A Quarter of a Trillion Dollars, Paid For and Not Running

Ed Zitron published Where’re All The AI Chips? on 22 September 2026, arguing that the AI buildout has produced far less running capacity than the spending implies, and that a great deal of hardware is sitting in warehouses and half-built facilities. It is a strong piece with a strong thesis, and parts of it rest on reporting from sources we cannot check.

The parts that rest on public filings we can check, so we did. This is what four of his load-bearing numbers look like when pulled from the primary documents rather than quoted.

One of them is stronger than he makes it. One of them is measuring the wrong balance sheet. And one of them turns out to be false in a way that is more interesting than if it had been true.

The number that does the work

Under US accounting rules an asset does not start depreciating until it is ready for its intended use. Until then its cost accumulates in a line called construction in progress. That makes it the closest thing in audited financial statements to a direct measure of “bought, not yet running”.

Extracted from each company’s most recent Form 10-K. Fiscal years differ; Oracle’s ends 31 May.

A quarter of a trillion dollars, capitalised, audited, and by the companies’ own accounting not yet earning anything. That is a stronger statement of the case than the chip-counting that surrounds it, and it does not depend on anybody’s source.

Three things that made it awkward to collect

Alphabet does not use the words. Its line reads “Add: assets not yet in service”. A search for “construction in progress” misses it entirely, which is how our first pass returned $104 billion instead of $250 billion — we had found only the two companies that tag the concept in machine-readable form.

Microsoft does not disclose it. There is no construction-in-progress component in its property and equipment table. It discloses contractual construction commitments of $34.6 billion, of which $29.8 billion falls due within a year, but that is an obligation to spend rather than an asset already paid for. We have left it out of the total rather than quietly making the two comparable.

That absence matters more than the other five entries combined, because Microsoft is the largest builder and the subject of most of the argument. The single best public measure of the thing being debated is not published by the company being debated.

Column order is not consistent. Amazon and Alphabet print the earlier year first; Meta, Oracle and CoreWeave print the later year first. Reading position rather than the header would have halved two of the five figures, and the result would have looked entirely plausible.

The number that is measuring the wrong company

NVIDIA’s inventory has grown sharply, and it is regularly offered as evidence that GPUs are piling up unused.

Days of inventory computed as inventory divided by quarterly cost of revenue, times ninety. Source: NVIDIA quarterly filings via the SEC’s XBRL company-concept API.

Inventory is up 5.4 times in two years and days of inventory has roughly doubled from the early-2025 low. Relative to sales the move is milder — from about 0.22 to 0.33 — because revenue more than tripled underneath it.

But this number cannot answer the question it is being asked. Inventory on NVIDIA’s balance sheet is product NVIDIA has not sold. A GPU shipped to a customer leaves NVIDIA’s books completely and reappears on the buyer’s, in property and equipment or in the construction-in-progress line above. Rising NVIDIA inventory says NVIDIA is building stock ahead of shipment, which is a fact about NVIDIA.

The chips allegedly sitting in warehouses are, by definition, chips that were sold. They are in the $250.2 billion, not in the $31.6 billion. Presenting the two together as though they were the same argument double-counts nothing but does confuse two different claims, and the weaker of them is the one that gets quoted.

The claim that turned out to be false

The standard version is that hyperscalers keep extending the assumed useful life of servers, which lowers reported depreciation and flatters earnings without anything changing in the data hall. It is a good argument and it has the advantage of being checkable.

Both of the companies that changed an estimate did so effective 1 January 2025. They moved in opposite directions.

Meta did what the argument predicts, and the effect is large: a dollar a share.

Amazon did the reverse, and said why:

The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.

That is a company telling its auditors that AI hardware wears out faster than it used to, and accepting $1.4 billion of additional depreciation to say so. It is the opposite of flattering the numbers, and it is more damaging to the overbuild case than the version where everybody extends: if accelerators obsolete faster than the six years previously assumed, then hardware waiting in a half-built facility is losing value while it waits.

