The AI selloff is an accounting argument, not a technology one
I do the books at a manufacturing company. So when I read that more than a trillion dollars came off chip stocks in a matter of days, my instinct was not to ask whether AI is real. It was to ask what the numbers look like.
They look like something I recognise from month-end: a very large capital expenditure, and a revenue line that has not caught up with it yet.
Nvidia reports its results on 26 August, and it is being treated as a verdict on the entire sector. So this is a good moment to explain what is actually being argued about, because most of the coverage frames it as a debate about technology when it is really a debate about accounting.
What happened
Late July was brutal. Chip stocks shed more than a trillion dollars in market value across a few sessions. Nvidia alone lost around $238 billion from one Friday close. The memory makers were hit hard too — SK Hynix, Samsung and Micron together lost several hundred billion more.
August has been choppy rather than catastrophic. Another heavy session in mid-month took roughly $130 billion off Nvidia in a day. The stock has spent the month somewhere around $219, up about 18 percent for the year — a perfectly respectable number that feels like a disappointment only because of what came before.
Analysts, for what it is worth, still have targets clustered well above that.
The part that is not in doubt
Here is what makes this different from an ordinary bubble story: demand is not falling.
Microsoft’s cloud business grew 43 percent in its latest quarter. Nvidia’s data centre revenue grew 92 percent. The company’s own guidance for the quarter it reports tomorrow was $91 billion, which would be roughly 95 percent growth year over year.
Those are not the numbers of a collapsing market. Nobody is selling because they think nobody wants AI chips.
So what is the argument about?
Look at what the big buyers are spending on infrastructure this year:
- Amazon — around $200 billion
- Alphabet — $195 to $205 billion
- Microsoft — around $175 billion
- Meta — $130 to $145 billion
That is roughly $700 billion of capital expenditure from four companies in a single year.
In accounting terms, capex is not an expense you take today. You capitalise it and depreciate it across the asset’s useful life. Which means the spending shows up on the balance sheet now, and the cost hits the income statement gradually for years afterwards — whether or not the asset earns anything.
That is the entire question the market is asking. Not “does AI work” but “will the cash these assets generate exceed the cost of the assets, before the depreciation schedule catches up?”
It is the same question I would ask about a new production line. The machine is impressive. The machine is running. Is the machine paying for itself?
Why nobody can answer it yet
Two reasons, and both are genuinely hard.
The useful life is unknown. When you depreciate a lathe over ten years, you are making a defensible estimate. How long is an AI accelerator economically useful? Three years? Five? If a better chip arrives in eighteen months and your customers migrate, the asset is not worn out — it is obsolete, which is worse, because you are still depreciating it. The assumptions being used here have not been tested by a full cycle.
The revenue is arriving somewhere else. Cloud revenue is growing strongly, so some of it is clearly landing. But a great deal of AI spending currently produces capability rather than cash — better products, retained users, competitive position. Real value, difficult to put a number on, and impossible to match against a specific depreciation line.
One analysis I came across claimed that for every dollar spent on AI tokens, only about eighteen cents produces user-facing value, with the rest going to rework and review. I would treat that figure carefully — it is one estimate and the methodology is not transparent. But the direction it points is the same one the market is worried about.
What I noticed in the coverage
A small aside, but relevant to anyone researching this.
While reading around for this piece, I found articles confidently stating Nvidia’s expected revenue as $91 billion, $93 to $95 billion, and — in one case — $28.7 billion. That last figure appears to be a stale number from a previous year, republished as current.
A lot of financial commentary is now machine-generated at volume, and it confidently reproduces figures that are simply wrong. When the numbers matter, go to the company’s own investor relations page. Nvidia’s guidance came from Nvidia in May: $91 billion, plus or minus two percent.
What to watch tomorrow
If you follow the earnings release, the headline revenue number is the least interesting part. It will almost certainly be enormous.
The things that actually address the argument above:
- Guidance for next quarter. Forward-looking, and the closest thing to management telling you what they see in the order book
- Gross margin. If it slips, competition is starting to bite. AMD and custom silicon from the cloud providers are real now
- Customer concentration. A handful of hyperscalers driving most of the revenue is a risk, because those four companies set their own capex budgets
- Anything said about depreciation assumptions — by Nvidia or by its customers. This is the quiet number that decides whether the maths works
The honest position
I do not know whether this is a bubble, and neither does anyone writing confidently that it is or is not.
What I would say is that the two questions have been getting muddled. “Is AI transformative?” and “are these companies worth what they cost today?” are separate questions with separate answers. The technology can be genuinely important and the stocks can still be expensive. Railways changed the world and ruined a great many investors on the way.
What strikes me from the ledger side is how ordinary the underlying question is. Strip out the language about frontier models and superintelligence and you have a business that has bought a very large amount of equipment on the expectation of future revenue. Every accountant has seen that situation. Sometimes it works.
The difference here is the size of the number, and that nobody has done this particular sum before.
I am an accountant, not a financial adviser, and nothing here is investment advice. Figures were accurate as of 25 August 2026 and are drawn from published reporting; verify anything you plan to act on against primary sources.
Related reading
