Nvidia's Megawatt Got 60% More Expensive and 30 Times More Productive in One Generation.

Buried in Nvidia's August call are the two numbers that decide whether an AI data centre is a good business, and they point in opposite directions.

The first is what a megawatt costs. Colette Kress walked the series generation by generation:

Since Hopper, our revenue opportunity has grown from roughly $18 billion per gigawatt to $25 billion with Blackwell, to $40 billion with Vera Rubin, which now spans Vera CPU, Rubin GPU, NVLink, InfiniBand or Ethernet, and Groq LPU.

The second is what that megawatt produces. From the same call, two sentences later:

Vera Rubin exemplifies this, delivering 30x higher throughput per megawatt and 35x lower token cost relative to Grace Blackwell Ultra.

Put them together and one generation of Nvidia hardware costs 60% more per megawatt and produces thirty times as much from it. That is an 18.8× improvement in throughput per dollar of hardware, in a single step, and it is the reason buyers keep paying more per megawatt rather than less.

But revenue per megawatt is not throughput per megawatt. It is throughput multiplied by what a unit of throughput sells for, and that second number is falling at least as fast as the first is rising. This piece is about what happens where those two curves meet, because that collision — not the price of the hardware — is what determines whether the megawatt pays for itself.

The points

• $18M → $25M → $40M a megawatt, Hopper to Blackwell to Vera Rubin. Nvidia's own revenue opportunity per gigawatt, disclosed. Up 2.2× across three generations, 60% in the last one.
• 30× the throughput per megawatt, Vera Rubin against Grace Blackwell Ultra. Disclosed, and it is per megawatt rather than per chip — the comparison that matters when power is the constraint.
• 35× lower cost per token on the same comparison. Disclosed, and it is the provider's cost, not the customer's price.
• 18.8× more throughput per dollar of hardware in one generation. Derived: 30× output against 1.6× cost.
• The platform got wider, not just faster. The $40 billion now spans Vera CPU, Rubin GPU, NVLink, InfiniBand or Ethernet and Groq LPU. Some of the 60% is more hardware, not dearer hardware.
• Revenue per megawatt holds flat only if the price of a unit of output falls exactly 30×. Derived, and it is the whole question.

What a megawatt costs, and what is in it

Two things are worth separating here, because they get conflated constantly.

This is not the cost of a data centre. It is what the buyer pays Nvidia — silicon, interconnect and networking. Land, shell, substation, cooling, construction and operations sit on top and are not in any of these numbers.

And part of the rise is scope. Hopper's $18 million bought GPUs. Vera Rubin's $40 million buys a CPU line Nvidia did not previously sell at scale, two networking fabrics, and an LPU from the Groq partnership. Nvidia is capturing a larger share of the same build, so the per-megawatt figure rises even where the underlying components do not. Kress said as much on the call — the full-stack platform is "expanding our share of the data center TAM."

What a megawatt now produces

The 30× is the number that makes the 60% rational, and its unit matters. It is throughput per megawatt, not per chip or per rack. When the binding constraint on an AI build is power — and Nvidia spent much of the call saying supply and power are exactly what bind — throughput per megawatt is the only efficiency figure that converts directly into revenue.

The companion figure, 35× lower token cost, is the same improvement seen from the operator's side. It is what it costs them to produce a token, and it falls faster than the throughput rises because power efficiency improves alongside compute density.

So the operator gets thirty times more output from a megawatt they paid 1.6 times more for. On the cost side of their P&L this is unambiguously good, and it is why a $40 million megawatt is a better purchase than an $18 million one despite costing more than twice as much.

The collision: what actually happens to revenue per megawatt

Here is the part that is not on the slide. Revenue per megawatt is:

throughput per megawatt × price per unit of throughput

Nvidia has told us the first term rose 30×. Nobody has told us what the second is doing, and it is falling fast — that is the entire direction of travel in inference pricing, and it is the subject of our token-economics work. So the outcome for the operator is a race:

Thirty times is the break-even. If prices fall by less than throughput rises, the operator captures the difference and revenue per megawatt goes up. If they fall by more, the hardware is getting better and the business is getting worse at the same time — which is a thing that can happen, and has happened in other commodity-compute markets.

Two observations sharpen it. First, Nvidia's disclosed 35× token-cost improvement is above the 30× throughput gain, which tells you the company itself expects unit costs — and therefore achievable prices — to fall faster than output rises. Second, prices in this market have been falling at rates that make 30× look modest: on ARC Prize's verified runs — the one place a solved reasoning task is priced the same way twice — the cost of the same 87.5% score fell from about $4,560 to $0.30 in twenty months.

That does not mean revenue per megawatt is falling. It means volume has to grow into the gap, which is precisely the Jevons argument — cheaper output opens workloads that were uneconomic before, and total tokens consumed rises faster than price per token falls. The 30× break-even is a clean way to state how much new demand the industry needs to find per generation just to stand still.

Who captures it

The same physical megawatt produces very different revenue depending on who owns the output, and the spread is larger than any of the hardware numbers above.

The frontier figure is an outside estimate, not a disclosure: roughly $50 million of revenue per megawatt for Anthropic in 2026, against a $10–15 million compute cost, published by Tomasz Tunguz and by Dylan Patel of SemiAnalysis. Anthropic's reported $10.9 billion of revenue and $559 million of operating profit are consistent with it. Nobody has filed it, and the row that uses it is directional.

The first two rows are ours, from our model of SpaceX — a capacity-driven model that treats it largely as a landlord: 55% of capacity leased at $2.03 million per megawatt-quarter, 15% reserved for the model layer Grok runs on. A landlord needs nine years of gross revenue to cover the hardware. A frontier lab needs ten months. That is the same silicon, drawing the same power, on the same site.

This is why the 30× matters more to a model provider than to whoever built the building. The throughput gain accrues to whoever sells the tokens. A landlord's rent does not rise 30× because the tenant's hardware got better; it rises with what the market will pay for a megawatt of hosted capacity, which is a different and much slower curve.

What would change the conclusion

• Any disclosed revenue-per-megawatt figure from a frontier lab. The ~$50 million rests on two analysts and it is the largest number in this piece. Anthropic's reported IPO filings would settle it and reprice every model that assumes AI capacity earns a landlord's rate.
• The next generation's per-gigawatt figure. The series has gone $18M → $25M → $40M a megawatt. A fourth step tells you whether Nvidia keeps widening the platform or whether the per-megawatt price is topping out.
• A throughput number on the same basis twice. The 30× is Vera Rubin against Grace Blackwell Ultra. Without a consistent series nobody can tell whether throughput per megawatt is compounding at 30× a generation or whether this was an unusually large step.
• Evidence on what a unit of output actually sells for. Everything above turns on the second term, and it is the one nobody discloses. Token list prices are published; realised revenue per token, after caching, batching, routing and enterprise discounts, is not.

The per-gigawatt series, the 30× throughput and 35× lower token cost for Vera Rubin against Grace Blackwell Ultra, and the description of what the platform spans are all disclosed by Nvidia on its fiscal Q2 2027 call and in the accompanying release, covered in our note on that print. The break-even table, the throughput-per-dollar figures and the payback months are ours, arithmetic on those disclosures. The roughly $50 million of revenue per megawatt for a frontier model provider and the $10–15 million compute cost beside it are estimates published by Tomasz Tunguz and by Dylan Patel of SemiAnalysis, not company disclosures. The landlord rows come from our SpaceX model, whose $2.03 million per megawatt-quarter price and split of capacity between leasing and the model layer are assumptions of ours; SpaceX is private and discloses no capacity, capital expenditure or supplier list.