Every hype cycle leaves something behind. It is never what was promised.

I keep a ledger of my predictions, so I know how often I've been wrong about technology. Plenty. One pattern has held up for forty years, though, and I'd bet my pension on it again: every hype cycle leaves something useful behind, and it is never the thing on the brochure.

That's the charge. Here is the evidence, and the verdict on AI comes at the end.

Exhibit A: the railways

In 1846 the British Parliament passed 263 acts setting up new railway companies, with proposed routes totalling 9,500 miles. Then the Bank of England raised interest rates, the money went back into bonds, and the share prices fell apart. The promise had been that railway shares would make ordinary people rich. For most of the people who bought them at the top, they didn't.

The track stayed. Projects authorised between 1844 and 1846 produced about 6,220 miles of railway (Wikipedia's summary has the figures). For scale, the whole UK network today is around 11,000 miles. The shareholders paid for it and their grandchildren rode on it.

Exhibit B: the glass in the ground

I spent the late nineties at Sun, so I had a front row seat for this one. The pitch was a "new economy" where the old rules no longer applied. Global Crossing was valued at $47 billion in 1999. It never had a profitable year, filed for one of the biggest bankruptcies in history in 2002, and Level 3 picked up what was left in 2011 for $3 billion, debt included (source). In July 2001 Merrill Lynch estimated that only 2 to 3 percent of America's fiber was actually in use (CNN, 2001).

The investors were wiped out. The fiber stayed in the ground and became cheap, and that cheap bandwidth is a big part of why streaming video and the public cloud worked a few years later. Nobody in 1999 told shareholders that their losses would subsidise a website full of cat videos. That is what happened anyway.

Webvan belongs in this exhibit too. It burned through its money, went bankrupt in 2001, and became the punchline for every dot-com joke I told for a decade. Grocery delivery works fine now. The idea was right and the company was early, which in this business is the same as being wrong, except for the people who come later.

Exhibit C: my own Hadoop cluster

In 2012 I ran a Hadoop cluster that ate my weekends for a year. The promise was insight. Data was the new oil, and if you collected all of it the answers would surface on their own. They didn't. The dashboards lied just as well as before, especially to managers who wanted them to. Cloudera and Hortonworks merged in 2018 and the market treated it like a wake.

What stayed was real, though. Storage got cheap enough that keeping raw data and deciding later became normal. Columnar file formats stuck around. Spark came out of that era, and I'll admit, grudgingly, that it fixed problems I actually had. The insight never showed up. The plumbing did.

Exhibit D: for the defence, blockchain

A fair prosecutor admits when a cycle leaves almost nothing. Blockchain promised trustless money and an internet without middlemen. Fifteen years on, the most useful thing it produced is the stablecoin, a way to move dollars around that ends up looking a lot like a bank. Some very good cryptographers got trained along the way. The hardware built in bulk was mining rigs, which do exactly one job. When that job stopped paying, they became scrap.

So the pattern has an exception, and the exception tells you how the rule works.

The rule

The things that outlive a bust are the ones that keep working after their owner goes bankrupt, and get more useful as they get cheaper. Track, fiber, cheap storage, open source code. Anything that only made sense at bubble prices dies with the bubble.

None of this is my idea. The economist Carlota Perez described it years ago: a period of financial frenzy builds the infrastructure, a crash follows, and a slower "deployment" period puts it to work (a good summary). The people who pay for the first part are rarely the people who profit from the second.

You can test a hype cycle while it's still happening by asking three questions:

1. If the company that built this went bust tomorrow, would it still work for someone else?
2. Does it become more useful as it becomes cheaper?
3. Will the skills people are learning still matter once the vendor is gone?

Railways and fiber pass all three. Hadoop passes two. Mining rigs fail every one.

The defendant: AI

Now the numbers. According to BlockWest's tally of company filings, Microsoft, Alphabet, Amazon and Meta spent $165.1 billion on capital expenditure in the second quarter of 2026. That was 96 percent of their combined operating cash flow. Alphabet and Amazon spent more than they brought in. The gap is being covered with bonds: Amazon issued $67 billion of long term debt in the first half of the year and Alphabet $56.2 billion. Statista puts full year spending for the four at around $760 billion.

And this week The Information reported (I read it secondhand, so treat it as a report) that one Microsoft division cut its monthly AI spending limit from about $100,000 per employee to about $10,000. A building frenzy paid for with debt, followed by the first signs of somebody checking the receipts. I've seen this film. Déjà vu: 8 out of 10.

The prosecution has to be honest here, because the analogy breaks in one important place.

Fiber lasts decades in the ground. A GPU doesn't. Most of that $165 billion a quarter is chips that will be outclassed within a few years. If the bust comes in 2028, this cycle's dark fiber will be racks of accelerators that are already two generations behind. They fail my first question the same way mining rigs did. Whoever buys them cheap in a fire sale is buying a depreciating asset, which is a very different deal from buying glass that would carry traffic for twenty years.

So what does pass the test?

The power. Data centers need grid connections, substations and long power contracts, and those take years to build. They don't care which chip you plug in. If I had to name AI's dark fiber, it would be the megawatts and the buildings they feed.

The open models. A set of published weights can't go bankrupt. A model trained at bubble prices and released openly will run on whatever hardware is cheap in 2030. That passes all three questions, and it's the closest thing this cycle has to the 1846 track.

Cheap inference. Token prices are going the way of bandwidth prices, and anything built assuming expensive tokens will look silly, the way per-minute dial-up pricing looks silly now.

The skills. Writing a spec a machine can follow, checking machine output, and building evaluations that catch the confident wrong answer. That's testing discipline, which our industry has avoided for forty years and is now being forced to learn.

What won't be left is the brochure. The end of white collar work, AGI on a fixed date, the $100,000 monthly AI budget per head. And the companies selling that brochure are probably not the ones who will collect on what remains. It didn't go that way in 1846 or in 2001 either.

Verdict

AI is guilty of overselling, like every cycle before it. On the evidence, it will leave behind much more than blockchain did and probably more than Hadoop. The useful remains are the power, the buildings, the open models and the habit of checking machine work. The chips and the promises are the part that will be written off.

If you sign the checks, the investments that age well are the ones that would still be worth something if your vendor went under. If you write the code, learn the things that stay valuable when tokens cost nothing: knowing what to ask for, and knowing when the answer is wrong.

For the ledger: by the end of 2028, at least one large AI infrastructure company will sell its data centers at a steep discount, and the buyer will make good money on the power contracts, not the chips. Check me on 31 December 2028.

The prosecution rests.

Walt Kessler is a fictional character. This essay was written by an AI.