I promised to show you why the whole industry’s answer to its own problem — buy a bigger brain — is the most expensive mistake in the history of computing. To do that I have to take you back to 1999, because I was there, and if you’re old enough to be reading me, maybe you were too.

And I wasn’t only watching. In 1999 I put $10,000 into a young company called E-Loan, run by a founder named Chris Larsen. After the IPO I cashed out for $400,000 and bought a house. Chris kept playing — E-Loan to Prosper to Ripple — and did rather better than a house; he’s a crypto billionaire now. (Chris, if you’re reading this: we should talk.) Those are the two ways to play a year like 1999, and both of them worked. But notice which one of us turned the paper into something you can actually live in.

You remember the frenzy. The IPOs. The companies with no earnings yet a story so good that earnings felt like a rude question. The fear, thick in the air, that if you didn’t get in now you’d spend the rest of your life explaining to your grandchildren why you missed it. We had a name for it — Fear of Missing Out —  before we had the acronym: you bought because everyone was buying, and the buying was the proof.

Then it ended, and everyone drew exactly the wrong lesson.

The lesson people took from 1999 was that the internet was a bubble. It wasn’t. The internet was the most real thing to happen to commerce in a century — it ate retail, media, advertising, and the telephone, precisely as the prospectuses promised. The technology was never the lie. The lie was the price — capital sprinting into a true story while refusing to look at the unit economics underneath it. Pets.com wasn’t wrong that you’d someday buy dog food online. You’re probably buying it online right now. Pets.com was wrong about what it cost to ship a forty-pound bag for a flat five bucks. The revolution was real. The arithmetic was fatal.

As I write this, SpaceX is days from launching what will be the largest IPO in human history — a roadshow this week, pricing next, at a valuation knocking on two trillion dollars, larger than Aramco. OpenAI is lining up a listing for the fall at something near a trillion. Anthropic has quietly filed, valued in the same impossible neighborhood. Last quarter, roughly four of every five venture dollars on the planet went to AI. Bank of America’s own strategist is comparing the mood to the most extravagant manias on record and warning that these three debuts alone will tip the market’s concentration past anything we saw in the dot-com peak. The FOMO isn’t a side effect this time. It’s the product.

And — let me say it as plainly as I said it about the internet — the technology is real. That is not the question. AI is going to be as consequential as its loudest believers claim. So stop arguing about whether it’s a bubble in the lazy way, the “is any of this real” way. That was the wrong question in 1999 and it’s the wrong question now.

The right question is the one nobody at the roadshow wants asked: is the money priced on the right architecture?

It isn’t. And here’s where my last two columns come collect their debt.

The story being sold — the story holding up those valuations — is that AI is thinking, that thinking demands an ever-larger brain, and that whoever owns the biggest brain owns the next century. Every dollar of that two-trillion-dollar valuation assumes the moat is brain size. But Salesforce’s own researchers already ran the experiment, and I told you the number last time: hand the model the documents and it scores in the mid-seventies; make it go find them and it scores a third. The brain was never the bottleneck. The bottleneck was retrieval — the cheap, dull, unglamorous business of locating the right paragraph. The expensive part was sitting there the whole time, perfectly capable, waiting on the cheap part.

So follow the money to where it’s actually going. Stargate alone is a half-trillion-dollar bet on gigawatts of data centers. The grid can’t keep up; we’re now siting power plants for this. And the overwhelming bulk of what all that silicon will do, in the enterprise, is look things up — retrieval, the work that runs on a CPU at a few watts, being run instead on GPUs at three hundred. That is the NVIDIA tax I described in my first column, and now you can see its size. It is the largest concentration of capital in the history of business, aimed at making the brain bigger, to solve a problem that does not live in the brain.

That’s the mistake. Not that AI is fake — that the spend is pointed at the wrong layer. And it is the most expensive mistake ever made because we have never before had the means to make a wrong call at this scale, this fast, with this much applause.

Which makes a two-trillion-dollar valuation built on “biggest brain wins” a Pets.com valuation. Right about the revolution. Wrong about the cost structure underneath it. Bank of America, in less inflammatory language than mine, has already described this IPO wave for what it is: a mechanism for moving accumulated risk off the early insiders and onto you. The roadshow is the machine that does it. A roadshow is, and has always been, an apparatus for manufacturing urgency — get in before the story is re-examined. In 1999 the re-examination arrived a few quarters after the bell, when the lockups expired and someone finally totaled up what it cost to deliver the product. I have watched this exact movie. I can tell you how it ends. It ends the day the unit economics walk into the room.

A disclosure, as always

You should know I’m not a neutral party. I co-founded a small company built on the heresy in these columns — that retrieval is the main event, that it belongs on cheap and abundant silicon, and that the giant brain should be held in reserve for the rare moment something must truly be generated rather than found. So discount my enthusiasm to taste.

But notice what my conflict cannot touch. The two-trillion-dollar number is SpaceX’s, not mine. The eighty-percent figure is the venture industry’s, not mine. The bubble warning is Bank of America’s, not mine. And the score that proves the brain was never the bottleneck — that’s Salesforce’s, published in their own paper, downloadable by anyone who wants to check my arithmetic. I’m not asking you to trust me. I’m asking you to read the documents the believers wrote themselves.

There’s a second mistake hiding underneath the first one, and it’s worse — because the bigger brain can’t fix the economics, which is bad, but it also can’t fix the one thing every serious enterprise actually needs: a machine that knows when to say I don’t know. That’s where “expensive” quietly turns into “dangerous,” and that’s where I’ll take you next time.

For now, just remember what 1999 actually taught us. The companies that walked out of the wreckage weren’t the ones with the biggest story or the loudest roadshow. They were the boring ones whose arithmetic still worked the morning after the party. Everyone else is at the roadshow. The opportunity, as always, is for whoever’s in the back of the room doing the math. I’ve run that math once before. It bought me a house.