%, still anemic). Shareholders sued.

Chegg sells homework help. It was worth about $14 billion at its peak. It is now worth roughly $100 million, a decline of about 99%, and it has cut 45% of its staff. Students who once paid a subscription to see worked solutions now ask a chatbot for free.

Accenture sells expertise by the hour. Its shares fell sharply through 2025, including an 18% single-day drop, and the company announced cuts of 22 thousand people. KPMG cut 4% of its US advisory practice. In India, the combined market capitalization of the five largest IT services firms fell 46% from its 2024 peak. Intuit, whose TurboTax business stands between American citizens and their own tax returns, became one of the worst performers in the S&P 500 and fell to a five-year low after a Goldman analyst pointed out that an AI model can process a return for roughly twelve cents against TurboTax’s average revenue per user of about $162.

These read like four unrelated corporate misfortunes. They are one event.

AI has redefined what is scarce

Every business on that list occupied the same structural position. Each one stood between a person and something that person wanted and charged for the passage. Gartner stood between an executive and industry knowledge. Chegg stood between a student and an answer. Accenture stood between a corporation and skilled labor. Intuit stood between a citizen and the tax code. None of them owned the thing on the other side. What they owned was the difficulty of getting there.

That difficulty is what AI dissolves. It leaves knowledge and skill intact. What it removes is the friction that makes either one expensive to reach, and a toll booth with no friction to sell has nothing left to charge for.

Henry George ran into a version of this problem in 1879. He was writing during the most productive decades in human history to that point. Railroads, telegraphs, steam shipping and industrial manufacturing had multiplied output per worker beyond anything previous generations could have imagined. And yet, George observed, poverty had not disappeared. In the most advanced cities it had deepened. His explanation was that the gains from productivity were being captured as rent by those who owned land, because land was fixed in supply. As everything else got more efficient, the value of standing in the right place absorbed the difference.

You do not have to accept George’s politics to find the mechanism useful. His insight was that productivity gains flow to whatever is scarce, and that the identity of the scarce thing can change while everyone is still watching the old one.

For the past forty years, the scarce thing in the developed economy was the ability to process information. That scarcity is what a professional credential certified. Law, medicine, accounting, consulting, financial analysis and academic research all rest on the same premise: the information exists, it is difficult to navigate and I have been trained to navigate it on your behalf. The premium those professions command is a rent on that difficulty. We called it human capital, a phrase that made it sound like something you own instead of somewhere you stand.

AI attacks that position directly. It does not have to be better than a good analyst to break the economics. It only has to be good enough to make the passage cheap. Meanwhile, the scarce thing is moving somewhere older and more physical, and you can see where, following the capital.

Oracle spent about $56 billion on capital expenditure in a single fiscal year. It ran roughly $24 billion of negative free cash flow. It carried something in the region of $130 billion of debt, pushed its debt-to-equity ratio toward 300%, and was downgraded by S&P to the lowest rung of investment grade. Its stock fell about 64% from its September peak, and June was its worst month since 1990.

Notice what the market punished and what it did not. Investors punished the leverage and the customer concentration. Almost nobody argued that the assets themselves were worthless. The debate was about whether Oracle could survive holding them, which is a very different debate from whether they were worth holding.

Physical assets and constraints are what matter

Across the industry, the same purchase is being made on an enormous scale. The money buys compute, and it also buys something less discussed: land near high-voltage transmission, positions in utility interconnection queues that can take years to clear, water allocations, permitted sites, substation capacity and the electrical generation to feed all of it. These are the assets that cannot be summoned by writing a check, because the constraint on them is legal and geological, and no amount of financing changes it.

The precision there matters, because “physical assets appreciate” is too loose to be true. Intel owned an enormous physical plant and had a miserable decade. Owning factories did not save it. What Intel lacked was a binding constraint that others could not route around, since its fabs were not at the leading edge and its competitors could get capacity elsewhere.

