New

The Ai Boom Has A Warning System. Here's What It's Saying Now

Every great boom ends. The railroads did. Radio did. The dot-coms did. Someday, the AI buildout will too.

The trouble is that nobody rings a bell at the top. Most investors learn a boom has peaked the hard way — months later, after the gains are gone.

I'd rather not learn it that way. So I built a warning system. I call it the Peak Watch Checklist: 13 specific, trackable signals that have historically flashed before tech booms turn into busts.

Right now, seven of those 13 signals are lit. That's not a sell signal, but it's no longer a green light, either. The AI boom is in its late innings, and it's time to start watching it the way a pilot watches the instrument panel on approach.

Why a peak checklist, and why now

The AI buildout is the biggest corporate spending spree in modern history. The five largest cloud companies are on track to spend more than $750 billion on AI infrastructure this year, with over $1 trillion expected in 2027.

For most of this boom, that spending was paid for with cash. Microsoft (MSFT), Alphabet (GOOGL), Amazon (AMZN) and Meta (META) are some of the most profitable companies on earth, and they could fund chips and data centers out of pocket.

That's changing. Spending has now outrun the cash these companies generate, so they're borrowing to fill the gap. Hyperscaler bond sales hit $159 billion by mid-year, already more than all of 2025. Alphabet sold a record $85 billion of new stock.

When a boom shifts from being paid for with profits to being paid for with borrowed money, the rules change. Lenders start to matter as much as customers. And lenders get nervous faster than customers do.

We got a taste of that in mid-October. Bonds issued by CoreWeave (CRWV), a company that rents out AI chips to tech giants, sank until they yielded around 15%, up from roughly 10% this summer. That's the kind of yield usually reserved for companies in real trouble.

The fear spread fast. The Philadelphia Semiconductor Index, or SOXX, fell 11.4% in the week ending October 16, on an eight-day losing streak. Stocks bounced the following Monday, but the episode made the point: the AI trade now has a credit-market pressure point it didn't have a year ago.

That's exactly why you need a checklist. In a moment like this, headlines swing from "bubble!" to "buy the dip!" by the hour. A checklist forces you to ask a better question: which specific things would have to go wrong for this boom to be over, and have they?

How the checklist works

Booms don't die all at once. They fail in a fairly predictable order, so the checklist sorts its 13 signals into the three stages a boom passes through on the way down:

Froth (late boom). Money is cheap and plentiful, and it starts circling back on itself. This stage can last a long time, and stocks often keep rising through it.
Cracks (the top forming). Good news stops lifting stocks. Raising money gets harder. Borrowing costs climb.
Confirmation (the top is in). The real economy rolls over: chip orders get cut, data centers sit empty, projects get canceled.

Each signal gets one of four ratings:

Clear: no sign of trouble yet
Watch: early or mixed evidence
Flashing: clearly present, but short of its tripwire
Triggered: the tripwire has been hit

The tripwire is the most important part. Every signal comes with a specific, observable event that would turn it red, written down in advance. That's what keeps the checklist honest. It's easy to explain away bad news after the fact. It's much harder when you've already committed to what bad news looks like.

Finally, I roll everything into one number: the Peak Watch Score, which counts how many of the 13 signals are lit (flashing or triggered). I update it every week.

Stage 1: Froth (3 of 3 flashing)

1. Debt-funded spending: FLASHING

A healthy boom is paid for with profits. A late one is paid for with debt. Oracle (ORCL), one of the most aggressive AI builders, posted negative $5.4 billion of free cash flow last quarter on $28.5 billion of spending, and carries about $125 billion of debt. Alphabet's planned $180 billion to $190 billion of 2026 spending roughly equals all the cash its business generates.

Turns to triggered when two or more of the big cloud companies burn more cash than they bring in, or new spending is funded mainly with bonds and stock.

2. Vendor financing: FLASHING

This is when a supplier lends its customer the money to buy the supplier's own products. It inflates sales, and it was a hallmark of the late-1990s telecom bubble. Chipmaker Broadcom (AVGO) has agreed to lend AI lab Anthropic up to $42 billion to help pay for a $125 billion lease of Broadcom-designed chips. Separately, Wall Street banks and Blackstone (BX) are lining up a record $60 billion debt package to buy chips that will be leased to Anthropic.

Turns to triggered when vendor loans fund a large share of a supplier's AI revenue, or a financed customer misses a payment.

3. Record stock sales: FLASHING

Smart companies sell stock when it's expensive. This year, Alphabet's $85 billion share sale and SpaceX's $75 billion IPO both set records. Anthropic (ANT) is preparing an IPO that prospective investors reportedly value at $1.8 trillion to $2 trillion, and AI cloud company Nscale (NSCL) has filed to list on the New York Stock Exchange.

Turns to triggered when mega-deals keep coming, priced above their expected ranges, with big first-day pops.

Stage 2: Cracks (2 triggered, 2 flashing)

This is where the action is. Two of the seven crack signals have hit their tripwires, and two more are flashing. Both triggered signals are about money, not demand.

