The $655 Billion Bet: AI Capex, Circular Deals, and What Breaks First

In my stablecoin piece I teased the AI trade as "crowded and priced for perfection." That deserves more than a throwaway line, because the numbers involved have stopped being tech-budget numbers and started being sovereign-budget numbers.

Four companies—Amazon, Alphabet, Meta, Microsoft—have guided to roughly $655 billion of capital expenditure for 2026. For scale: that's more than Italy's entire annual state budget, deployed by four boards, mostly into GPUs and buildings to put them in.

The question isn't whether this is impressive. It's whether it's rational—and what it does to your portfolio either way.

The Spend Is Real

Start with what's verifiable: the guidance.

FIG. 01 // CAPEX GUIDANCE 2026

Big-four hyperscaler capital expenditure, 2026 guidance

$655B

≈ +36% vs 2025 aggregate · ~75% AI infrastructure

DATA: COMPANY GUIDANCE · CREDITSIGHTS · TOM'S HARDWARE · JUL 2026

Amazon has guided ~$200 billion, more than doubling its 2025 outlay. Alphabet raised guidance to as much as $190 billion. Meta is at up to $145 billion, Microsoft tracking above $120 billion for its fiscal year. CreditSights estimates roughly 75% of aggregate hyperscaler capex funds AI-specific infrastructure.

Two things can be true at once: this is the largest private infrastructure build-out in history, and the companies doing it generate enough operating cash flow to afford it. This is the key difference from 1999—Cisco's customers were debt-financed startups; Nvidia's biggest customers print $100B+ of free cash flow a year.

Mostly. Which brings us to the uncomfortable part.

The Loop

The marginal dollar of AI "demand" is increasingly hard to distinguish from the marginal dollar of AI investment, because the same names sit on both sides of the trade.

FIG. 02 // THE LOOP

Circular deals: everyone is everyone's customer and investor

Capital flowsProduct / revenue flows

SOURCES: BLOOMBERG · COMPANY ANNOUNCEMENTS · 2025-2026

Follow the arrows: Nvidia commits $100 billion to OpenAI—which spends much of it on Nvidia GPUs. Oracle's Stargate program builds ~$300 billion of compute for OpenAI, on Nvidia hardware. Nvidia holds a stake in CoreWeave, which has a cloud deal with Oracle. Every hop recognizes revenue; every participant's stock re-rates on that revenue; the re-rated stock funds the next commitment.

This is not fraud—every deal is disclosed. But it's the same capital being counted more than once as "demand," and it makes the reported growth rates of everyone in the loop partially endogenous. OpenAI itself went from an $80 billion valuation in 2023 to ~$730 billion by early 2026 while projecting $140 billion of cumulative operating losses through 2029. The loop only works while the equity keeps re-rating.

The Concentration

Meanwhile, here's what the index you're benchmarked to actually looks like:

FIG. 03 // CONCENTRATION

How much of the market is one trade

Top-5 stocks' weight in the S&P 500

30%

highest concentration in ~50 years

Share of index returns driven by AI stocks since 2022

75%

the index is the AI trade

Nvidia market cap, 2026 peak

$4.5T

larger than every stock market outside the US and Japan

DATA: S&P GLOBAL · VANDERBILT/NBER ANALYSES · 2026

Top five stocks: 30% of the S&P 500, the highest concentration in roughly half a century. AI-related stocks have driven ~75% of index returns since 2022. And the fundamental case under it is still mostly projection: an NBER study from February 2026 found 90% of firms report no measurable productivity impact from AI so far—while their executives keep projecting gains.

If you hold a passive index fund, you are not diversified. You are long one theme with extra steps.

What a Quant Does With This

I don't find "bubble: yes/no" a useful research question. These are:

1. Dispersion is cheap relative to the risk. With 30% of the index in five names, index vol understates single-name vol. Correlation breaks in either direction—melt-up or derating—pay off dispersion structures.

2. Watch the financing channel, not the narrative. The equity story is loud; the debt story is quiet. Data-center construction is increasingly financed in private credit and ABS markets. When AI capex stops being self-funded from operating cash flow, the regime has changed—that's measurable in filings, not vibes.

3. Depreciation is the slow fuse. $655B/year of hardware with 3-5 year useful life converts into a permanent, growing D&A drag. Earnings quality at the hyperscalers will degrade mechanically even if revenue lands. Model it; most sell-side numbers don't.

4. The loop is a contagion map. The circular-deal diagram above doubles as a stress-propagation graph: if OpenAI's equity stalls, the impairments land at Nvidia, Oracle and CoreWeave simultaneously. Position sizing should treat them as one correlated exposure, not four.

The honest assessment: the cash flows funding this build-out are real, which is why the bubble hasn't popped on schedule for three years. But circular deals, projected-not-realized productivity, record concentration and a mechanical depreciation wave are each individually survivable and collectively fragile. You don't need to predict the pin. You need to not be the most crowded person in the room when it finds one.

---

Are you positioned in or against the AI trade—and what's your tell for when the regime turns? The sell-side tells one story; the trenches tell another.