So the useful-life claim, as usually stated, is not supported. One extended, one shortened, one unchanged. What replaces it is a narrower and sharper observation, which is that the two companies closest to the workload disagree by a year about how long the equipment lasts.

Nobody discloses the life of a GPU

We went looking for the useful life of an AI accelerator specifically and it is not in any of the six filings. Every disclosure is written at the level of “servers and network equipment” or “technical infrastructure”. Microsoft’s two-to-six-year range covers that whole category without separating an accelerator from an ordinary server, and a range that wide cannot be pinned to any particular asset.

This is worth stating plainly because the GPU depreciation schedule is central to almost every argument about the economics of AI infrastructure, including ours. It is not a disclosed number. Every figure in circulation for it, including any we have used, is an estimate.

The same gap, measured from Canada

We came to this from the other direction. Our own tracked set of Canadian data centre projects carries twenty sites with source URLs and confidence flags on every row, and it separates capacity into five columns because one column could not hold the question honestly.

Operating capacity is 1.1 per cent of what has been announced. The 112 MW is three projects: Vantage’s Montreal campuses at 91 MW, QScale Q01 at 14, and Bell’s Merritt site at 6. The headline gigawatts — Wonder Valley at 7.5 GW, Beacon at 4.5 GW across six sites — are proposals.

The American filings and the Canadian project data are measuring the same gap from opposite ends. There the money arrives before the capacity; here the announcement arrives before the concrete.

And the Canadian data suggests why the money cannot convert as fast as it is raised. Alberta capped new large-load grid connections at 1,200 MW to 2028 against 29 projects requesting more than 21 GW, against a provincial peak demand of about 12,000 MW. That 1,200 MW is already fully allocated to two projects. This is not a procurement delay that money can solve. It is a queue with a ceiling, described in our chapter on power.

Where we disagree with the piece

We have argued elsewhere that the dot-com fibre boom is a better analogy for this than 2008, and Zitron rejects that directly: GPUs, he writes, are nothing like dark fibre.

He is right about the asset and we were not careful enough. Dark fibre was the ideal overbuild — almost no carrying cost, no obsolescence, and it was eventually lit. A server depreciates on a schedule the filings put between two and six years, and Amazon has just shortened its own estimate. The half of this buildout that has been built has a clock on it that fibre never had.

What survives from that piece is the narrower claim it was actually making, which was about financial contagion rather than asset quality: there is no subprime-style transmission mechanism here, and none of this requires Microsoft to default on anything. That still looks right. “The fibre comparison is more interesting” was doing work our own data does not support.

What we would not claim

Construction in progress rising is what a buildout looks like. A company spending $115.9 billion a year in capital expenditure, as Microsoft did in fiscal 2026, will have a great deal in flight at any moment, and none of the five figures above is evidence of anything improper. They are evidence that a large amount of capital has been committed and has not yet become capacity.

We also looked at depreciation as a share of capital expenditure, which has fallen sharply across all four large filers — Microsoft from 70 per cent in 2019 to 30 in 2026, Oracle from 74 to 14. It reads like strong evidence and we are not leaning on it, because the ratio falls during any rapid buildout by construction: depreciation reflects the accumulated in-service base while capital expenditure is the current year’s flow. It would fall even if every dollar were installed the day it was spent. The magnitude is striking. The inference is not available.

The honest summary is that one of these measures is decisive, one is misdirected, one is false as usually stated, and one is suggestive and unusable. That is a less satisfying conclusion than the article’s, and it is what the filings support.

Frequently asked questions

Figures extracted from SEC Forms 10-K and from the SEC’s XBRL company-concept API, with accession numbers recorded per row in hyperscaler_cip_and_useful_life.csv and hyperscaler_capex_depreciation.csv. The Canadian project figures come from the tracked set behind Canada’s Data Centre Race. The piece that prompted this is Where’re All The AI Chips? by Ed Zitron.