On Nvidia's August call for its fiscal second quarter, Jensen Huang did the thing companies almost never do. He gave the numerator and the denominator in the same answer:
All of that increased our revenue contribution or revenue opportunity per gigawatt to $40 billion. Each gigawatt of data center increased from, say, $30 billion about 5 years ago to now $60 billion today. [...] But you are talking about a $60 billion investment.
The $40 billion has been quoted everywhere since the call, and we used it ourselves to work out what a megawatt now costs and what it produces. The $60 billion is the more interesting half, because it is the only figure anyone has put on the rest of the building — and it turns a big number into a share. Two thirds of the money spent on a gigawatt of AI infrastructure goes to one supplier, and the supplier is the one who told us.
The half that is not Nvidia got cheaper
Huang gave his own per-gigawatt figures generation by generation on the same call: about $3–5 billion in the general-purpose computing era, $18 billion with Hopper, $25 billion with Grace Blackwell, $40 billion with Vera Rubin, "and after that it is going to be higher."
Set the two ends of that series against the two costs of a gigawatt and the arithmetic is short:
- Then, a $30 billion gigawatt with $3–5 billion of Nvidia in it. Nvidia's share: a tenth to a sixth.
- Now, a $60 billion gigawatt with $40 billion of Nvidia in it. Nvidia's share: two thirds.
- Everything else — land, shell, substation, cooling, construction, storage, non-Nvidia silicon, the people who build it — was $25–27 billion a gigawatt and is now $20 billion. Down roughly a fifth to a quarter, in nominal dollars, while the total doubled.
- Nvidia's slice grew by $35–37 billion. The whole gigawatt grew by $30 billion. The increase in Nvidia's take is larger than the increase in the bill.
For scale on how far that has moved: in calendar 2021 — the era of the $30 billion gigawatt — the whole of Nvidia's revenue, gaming included, was $24.3 billion. One gigawatt of Vera Rubin is now worth more to the company than everything it sold that year.
The pairing at the older end is ours, and it deserves its caveat in the sentence that uses it rather than in a footnote. Huang dated the $30 billion gigawatt only as "about 5 years ago", and described the $3–5 billion as the general-purpose era "during Moore's Law" — two loose dates, in two different answers, on the same call. Putting them side by side is our reading, not his. What is not ours is the ordering: on his own sequence, general purpose comes before Hopper, and Hopper's $18 billion is already three generations old.
It is not a rack price, and it cannot be made into one
The obvious next move is to divide. Nvidia sells the data centre as rack-scale systems — Huang called NVL72 "the world's first rack scale architecture" and said Nvidia is now on its third generation of it — so $40 billion a gigawatt looks one division away from a price per rack.
It is not, for two reasons the company supplied itself.
The $40 billion is wider than a rack of GPUs. Colette Kress said on the same call that it "now spans Vera CPU, Rubin GPU, NVLink, InfiniBand or Ethernet, and Groq LPU" — a platform, not a product.
And the conversion needs a number nobody has published. Dollars per gigawatt become dollars per rack only through rack power, and Nvidia gave no power figure, no rack count and no unit volumes at all. Every choice produces a different answer:
| If the average rack draws | A gigawatt holds | Nvidia's content per rack |
|---|---|---|
| 92 kW | 10,870 racks | $3.68M |
| 120 kW | 8,333 racks | $4.80M |
| 140 kW | 7,143 racks | $5.60M |
| 200 kW | 5,000 racks | $8.00M |
The gigawatt itself is ambiguous too: IT load, or the whole facility with its cooling overhead. If it is the facility, every row holds fewer racks and every price in the right-hand column goes up. This is where most analysis quietly picks a row. We are not going to — the honest statement is that dollars per gigawatt and dollars per rack are separated by a figure Nvidia has not disclosed, and anyone quoting a rack price off this number has chosen one.
Where it meets our model
Our Nvidia model drives both data-centre lines off a single rack count: $2.97 million of compute and $725,000 of networking per rack-equivalent, $3.695 million in all. That is close to the bundle Nvidia says the $40 billion spans, less the Groq accelerator. Multiply it by the 24,095 rack-equivalents the model carries for the June quarter and it produces $89.0 billion — the data-centre revenue Nvidia reported.
So the disclosure meets our model at exactly one row of that table: the first. $40 billion divided by $3.695 million is 10,825 racks in a gigawatt, an average draw of about 92 kilowatts each. Anyone who thinks a Vera Rubin rack draws more than 92 kilowatts thinks our content-per-rack figure is low, and the model's own note already says as much.
What that would change is not what it looks like. The rack count is backed out of revenue at the assumed price rather than disclosed, so a dearer rack means fewer racks behind the same reported revenue, and the near-term revenue path does not move at all. What moves is the distance to the 65,000-rack-a-quarter shipment ceiling — the largest single lever in the model, and worth more than seven times what losing a quarter of today's shipments costs. At $3.695 million a rack the line reaches that ceiling in about eight quarters; at $4.80 million, ten; at $5.60 million, eleven. A higher rack price pushes the model's binding constraint further out, not closer.
That is also why neither number has moved. A denser, dearer rack should mean fewer of them per quarter of packaging and memory, so the same disclosure that argues our price is low argues the 65,000-rack ceiling is high. Changing one without the other would be picking the half that flatters the answer.
What settles it
- A published power figure for a Vera Rubin rack. It is the one number that converts $40 billion a gigawatt into a price per rack, and until it exists every such conversion in public is a choice dressed as arithmetic.
- The next generation's per-gigawatt figure. Asked directly whether $40 billion becomes $60 billion or $80 billion, Huang said only that it goes higher. At two thirds there is not much room left above Nvidia inside a $60 billion gigawatt — the next step has to come mostly from the gigawatt costing more, not from Nvidia taking more of it.
- Whether the $60 billion is a number Nvidia stands behind. It came in an unscripted answer, with a "say" in front of it, and it appears in neither the results release nor the CFO commentary. It is the only figure here the company has not repeated anywhere else.
- What anyone else says a gigawatt costs. Data-centre developers and landlords publish cost per megawatt for shells, land and power. If those add to something near $20 billion a gigawatt, the $60 billion holds together; if they add to far more, Nvidia's two thirds is too high.
The per-gigawatt series — $3–5 billion for general-purpose computing, $18 billion for Hopper, $25 billion for Grace Blackwell, $40 billion for Vera Rubin — the $30 billion and $60 billion cost of a gigawatt of data centre, and the description of what the $40 billion spans are all Nvidia's, from its fiscal Q2 2027 earnings call on 26 August 2026; the call text is a third-party transcription and every figure quoted from it was checked against the company's own results. Reported data-centre revenue and calendar-2021 revenue are as Nvidia reported them. Everything else is ours: the share percentages, the $20 billion and $25–27 billion of non-Nvidia content, and the pairing of a general-purpose-era per-gigawatt figure with a gigawatt cost Huang dated only as about five years ago. The rack table, the 92-kilowatt average and the quarters-to-ceiling figures are arithmetic on our Nvidia model, whose rack price, networking content per rack and 65,000-rack ceiling are assumptions of ours, not disclosures. Nvidia publishes no unit volumes and no rack power figure.
