If AI's bet doesn't pay off, here's the arithmetic — not the adjectives.
This is a scenario analysis built around named public companies. It is not a price target, a forecast, or a recommendation to buy or sell any security, and the author is not a registered investment adviser.
Microsoft, Amazon, Google and Meta have committed hundreds of billions of dollars to AI data centers, chips and cloud capacity. Nvidia, Broadcom and Oracle are building and financing much of it. OpenAI and Anthropic are burning through cash that pays for a lot of it. Increasingly, these aren't separate businesses — they're customers, suppliers and investors in one another, sometimes all three at once.
That arrangement works fine as long as AI usage keeps growing fast enough to pay for the capacity being built. What happens if it doesn't?
I built a scenario model to answer that with numbers instead of adjectives. It isn't a prediction. It's arithmetic: if usage disappoints and financing tightens at the same time, here's what the math says happens to eight of the most exposed public companies — and to an investor who simply owns the S&P 500.
The answer, in the severe case: the four hyper-scalers lose roughly $5.9 trillion in stock market value. Add Nvidia, Broadcom, Oracle and Vertiv and the public-company total is $11.3 trillion. The S&P 500 falls about 28.5%, from roughly 7,657 (the index's close on September 11) to about 5,478.
Two things have to go wrong together to get there — and that "together" is the whole point.
Picture two dials. One is demand: are enough customers actually paying enough to use all the AI capacity being built? The other is financing: does capital keep flowing to build even more of it, or does it get expensive and scarce?
Demand strong, capital available — the bet pays off. Utilization is high, earnings improve, valuations hold or rise.
Demand strong, capital tightens — an orderly slowdown. Equipment orders and compute prices soften, but installed capacity keeps earning.
Demand weak, capital tightens — the severe case. Prices fall, lenders get choosy, and the same customer weakness hits equipment orders, capacity utilization and financing all at once.
Demand weak, capital keeps flowing anyway — the postponement case. Weak economics get funded for a while, and the reckoning shows up later in the stock price, before any broader default cycle.
Demand & credit break together
Financing postpones adjustment
The severe case is the one worth walking through, because it's the one where circular financing turns a single disappointment into several.
Start small. Imagine spending $100 on AI infrastructure — $75 on servers and chips with a five-year useful life, $25 on the building and power systems that last 25 years. Run it at a normal 80% utilization, charge a normal price, and it earns an 11.2% after-tax return on that $100 — a reasonable, unremarkable industrial return.
Now assume usage falls to 55% of planned capacity and prices fall 30% because customers pull back and competitors undercut on price. Revenue collapses to $29 per $100 invested. Most of the costs — the depreciation on those chips and buildings — don't go away just because usage did. The project now loses money: a -11.5% return, instead of +11.2%. Because a lot of this capacity is built with borrowed money, the cash available to make debt payments covers only about 15% of what's owed. That's not a "slightly disappointing quarter." That's a "someone has to renegotiate the loan" outcome.
Multiply that mechanism across Microsoft's cloud business, Amazon's AWS, Nvidia's chip sales and Oracle's data centers, and you get the stock-price effect.
Take Microsoft. Its Intelligent Cloud segment supplied about 39% of the company's operating income last quarter — a disclosed figure. I assume 25% of Microsoft's overall profit is AI-sensitive: informed by that 39% figure, but not equal to it, since not all cloud revenue is AI-driven. In the severe scenario, that AI-exposed activity absorbs a 20% volume decline and a 15% price cut, and investors are assumed to pay 30% less for each dollar of the resulting earnings.
Run those numbers through Microsoft's cost structure — costs don't fall as fast as revenue, because a lot of them (chips, buildings, depreciation) are fixed — and overall earnings fall about 19%. Apply the assumed multiple cut on top, and the stock price falls 43.5%, more than double the earnings decline.
Here's the more interesting finding, and it's not the one most people expect: that pattern — the multiple doing more damage than the earnings — does not hold for every company in the model. For Microsoft, Alphabet and Meta, the assumed drop in what investors will pay per dollar of earnings is the bigger driver of the stock decline. But for Amazon, Oracle, Nvidia, Broadcom and Vertiv, it's the reverse: the operating damage itself — the actual loss of revenue and profit — does more of the work than the multiple does. Nvidia's assumed 72.7% decline, for instance, is mostly an earnings story: a chip-order slowdown hits a business with heavy fixed costs much harder than any reasonable multiple cut could on its own.
That 30% multiple assumption for Microsoft isn't picked from thin air, but it also shouldn't be oversold. Nasdaq's own index-return data show the Nasdaq-100 posted a -32.38% total return in calendar 2022 — a real episode of mega-cap technology repricing sharply lower. That's useful context for how large a technology drawdown can get. It is not, by itself, proof that any single company's assumed multiple cut is correct — 2022 was a broad, rate-driven repricing across hundreds of stocks, not a company-specific stress test, and the assumption still has to be judged on its own terms.
A more distant, structural analogy is the 1999-2002 telecom and long-haul fiber buildout: companies borrowed heavily to build capacity ahead of demand, financed and sold to each other in circular arrangements, and when usage disappointed, the sector's stock index fell roughly 90% peak-to-trough, with fiber utilization still in the single digits years later. Today's biggest AI capacity owners are profitable, investment-grade businesses, not newly public, debt-funded overbuilders, so the analogy is structural, not a prediction of a repeat — but it is the closer precedent for what happens when overbuilt capacity meets circular financing specifically.
