NVIDIA sold $96 billion of chips in three months. Follow the money backwards.

On 26 August NVIDIA reported quarterly revenue of $96.2 billion, up 106% year on year and 18% on the previous quarter. Data-center revenue was $89.0 billion, up 117%. Gross margin was 75%. The outlook for the following quarter is $108 billion, give or take 2%. Jensen Huang's framing of it was that compute is now revenue.

Numbers that large stop meaning anything after a while, so here is a way to make this one mean something: run it backwards through the map.

Backwards from the invoice

Someone manufactured the chips. NVIDIA designs; it does not fabricate. Every Blackwell part in that $89 billion was made by TSMC, which is the single most load-bearing relationship on the whole map — and, by a wide margin, the one with the most companies standing behind it. The lithography machines are ASML's. The deposition and etch tools come from Applied Materials and Lam Research. The inspection is KLA's. The chip was designed with Synopsys and Cadence software.

Someone made the memory. A modern AI accelerator is mostly a memory problem wearing a logic costume. SK hynix, Micron and Samsung make the high-bandwidth memory stacked beside the processor, and their production capacity is as real a constraint on this number as wafer supply.

Someone assembled the systems. Dell, Supermicro, HPE, Hon Hai and Quanta turn chips into racks. Vertiv and Eaton supply the power and cooling those racks need. Somebody ran a wire to the building, which is where the energy layer of the map starts.

Someone bought it. Overwhelmingly the clouds and the labs — Microsoft, Amazon Web Services, Google, Oracle, CoreWeave, Meta — spending capital today against revenue they expect later.

That chain is the whole argument of this site in one sentence: the interesting thing about the biggest company in the AI economy is that none of its quarter happens alone.

And the gains are broadening

A week later, on 2 September, Broadcom reported revenue of $29.6 billion, up 86%, with AI semiconductor revenue of $16.7 billion — up 221% year on year and 54% quarter on quarter — and guided to $21.7 billion next quarter. Broadcom's AI business is largely custom accelerators designed with and for individual customers, Google's TPU line being the best known.

That matters for a reason beyond Broadcom. One company growing spectacularly can be a bubble; a second, structurally different company growing spectacularly in the same quarter is harder to explain that way. It is evidence that demand is broad enough to support more than one architecture, which is the half of the valuation argument that usually goes untested.

What this is evidence for, precisely

On the thesis tracker, both prints land on the valuation row: the claim that today's AI valuations are grounded in reported earnings rather than in a 2000-style story about the future. Earnings of this size, at this growth rate, with this margin, are the strongest available evidence for that claim, and we file them as supporting it.

What they are not evidence for is the row underneath, the one asking whether the spending is sustainable. NVIDIA's revenue is its customers' capital expenditure. A supplier's record quarter proves money was spent; it cannot prove the spenders can keep spending, or that their own customers will eventually cover it. That question is answered in the buyers' cash-flow statements — which is exactly why Oracle's results two weeks later were filed as a mixed signal rather than as more good news.

The optimism, with the receipts attached

It would be easy to read this site's signal feed — permit freezes, rate rises — and conclude the AI build-out is in trouble. It is worth saying plainly that the two largest evidence points of the last month were both enormous, audited, reported revenue increases from companies selling real hardware to real buyers.

The discipline is to keep both in view. The constraints are getting harder. The demand, so far, is getting bigger.