To The Moon

AI Spending Isn't Peaking. Its Growth Is.

Yesterday we looked at AMD into its print. Here is the spending wave it rides on.

$201.9B. That is what Amazon, Microsoft, Alphabet and Meta spent in cash on capex in one quarter, Q3 2026.

The usual question is whether that number has peaked. I think it is the wrong question. Spending and the growth of spending peak at different times, and markets have tended to care about the second one.

TWO PEAKS, NOT ONE

A boom has a level (dollars spent) and a growth rate (how fast those dollars rise). The growth rate can top out years before the level does.

Goldman Sachs (24 Sep) noted that consensus has the big US cloud spenders at about $800B of capex in 2026 and about $1.1T in 2027, with growth roughly halving each year.

UBS lifted its own forecasts in late October and expects growth to normalise in 2028.

The two use different company sets, so I am not adding them together. What they share is the shape: dollars still rising, growth slowing.

For the four biggest US spenders alone, capex rose from $376B in 2025 to an estimated $711B in 2026, up 89%. That would make 2026 the growth peak, if estimates hold.

Analysts see $0.9-1.1T in 2027, so growth slows even as dollars keep climbing.

[CHART: mm_capex_B_1nov_tiger_v1.png]

WHAT PAST BOOMS DID

Shale is the slow version. US oil and gas well investment grew fastest in 2011, at +31.0%. Dollars kept climbing for three more years, to a 2014 high about 50% above 2011.

Then it fell 62% from 2014 to 2016, after the oil price broke. The energy producers' ETF, XOP, rose 58% from end-2011 to its June 2014 peak (price basis), then fell 72% by February 2016 (closing basis).

Telecom is the fast version. US communication equipment investment grew fastest in 2000, at +24.9%, the same year it peaked in dollars. It was down 32% by 2002.

The Nasdaq fell 78% from its March 2000 closing high to October 2002. There was no gap between the two peaks, and no warning.

Memory is the closest match to today. Micron's capex growth peaked in FY2018, at +87.6%. Its spending in dollars peaked a year later, in FY2019.

The stock did not wait. Micron closed at its high on 29 May 2018 and fell 54% by 24 December. The chip index fell 26% over a similar stretch.

So the gap between the two peaks ran from zero to three years. In two of the three cases, stocks topped near the growth peak, not the spend peak.

[CHART: mm_pastbooms_D_1nov_tiger_v1.png]

WHAT IS DIFFERENT THIS TIME

About $1.0T of signed leases at Microsoft, Meta, Alphabet and Amazon have not started yet. They sit outside the cash capex figure above.

Only Microsoft's and Meta's, about $720B together, run as late as 2034-36. Alphabet's start by 2031, and Amazon gives no end year for the bulk of its share.

Neither telecom in 1999 nor shale in 2013 had a contracted backlog like that. Contracts can still be repriced, so this is a cushion, not a guarantee.

There is also a fifth spender. SpaceX's AI segment spent $16.4B on capex in Q3, so SpaceX now spends like a hyperscaler. It guides $35-40B of total company capex for Q4, not AI alone.

MEMORY: THE LAYER I THINK IS UNDER-PRICED

Here is what the three memory makers say about their own order books. SK hynix said on 28 Oct that its DRAM and HBM are effectively sold out through 2029.

Samsung said on 29 Oct that its 2027 HBM is already sold out. Micron said on 30 Sep that the vast majority of its 2027 HBM supply is agreed.

Yet Micron trades on a single-digit forward P/E. Data vendors disagree on the exact multiple, so I am not printing one.

My hypothesis: the market is pricing memory as if 2018 repeats, a growth peak followed by a fast fall. The suppliers' order books do not show that cliff yet. That is my read, not a forecast, and a low multiple is not the same as cheap.

THE COUNTER-CASE

Memory has looked inexpensive at past peaks, because earnings were at their peak too. In 2018 the low multiple was the warning, not the opportunity.

"Sold out" contracts can be renegotiated if demand slips. And telecom shows a growth peak can arrive with no years of rising spend behind it.

What would change my mind: 2027 capex plans that fall in dollars, not just in growth, or a memory maker walking back a sold-out statement.

THE DATED TEST

Based on past timing, Microsoft and Meta typically report in the last week of January, Alphabet and Amazon in the first week of February (dates not yet confirmed). Those Q4 2026 reports are when 2027 plans get firmer.

If 2027 dollars still rise while growth slows, the two-peak reading holds. If dollars fall, it was the telecom shape, and I was wrong. I will score it either way.

HOW TO RUN THIS ON ANY STOCK

1. Find the capex series that drives the stock's revenue, in dollars, for several years.

2. Work out the yearly growth rate, and mark the year growth peaked and the year dollars peaked.

3. Check where the stock topped against those two dates, in past cycles of the same industry.

A hypothetical example, in made-up units: spending goes 100, 190, 260, 300. Dollars rise every year, but growth goes +90%, +37%, +15%. A market that prices growth can fall while spending is still rising.

When you hold a supplier to a spending boom, do you track the dollars your customers spend, or how fast those dollars are growing?

Not advice. I may hold positions in the names mentioned.

Stop guessing. Start calculating.

Live to fight another day. 🤙

#options #tradingpsychology #riskmanagement #optionstrading