The assets appreciating now are the ones where the constraint binds. There is one company in the world that makes extreme ultraviolet lithography machines. There is a finite amount of firm power that can be delivered to a given county before the transmission system needs a decade of upgrades. Water rights in the arid West are legally allocated and largely spoken for. Permitted land within reach of existing transmission is a fixed inventory that no amount of enthusiasm expands. You can have all the capital in the world and still wait in line.

This is the trade being made across the economy right now, and almost nobody has described it plainly. Capital is fleeing rents on information and racing toward rents on physical constraint. The people who own credentials are on one side of it. The people who own ground, water and electrical interconnection are on the other.

None of which means the market carnage of 2026 proves the story. A substantial part of it does not, and the honest version has to say so.

Markets are repricing, not reacting

The software sector had a brutal year. The main software index fell more than 20% while the S&P 500 was roughly flat. Median valuations for public software companies fell to about three times revenue, down from nearly five times a year earlier and more than six times the year before that, which puts the sector back where it traded in 2015. Roughly a trillion dollars of market value evaporated in the first quarter alone.

Very little of that came from AI eating anyone’s business. It came from investors deciding to pay less for the same earnings.

The tell is that the companies kept beating expectations while their shares fell. Salesforce grew revenue 13% and reported that its AI product line had passed a billion dollars in annualized revenue, growing more than 200%. ServiceNow grew subscription revenue nearly 25%, raised its AI target by 50% one quarter into the year, and reported that the number of customers running AI agents in production had grown ninefold in nine months. Figma grew revenue 46% with customers spending 39% more than the year before, and more than three-quarters of its enterprise users who hit their AI usage caps chose to pay for more. Atlassian’s revenue accelerated, and its seat counts kept expanding, which is exactly the thing the bear case said could not happen.

All of those stocks fell hard anyway. When strong results cannot move a share price, the market has stopped reacting to the quarter and started repricing the distant future.

There is also good evidence that the AI disruption everyone is pricing has not actually arrived in most enterprises. A study out of MIT examining hundreds of deployments found that most custom enterprise AI projects never reached production. Gartner, before its own troubles, predicted that more than 40% of agentic AI projects would be canceled by the end of 2027 on grounds of cost and unclear business value. Corporate IT budgets grew only about 3.5%, and the growth was reallocated toward compute instead of added on top, which is why both hardware and services suffered.

The question of loss is a question of transfer, not redistribution

So, the picture is mixed, and any honest account has to hold two things at once. The broad software selloff was mostly a valuation event driven by fear of a future that has not happened yet. The specific destruction of the intermediaries was real and permanent, and it showed up in their financial statements.

That distinction is the whole argument. The businesses that broke were the ones whose only asset was standing in the way. The businesses that merely got cheap were the ones that own proprietary data, regulated workflow or something a customer cannot rebuild over a weekend. The market has not finished telling those two groups apart, which is why the sorting looks so chaotic.

That unfinished sorting is one reason the word “bubble” keeps surfacing, and the comparison everyone reaches for when it does is the railroads, or the fiber boom of the late 1990s. It is offered as reassurance. Yes, investors were wiped out, the story goes, but the track and the fiber remained, and the productivity arrived a decade later for everyone else. Manias are how societies finance infrastructure they would never approve rationally.

The consolation does not fully transfer, because of depreciation. Track laid in 1873 carried freight for a century. Fiber laid in 1999 sat dark for five years and then ran the internet for the next twenty. The largest single line-item in AI capital expenditure today is advanced semiconductors, and those have an economic life of roughly three to six years, depreciating against an improvement curve that makes each generation obsolete faster than concrete or steel ever aged. If this cycle breaks in 2027, nobody lights up the 2025-vintage chips in 2033. They are simply impaired.

But split the spending into two layers and something more interesting appears. The fast-depreciating layer is compute. The durable layer is everything underneath it: the land, the buildings, the substations, the transmission interconnections, the water allocations and the permits. Those have useful lives measured in decades. They are the railroad-like assets of this cycle, and they will outlive whatever happens to the chips inside them.

Which means a crash does not distribute anything. It transfers.

