You Can't Lay Off a Datacenter | Okulez
Research Note<br>Published July 22, 2026
You Can't Lay Off a Datacenter
Three companies missed on free cash flow this week. Google, IBM, ServiceNow. Same week, same metric, and<br>the misses point in opposite directions. Google is spending the cash. IBM's customers are taking cash they<br>used to spend on IBM and spending it on AI hardware instead. ServiceNow got caught between them.
Revenue can look fine and earnings can look fine while this happens, because capital spending does not<br>hit the income statement up front. It hits the cash. And the cash is what the market caps were built on.
Now hold that next to the other set of numbers from the very same week. GE Vernova, which sells the<br>turbines that power datacenters, took in 88 percent more orders than a year ago and is sitting on the<br>largest pile of future work in its history, with turbine capacity effectively sold out for years. Texas Instruments' data center chip revenue is<br>up around 90 percent from a year ago. The companies selling into the buildout are having the best<br>quarters they have ever had, in the same week the companies funding it watched their free cash flow<br>halve. Both things are true at once. I wrote earlier this year about<br>what euphoria looks like from the inside. It looks like this:<br>record orders on one side of the ledger, vanishing cash on the other, and a market cheering both.
Here is what it reminds me of. When COVID hit, the government poured trillions into the economy, and<br>every tech company started hiring. Not because each one had carefully counted the work. Each company was<br>hiring because its competitor was hiring. Headcount became a signal. No one could stand still while<br>rivals grabbed talent, so everyone ended up with a lot of people. I watched it from inside the industry.<br>Firms outbid each other with bigger and bigger offers, and people learned to switch jobs every six<br>months, because every switch was another raise. The companies were not really hiring workers and the<br>workers were not really choosing employers. Both sides were trading the frenzy. Then the money<br>normalized, the demand normalized, and the same companies spent two years firing the people they fought<br>over.
Swap engineers for GPUs and you are looking at today. Nobody can be the one who underinvests in AI. Meta<br>has said it in the open: the risk of spending too little is bigger than the risk of spending too much.<br>Every rival's datacenter announcement forces the next one. It is the same mirror, the same arms race,<br>the same inability to stop, because stopping means conceding the race while the market still rewards<br>the spend.
But there is one difference, and it is the whole story. The COVID mistake was made out of people. People<br>are painful to let go, but on a balance sheet they are an expense, and an expense can be cut in a<br>quarter. That is why the 2022 layoffs stung and then passed, and margins were repaired inside two years.<br>This mistake is being made out of concrete, chips, leases, and debt. Google spent $44.9 billion on<br>construction and chips in a single quarter, double a year ago, and now plans as much as $190 billion<br>for the year. A year and a half ago, for every dollar Google took in, about 21 cents came out the other<br>end as cash. Now it is about 9 cents, and the company is telling you next year will be bigger. And that<br>is only the visible part. Nikkei's reporting puts roughly $1.65 trillion of AI-related financing off the balance<br>sheets entirely, in special vehicles and lease structures that are economically debt but do not read as<br>debt, on top of about $1.35 trillion that does. The numbers beyond the confirmed cases are estimates.<br>The structure is not.
You can lay off an engineer. You cannot lay off a datacenter. If the demand disappoints, this excess<br>does not clear through a hiring freeze. It clears through writedowns, stranded assets, and lease chains<br>that stop paying. We have seen that version before too. In 1999 the telecom companies borrowed fortunes<br>to lay fiber for an internet that was genuinely coming. The internet came. The fiber got used. The<br>companies that laid it did not survive to see it, and the people who lent against it ate the loss. Real<br>demand does not protect the builders who overpaid to meet it.
The bulls have two good answers here, and they deserve saying out loud. First, Google, Meta, and<br>Microsoft are not 1999 telecom startups. They throw off hundreds of billions a year from search, ads,<br>and software, and they can absorb their own writedowns without a solvency scare. True. Second, a<br>datacenter is not dead fiber: the land, the shell, the power hookups hold value even if the chips<br>inside, which are now the majority of the spend, age out in a few years. Also true. But notice what<br>each concedes. If the cash cows have to absorb the capex, free cash flow stays depressed, and the<br>valuations were built on the old cash flows, so the cushion protects solvency, not the price. And I am<br>not arguing Google goes...