I’m sure that every METAR has played with spreadsheets.
Many of us have probably used spreadsheets while concocting forecasts in our jobs (e.g. of future sales, etc.).
It sure is fun to enter numbers on the top line and the bottom line and adjust numbers in between to make those numbers work, isn’t it?
https://www.wsj.com/finance/big-tech-stocks-are-pricing-in-a-miracle-on-costs-6d93664f
Big Tech Stocks Are Pricing In a Miracle on Costs
Analysts’ models are assuming new efficiencies, but some of the numbers look too good to be true
By Jonathan Weil, The Wall Street Journal, July 28, 2026
Wall Street’s forecasts for Big Tech require a major leap of faith: that the biggest AI hyperscalers can boost revenue much faster than the costs of running their businesses. Some of the numbers look too good to be true…
The bullish earnings narrative requires these companies to achieve newfound operating efficiency even as they spend trillions of dollars in capex over the next few years. The premise is evident in the consensus estimates for the five hyperscalers when combined as a group.
[snip detailed section on revenues and SGA (sales, general and administrative expenses)]
But with some of the SG&A estimates, another possible explanation is that Wall Street analysts locked in their revenue and earnings estimates based on management guidance, then reverse-engineered the expense projections so they would fit…
The expenses hit cash, not just earnings. Most of SG&A consists of cash outlays for costs such as payroll, commissions and marketing. Overruns would reduce the operating cash flow needed to fund the AI build-out, making it more likely the companies would have to issue additional debt or sell stock to cover shortfalls. Amazon and Oracle have already turned free-cash-flow negative. So has Alphabet, which halted stock buybacks and issued equity to fund its spending…[end quote]
The article makes the case that depreciation will skyrocket over the next few years due to huge capital expenses. And that the forecast decline in expenses that yields a positive cash flow isn’t likely since growing businesses usually need more support. Will efficiencies from AI really increase productivity so much that SGA will decline faster than depreciation of delicate AI chips?
It’s hard to make predictions, especially about the future.
But spending almost $1.2 Trillion over the next 3 years will need a LOT of expense adjustments to make financial sense.
The situation is even worse than the WSJ shows because the article only analyzed the strong hyperscalers which have massive cash flows from their existing businesses.
There’s a riskier layer under the hyperscalers. These are called “neoclouds.” A “neocloud” (or specialized AI cloud provider) is a newer type of cloud company that focuses on doing one thing: renting out massive clusters of AI chips (GPUs). Examples include companies like CoreWeave, Lambda Labs, or Together AI.
The neoclouds borrow hugely to buy the chips and they don’t have any other income besides the rentals. Their customers are often startups which are vulnerable - like the dot-coms in 1999. But the chips become obsolete or physically fail in 3 -5 years. If they can’t keep business growing they won’t be able to pay off their debts.
In aggregate, neoclouds are dropping $60 billion to $80+ billion a year on chips and related hardware. Nvidia itself invests equity or provides credit support to neoclouds. The neocloud then turns right around and spends that capital buying more GPUs from Nvidia. That’s a lot like the vendor financing that sunk GE Capital, Lucent and Nortel in 2000.
Put the hyperscaler fun with spreadsheets together with the leveraged neoclouds and you have a lot of money at risk.
And don’t forget the paradigm shift of distilled AI. Once the customer has a distilled program they may call the neocloud that was hosting them and say, “Thanks. We still need you but only 1/10 of the volume we used to use.” But the neocloud still owes big time on the fixed cost of all those chips the customer isn’t using anymore.
When the music stops, the hyperscalers can simply let their neocloud leases expire or walk away from guarantees, leaving private credit markets and vendor-financing balance sheets holding billions in rapidly depreciating silicon.
The hyperscaler stock prices will be hurt but their foundational businesses will be OK. The neoclouds and their lenders could be crushed. Worse than the telecom fiasco (like Global Crossing and WorldCom) because the fiber optics don’t degrade the way chips do.
Wendy