4. Spending growth slows: CLEAR

AI spending doesn't have to fall for a boom to peak. It just has to stop accelerating, because stock prices are built on growth rates. So far, it hasn't. Alphabet's quarterly spending jumped from $35.7 billion in the first quarter to $44.9 billion in the second, and Microsoft guided to more growth next fiscal year.

Turns to triggered when two or more of the big four cloud companies guide to slower spending growth in the same quarter. Their third-quarter reports, due over the next week or so, are the next test.

5. Good news gets sold: FLASHING

When stocks fall on great news, the market is telling you the good news is already in the price. On October 15, Taiwan Semiconductor (TSMC), which makes nearly every advanced AI chip, reported record profit, up 68.7%. It said AI demand still exceeds what it can supply. Its U.S. shares fell anyway.

Turns to triggered when most AI leaders sell off on strong earnings within a single reporting season. We'll know soon.

6. Revenue lags spending: WATCH

Eventually, AI has to pay for itself. Right now, the gap is wide: roughly $750 billion of annual spending, against AI revenue that is growing fast but from a much smaller base. In early October, OpenAI told investors its annualized revenue was about $50 billion, not the widely reported $68 billion, and AI chip stocks fell on the news.

Turns to triggered when the leading AI labs' revenue growth slows for two straight quarters while spending plans keep rising.

(Image credit: Getty Images)

7. Pricing power erodes: WATCH

The cost of running AI models is collapsing. A usage-weighted index of AI "token" prices, the units AI services charge by, hit a record low in September. That's great for users, and it's fine for providers as long as usage grows faster. Rental prices for the chips themselves remain firm, which says demand is keeping up for now.

Turns to triggered chip rental prices fall for a sustained stretch while token prices keep sliding.

8. AI credit spreads widen: TRIGGERED

A credit spread is the extra interest a borrower pays over safe U.S. Treasury bonds. It's how lenders price risk, and bond markets often spot trouble before stock markets do. In mid-October, the cost to insure CoreWeave's debt against default topped 10 percentage points a year. Oracle's default insurance hit a record, and so did Amazon's, Alphabet's, Meta's and Nvidia's, though theirs remain far lower.

That's stress across several AI borrowers at once, which is this signal's tripwire. To be clear, nobody thinks Amazon is going broke. But lenders are demanding more to fund the buildout, and that raises its cost.

Next level: a big cloud company has to cut or reprice a bond deal.

9. Capital raises fail: TRIGGERED - SELECTIVELY

In a true boom, every deal gets done. On October 9, Australian AI data center developer Firmus pulled its planned $5 billion IPO after weak investor demand. That's one company, not a market-wide freeze. But it's a real crack.

Goes system-wide if a second major AI infrastructure deal is pulled or slashed. The Nscale IPO and the $60 billion Anthropic chip financing are the next tests.

10. The Fed tightens: FLASHING

Historically, booms die from rising interest rates more often than from falling demand. The Federal Reserve raised rates in September for the first time since 2023. September consumer prices then rose 3.7% from a year earlier, hotter than expected, and the 10-year Treasury yield touched about 5.2% in mid-October, its highest since 2007. Fed Chair Kevin Warsh has declined to tip his hand on the October meeting, and traders have scaled back bets on a hike there.

Turns to triggered when a second hike lands while long-term rates and credit spreads rise together.

Stage 3: Confirmation (mostly clear)

This is the most important stage, and it's the good news. If the AI boom were truly ending, you'd see it here first in the real economy: fewer chip orders, emptier data centers, canceled projects. None of that is happening at scale.

11. Chip orders slow: CLEAR

The companies that make AI's picks and shovels are still raising forecasts. TSMC guided fourth-quarter revenue to $52.3 billion to $54.3 billion and said demand still outruns its supply. Memory maker Micron (MU) just posted a record $54.2 billion quarter, more than four times its year-ago revenue, and guided higher still.

Turns to triggered whenchipmakers cut guidance, customers push out orders, or inventories pile up.

12. Capacity sits idle: CLEAR

In a bust, you see empty buildings. Instead, North American data-center vacancy was just 1.4% in the first half of 2026, per real estate firm CBRE (CBRE), down from 1.6% a year earlier, even as supply grew by about a third. Everything being built is getting filled.

Turns to triggered when vacancy rises for two periods in a row, or companies start disclosing idle chip clusters.

13. Projects get canceled: WATCH

Firmus and its partner CDC Data Centres scrapped a planned 1.6-gigawatt buildout in Australia after completing only about 42 megawatts. It's the same company that pulled its IPO, so this looks like one overextended developer, not a trend.

Turns to triggered when several developers cancel projects, or a big cloud company walks away from signed leases.

Where it stands now, and how I see it from here

Add it all up and the Peak Watch Score sits at seven of 13, up from six a week earlier. My reading: late cycle, not a confirmed peak.