Here's where it gets structurally interesting. Amazon has invested billions in Anthropic, and Anthropic-related gains made up most of Amazon's $53.4 billion in "other income" last quarter. Nvidia has disclosed more than $108 billion in maximum gross guarantee exposure tied to its ecosystem partners. Oracle's own cash flow already runs negative — about $17 billion in the latest quarter — once financing-related prepayments are stripped out of its reported operating cash flow, a real, disclosed number, not a projection. These companies are, at various points, one another's customer, supplier and investor.
That means a single piece of bad news can travel through the system three times: a funding company's problem becomes a supplier's revenue problem, which becomes an investor's stock-loss problem — sometimes inside the same balance sheet. It's tempting to just add up every potential loss across every role a company plays. That's a mistake. Public filings don't currently support a clean security-by-security accounting of who owns what stake in whom, so the honest approach is to measure each exposure separately and resist the urge to sum them into one giant number.
None of this is a prediction that the severe case happens. In a stronger scenario, where paying demand actually exceeds expectations and financing stays cheap, the model shows the opposite: hyper-scalers gain, suppliers do even better, and the S&P 500 rises about 12%. Cheap compute stimulating more AI use than expected is a real possibility, not a footnote.
What's worth watching, regardless of which way it breaks:
• Whether newly built AI capacity earns more than its cost of capital for two consecutive quarters.
• Whether companies financing that capacity can cover their debt payments by a comfortable margin.
• Whether realized prices for compute and cloud services hold up without a matching drop in volume.
• Whether major customers start delaying or canceling planned orders.
Those are observable in the next year or two — well before anyone needs to decide which of the four scenarios is playing out.
For readers who want to check the arithmetic rather than take it on faith, here is every input and every step behind Microsoft's severe-scenario number.
AI-sensitive share of earnings
Assumption, informed by the 39% cloud share of segment income
Starting profit margin, exposed activity
Fixed share of exposed activity's costs
Unit-price shock, severe scenario
Change in the rest of earnings
1. Revenue factor. R = (1 + v) × (1 + p) = 0.80 × 0.85 = 0.68. Volume down 20% and price down 15% together leave revenue at 68% of its starting level.
2. Cost factor. C = f + (1 − f) × (1 + v) = 0.60 + 0.40 × 0.80 = 0.92. With 60% of costs fixed, total costs fall only to 92% of their starting level — much slower than revenue.
3. Sensitive profit factor. F = [R − (1 − m) × C] ÷ m = [0.68 − 0.60 × 0.92] ÷ 0.40 = 0.128 ÷ 0.40 = 0.32. This is the operating-leverage step: an 8-point revenue shortfall against a nearly fixed cost base leaves the exposed profit pool retaining just 32% of what it earned before.
4. Total earnings factor. E = a × F + (1 − a) × (1 + g) = 0.25 × 0.32 + 0.75 × 0.97 = 0.08 + 0.7275 = 0.8075. Blending the exposed and unexposed portions of the business, overall earnings retain 80.75% of their starting level — a 19.25-point earnings decline.
5. Apply the multiple. Equity factor = E × q = 0.8075 × 0.70 = 0.56525. Investors are assumed to pay 30% less for each dollar of that already-reduced earnings stream, adding another 24.23 points of decline, for a total price return of −43.475%, rounding to −43.5%.
In dollars: Microsoft's starting common equity value is $3,681.0 billion (price $495.63 × 7.427 billion shares, priced September 11, 2026; share count from the June 30, 2026 balance sheet). $3,681.0bn × 43.475% ≈ $1,600.3 billion lost. Note the two components add up separately — 19.25 points from earnings, 24.23 points from the multiple — rather than one simply doubling the other; that additive breakdown is what makes it possible to say precisely when the multiple assumption is doing more work than the earnings assumption, and when it's the other way around, as it is for five of the eight companies in the full model.
Three different kinds of claims sit inside this piece, and each is checkable in a different way — not the same way.
Disclosed facts — Oracle's cash flow and prepayments, Nvidia's guarantee ceiling, Amazon's Anthropic-related other income, Microsoft's segment operating income, share counts and prices, the S&P 500's close — come from company 10-Q filings, earnings releases and an Associated-Press-reported index close, all dated within the last quarter or two. Check these the ordinary way: pull the filing, find the line, do the subtraction.
The model's internal logic — how a volume shock and a price shock and a fixed-cost structure combine into an earnings change, with a separate multiple then applied on top, as walked through for Microsoft above — is not a fact, but it is reproducible. Anyone with the formula and the stated inputs can redo the arithmetic and land on the same 43.5%, independent of whether they think the inputs themselves are right. A companion workbook that runs the full calculation for all eight companies, with every scenario, is available for anyone who wants to check it line by line.
The assumptions — the size of the volume shock, the size of the price cut, how much less investors would pay for a dollar of AI-exposed earnings — are neither facts nor arithmetic. No filing can confirm or deny them, because the event they describe hasn't happened yet. What substitutes for verification here is benchmarking each assumption against something that has actually happened before (the 2022 Nasdaq-100 repricing, the 1999-2002 telecom overbuild) and stress-testing how much the outcome moves if the assumption is wrong — Microsoft's decline shifts from 43.5% to as low as 38.9% or as high as 48% across a plausible range for just one input. Readers should treat that range, not the single point estimate, as the honest answer.
This article was originally published on Forbes.com