Revenue per megawatt needs a time period and a cost boundary

Before I accept revenue per megawatt as a business case, I want two labels: the time period and the costs inside the boundary.

Tomasz Tunguz’s August 27 essay relays Dylan Patel’s estimates contrasting compute cost per megawatt with higher revenue attributed to inference capacity. The essay explicitly discusses using inference proceeds to fund training. It does not ignore training altogether.

The problem is what happens when the appealing ratio escapes that explanation. A revenue-to-capacity comparison can begin to sound like a company-wide margin, even though the costs and time periods needed to calculate that margin have not been reconciled.

My correction is to keep the narrow metric and reject the expanded conclusion until the missing bridge is supplied. A useful factory measure should tell us which factory activity it measures.

Capacity is not energy, and neither is a billing period

A megawatt is a measure of power. It does not, by itself, tell us how long the capacity ran, how much of it was utilized or how much useful work it produced.

Revenue needs a period. An amount earned over a month cannot be compared directly with a cost covering a year or an asset’s purchase price. The units can look compatible because both sides mention dollars per megawatt while still describing different economic quantities.

I would ask whether the denominator is installed capacity, contracted capacity, average power draw or capacity allocated to a specific workload. Then I would ask which portion was productive during the revenue period. Those definitions affect the result before any strategic argument about AI begins.

A strong month on fully utilized inference infrastructure may be a useful operating observation. It cannot automatically price a new facility that will take time to fill, or a company that also spends heavily on research and training.

Keep contribution separate from the whole business

An inference operation may generate revenue above its directly attributed serving cost. That contribution can help fund other activities. It is not the same as profit after every activity has been included.

The distinction does not make the inference business less valuable. It tells us how to connect its performance to the broader organization. Training, research, support, unused capacity and other costs need a consistent accounting treatment before the ratio becomes a company-margin claim.

I would not assume that exactly half the factory is missing. The split between training and inference is an empirical question, and it can change. Nor would I infer a hidden accounting problem merely because a narrow contribution estimate differs from an operating margin. They are supposed to differ when their boundaries differ.

The task is reconciliation: identify what each measure includes and show how one becomes the other. If the data is unavailable, keep the limitation visible instead of filling it with a dramatic fraction.

Model a facility through its ramp

For a hypothetical AI infrastructure proposal, I would build a monthly model rather than begin with one peak revenue-per-megawatt figure. Record available capacity, usable capacity, utilization and the realized revenue for the work completed.

Costs would be entered with their actual timing and category. A capital purchase belongs in the investment and financing view; depreciation belongs in the relevant accounting view; recurring operating costs belong in the period they support. The analysis should connect those views without casually adding incompatible numbers.

Then I would test slower demand, lower realized prices and delayed capacity availability separately. Each can weaken the economics even when a best-case serving ratio looks attractive.

rendering diagram…

This is a proposed analytical structure, not an estimate of a private model company’s finances. The missing disclosures are part of the reason to keep the structure explicit.

Forecasts need their own column

The same discipline applies to hardware earnings. Nvidia’s Q2 fiscal 2027 release reports $96.2 billion in revenue and gives a $108 billion revenue outlook for the following quarter. The second number is guidance, not a booked quarter.

A forward annualization may be useful for a scenario. It should not be placed beside another company’s trailing revenue as though both were realized results over the same interval. The comparison needs to retain the distinction between observed and expected activity.

I would use the same labels in the infrastructure proposal: actual, contracted, forecast and peak scenario. A buyer should be able to see which assumptions are supported by existing receipts and which depend on a future utilization or pricing outcome.

Make the attractive number earn its use

The right response to a promising unit metric is to ask where it helps. Revenue per unit of constrained power might be valuable for comparing uses of an operating facility under matched conditions. It becomes less useful when it is asked to explain profitability across different time periods and cost structures.

A narrower measure with clear units can support a better decision than a sweeping margin story. Keep the numerator, denominator and accounting bridge visible, and the reader can decide what the number actually buys.

Revenue per megawatt becomes useful only when the time period, utilization and included costs travel with it.