If the cycle breaks, the over-levered marginal players sell the durable layer at distressed prices to whoever has the balance sheet to hold it through the trough. The buyers will be the same handful of entities with enormous operating cash flow who are already the largest owners. The permits do not get reissued to the public. The interconnection rights do not go back in the queue for a community solar project. They change hands, cheaply, in the direction of concentration.

This inverts the intuition most people bring to the word bubble. We think of a crash as a correction, a moment when things that got too expensive become available again and the overreach is punished. In an economy where the scarce asset is physical position, a crash is an acquisition event. It would accelerate the concentration it appears to punish.

Where exactly is spending shifting to?

On whether it is a bubble at all, the strongest counterargument deserves airing. Unlike 1999, most of this spending comes from companies with real profits, funded out of operating cash flow instead of speculative equity issuance. That is a genuine and important difference. The rebuttal is that bubbles break at the margin, and the margin here is levered. It includes companies carrying credit downgrades, a set of smaller specialized cloud providers financed largely with debt, a fast-growing private credit market with limited disclosure, and a web of circular arrangements in which model developers, cloud providers and chip manufacturers invest in one another and then book each other’s spending as revenue. The aggregate looks solid, the edge looks like 2007, and the edge is where these things always start.

If it does start there, the question becomes who absorbs it, and here the language gets loose. “Socializing the losses” invites a fair objection, since a large technology company is not a bank and a write-down at a profitable firm requires no bailout. For the claim to hold, the mechanism has to be named. There are at least five, and they are unequal in strength.

The most concrete is electricity. When a utility builds generation and transmission to serve enormous new industrial loads, that investment goes into the rate base, and the rate base is recovered from customers. Regulators have been attempting to design special tariffs that isolate large data center loads and make them bear their own costs. Whether those tariffs hold is being decided right now in public utility commission proceedings across the country, and the outcomes are mixed. Where they fail or get negotiated down, the cost of infrastructure built for a handful of firms appears on the bills of everyone else on the system. This is a live administrative fight with a documented paper trail.

The second is municipal. Local governments compete for data center projects with property tax abatements, sales tax exemptions on equipment, and publicly funded road, water and sewer upgrades, and they justify these against projected employment. Data centers are exceptionally capital-intensive and employ very few people once built. When the jobs do not arrive at the promised scale, the community has already spent the money and forgone the revenue.

The third is portfolio concentration. Public equity indices are now heavily weighted toward a small number of firms tied to this buildout. Ordinary retirement savings, held passively in index funds and target-date products by people who made no active bet on artificial intelligence, are exposed to the outcome. A drawdown transmits directly into pension funds and 401(k) balances.

The fourth is credit, and it is the least examined and the most likely to produce genuine contagion. Private credit has expanded enormously with thin public disclosure, and it has been an active lender to data center developers and specialized cloud providers. Vendor financing arrangements, where a supplier funds its customers’ purchases of its own products, have historically been a reliable warning sign, and they were a feature of the telecom collapse in 2001. When these unwind, the losses land in insurance portfolios, pension allocations and bank credit lines, well out of sight of the public equity markets.

The fifth, still emerging, is direct public participation, where governments take equity stakes, offer loan guarantees or extend sovereign support to strategic projects.

Aggregate versus localized arguments

The ratepayer channel is the one to lead with, because it is happening now and it is legible to anyone who opens an electricity bill.

That bill is also where the public conversation about the physical costs of all this has gone wrong, in both directions. People have been loudly worried about land, water, energy and carbon, and the response from industry has been to treat the worry as innumeracy. Both sides are partly wrong, and the errors are specific enough to separate, because those four resources do not behave the same way.

Water is where public alarm most exceeds the aggregate evidence. Much of the frightening coverage relies on consumption figures from evaporative cooling designs that newer closed-loop facilities have largely moved past, and at national scale, data center water use remains small next to agriculture and thermoelectric power generation. And yet the alarm is frequently correct in the specific instance, because water is a local resource and the national total says nothing about the county. A facility drawing on a single stressed aquifer in a drought-prone county imposes a real cost on that county, whatever the continental average shows. The honest formulation is that the fear is wrong at the scale of the country and often right at the scale of the community, and answering it with a national average answers nothing.

