Two red days and a 20% fine on a perfect quarter, and the mailbag fills with one question: is it over? Here's this desk's answer, argued with receipts: the AI bull market has a minimum of two years left, and the interesting question was never "when does it end." It's "where does the money go next." Because it moves. It always moves.

The arms dealers get paid first, the landlords second, the factories last — and we can already see the deed transfers being signed. Lily Caruso, on the game that hasn't actually started yet. By Lily Caruso · TrendyVest Deep Analysis · August 18, 2026, evening Every few days someone writes to this desk and asks what inning we're in.

I understand the question. It's a good question, asked by people who lived through 2000 and 2021 and have the scar tissue to prove it, and after a week like this one — memory down 8%, a perfect quarter fined a fifth of its market cap, yields at pre-crisis highs — it deserves better than a slogan. So here's my honest answer, and then fourteen hundred words of showing my work: you're asking about innings while they're still pouring the concrete for the stadium.

The game you're worried about ending hasn't started. What you've been watching for two years is construction. Follow the concrete.

Let me give you the numbers that reframed this for me when I laid them side by side this week. The five largest American cloud and AI companies are planning to spend roughly $660 to $690 billion on infrastructure in 2026. Not over a decade — next year alone.

Amazon around $200 billion. Alphabet as much as $185 billion. Microsoft $120 billion and up, Meta up to $135 billion, Oracle $50 billion.

That's nearly double what the same five spent in 2025, which was itself nearly double the year before. Beyond them, Stargate targets half a trillion dollars by 2029. The Ohio campus Nvidia backstopped on Monday doesn't switch on its first phase until 2028.

SanDisk's supply contracts run through fiscal 2028 with two-thirds of its production already spoken for. Nvidia signed a twenty-year lease guarantee this week. Read those dates again, because they're the argument.

The people writing the largest checks in the history of corporate capitalism have contractually committed themselves to a buildout whose first phases come online in 2028. When someone asks me whether the AI trade has two years left, I point at the paper: the next two years are already signed for. You can debate what happens in 2029.

The spine of 2027 and 2028 is under contract, with penalties. Now the uncomfortable chart, because you deserve the bear's best card too. Against that $660-plus billion of annual spending, the pure-play AI model companies — OpenAI, Anthropic, and everyone smaller — are projected to generate less than $35 billion in combined 2026 revenue by one careful estimate.

I'll be straight with you: the estimates disagree wildly. One market wrap this week credited Anthropic with $11.5 billion in a single quarter, triple what the cautious models assume for its year. When the scorekeepers disagree by multiples, that itself tells you how early in the game the scoring is.

But take the conservative number. Spending twenty dollars to earn one is either the stupidest thing capitalism has ever done, or it's what building a railroad looks like from the middle. The way you tell the difference is the backlogs — is anyone actually lined up to use this stuff?

Alphabet's cloud backlog just grew 55% in a single quarter, to $240 billion. Microsoft reports $80 billion of Azure demand it cannot fulfill — not for lack of customers, for lack of electricity. When your problem is that demand exceeds your ability to generate power, you are not in a demand bubble.

You are early. The pattern that repeats every time humans build a general technology. Here's the part I find beautiful, and it's the actual thesis of this piece.

The money in a buildout doesn't stay where it starts. It migrates, in the same order, every single time. The railroads: first the money went to the men selling rails and locomotives; then to the operators who owned the track; and finally — for decades — to everyone whose farm, factory, and town the track connected.

Electricity: the dynamo makers got rich first, the utilities second, and then the real fortune, the one that dwarfed the others, went to the factories that reorganized themselves around the electric motor — a process that took twenty years and rewarded the patient. The internet: Cisco was the best stock on earth while the fiber went in; then the platforms that sat on the fiber — the hyperscalers themselves, as it happens — collected for two decades and are collecting still. The arms dealers.

Then the landlords. Then the users. Three payments, in order, every time.

Map it onto today. Phase one is what you've owned for two years: Nvidia, the memory complex, the optics makers — the arms dealers of the stack, the sellers of rails. That trade isn't over; Fabrinet just grew 45% and Nvidia's audit is next Wednesday.

But it is maturing, and this week showed you the symptoms of maturity: verticals that breathe, fines for decoration, a market demanding signed contracts instead of stories. Phase two is beginning under your feet, and the backlogs are its birth certificate: the hyperscalers — the landlords — are converting the buildout into rent. A $240 billion cloud backlog is rent receivable.

An $80 billion power-constrained queue is a landlord with a waiting list. The money is already migrating from the people who sell the shovels to the people who own the land, and the crowd is still staring at the shovel store. And phase three — the longest, largest, slowest payment — goes to the users: the manufacturers, the industrial companies, the physical economy that reorganizes itself around machine intelligence the way factories once reorganized around the electric motor.

Want to know how early that phase is? Twenty-nine percent of manufacturers report AI in actual facility-level use. Only a third of all enterprises have scaled anything beyond a pilot.

Two-thirds of the economy hasn't taken the field. That's not an aging bull. That's a stadium with the seats still wrapped in plastic.

So: two years minimum. Here's the number and the wager. This desk's Burst Cycle models have carried 70% odds of at least eighteen more months since August 14, with published tripwires.

Tonight, this analysis extends the argument on structural grounds: I put the odds that the AI-led bull runs at least another twenty-four months at 65% — slightly below my colleague's shorter-horizon number, because honesty compounds and two years is a longer walk. The case rests on three legs. The spending is contractual through 2028, not sentimental.

The demand is supply-constrained, not imagined — you cannot fake a power shortage. And the migration has two full phases left to pay, with the biggest checks — the hyperscalers' rent collection and the manufacturers' reorganization — still substantially unwritten. Expect the bursts and the pullbacks both; the immune system that fined Fabrinet on Tuesday is the same one that keeps this organism alive.

Higher highs, ugly valleys, and a rotation your grandchildren will read about. What would make me wrong, stated now, in writing. Four things, and I'll grade them in public like everything else here.

If the pure-play model companies' revenue hasn't at least tripled by the end of 2027, the payback math starts failing and the landlords' rent never arrives — watch that gap close or widen like your portfolio depends on it, because it does. If the cloud backlogs stop growing — two consecutive quarters of contraction at two majors is the desk's standing tripwire — the migration stalls at phase one and this piece ages badly. If the 30-year Treasury breaks through 5.5% and stays there, the Margin Loan financing this entire buildout reprices, and my colleague's warning that the AI trade now funds its own opposition becomes the story of 2027.

And if the manufacturers simply don't adopt — if that 29% is a ceiling instead of a floor — then phase three was a theory, and the whole structure ends the way railroads ended in 1873 instead of the way electricity ended in prosperity. I don't believe any of these fire. I've written them down so you can check.

The honest coda, because Lily's column doesn't do victory laps in advance: this desk is on the record as bullish, with graded calls that include this week's visible miss, and you should weight our two-year confidence accordingly — conviction disclosed is conviction discounted. But when I sit with the numbers on a Tuesday night — the signed capex, the power queues, the two-thirds of the economy still in the dugout — the conclusion won't move for me. The question was never what inning it is.

They're still pouring the concrete. Buy your seat before the plastic comes off. Tickers in play: NVDA · MSFT · GOOGL · AMZN · META · ORCL · MU · SNDK · FN · MRVL · CAT · ETN This is TrendyVest's analysis and opinion — a deep analysis, for informational purposes only, Do your own research.

Markets. Tech. The Edge.

Research with receipts.