Electricity runs the other way. Here the evidence is ahead of the sentiment. Public opposition to data centers is organized around visible grievances: the appearance of the buildings, the noise of the cooling systems, the truck traffic, the water. Meanwhile, the mechanism most likely to remove money from a household budget is invisible and procedural, consisting of capacity market pricing, transmission cost allocation, rate case outcomes and the terms of interconnection agreements. People are angry; they are right to be angry, and their anger is aimed slightly to the left of the injury. Nothing about that is irrational. The public simply cannot see the pipe the cost flows through, because that pipe runs through proceedings that are technically public and effectively invisible.

Carbon is genuinely contested. The buildout has driven new natural gas capacity and delayed the retirement of coal plants that were scheduled to close. It has also driven the largest corporate procurement of clean energy in history and revived commercial interest in nuclear power for the first time in a generation. Serious people looking at the same data reach different conclusions about the net effect. Admitting that buys credibility for the places where the evidence is one-sided.

The same lesson runs through all of it. Aggregate statistics are the wrong instrument for evaluating a phenomenon whose entire character is local concentration. The whole point of a data center campus is that it puts an extraordinary amount of demand in one specific place. Averaging that across a continent is a way of not looking at it.

Ownership is moving, and the professional class must be aware

A peculiar feature of the moment sits on top of all this. The most prominent voices warning that AI will displace enormous numbers of workers are, in many cases, the people building the systems that would do the displacing. And the remedies they propose, in their various forms, share a structural characteristic: Each one transfers a portion of the income while leaving ownership of the underlying assets entirely intact.

That has to be stated carefully, because these individuals hold genuinely different views and flattening them is both unfair and easy to disprove. Some have funded serious empirical research into cash transfers. Others have argued for very different mechanisms. The disagreements among them are real.

The structural observation survives all of it. Every proposal on the table redistributes output. None redistributes position. A dividend paid from the returns on an asset is not the same thing as a claim on the asset, and the difference compounds forever. If the argument here is right, and the durable scarcity is shifting toward physical constraint, then a policy debate confined to income transfers is a debate about the smaller half of the question. The larger question is who ends up holding the land, the water rights, the interconnection capacity and the generation. That question is being settled right now, in permitting hearings and rate cases, while the public conversation is about robots and universal basic income.

The strongest objection to everything above is that it sounds inevitable, and inevitability is usually a sign that someone has stopped looking at the mechanism.

Almost nothing in this account is technologically determined. The specific outcomes depend on decisions being made by identifiable people in identifiable rooms, and most of those decisions could go the other way. Whether large loads pay their own infrastructure costs is a tariff design question, settled by state utility commissions. Whether interconnection queues favor whoever arrived first with capital or are restructured to weigh public benefit is a matter of federal and regional grid policy. Whether municipalities keep bidding against each other with tax abatements is a question of state law, and states have limited such competition before. Whether water allocations are treated as permanent private entitlements or as revocable public licenses is unsettled law in much of the West. Whether the credit exposure in private markets is disclosed at all is a securities regulation choice.

Each of these is a live and boring administrative question, and the boredom is why they are being resolved in favor of concentrated capital. The parties with billions of dollars at stake send lawyers to every hearing. The parties with electricity bills do not know the hearings exist.

The professional class has spent decades assuming its position was secured by education, and it has spent its political energy on cultural questions while the infrastructural ones went unattended. That assumption is now failing on the economic side, and the failure is arriving in the specific form of falling demand for intermediation. The lawyers, analysts, consultants and accountants who thought they had bought their way onto high ground are discovering that the ground they bought is being repriced, while the ground that is appreciating is measured in acres, acre-feet, megawatts and queue position.

Henry George’s mistake, if he made one, was believing that naming the mechanism would be enough to change it. It was not. The rent moved, and it kept moving, and the people who understood where it was going got there first.

It is moving again. The only remaining question is who is in the room when it is decided where it lands.

[Cheyenne Torres edited this piece.]

The views expressed in this article are the author’s own and do not necessarily reflect Fair Observer’s editorial policy.