A table has been going round since yesterday titled "Top 10 AI Neoclouds by Contracted Revenue per Megawatt." It ranks $IREN at ~$40M/MW, $GLXY at ~$29M, $CRWV and $HUT at ~$28M, $APLD at ~$26M, $WULF at ~$25M, $NBIS at $20–25M, $CORZ at ~$21M and $CIFR at ~$13M, each with a "Total Platform" gigawatt figure in brackets. The original post has a few hundred views. The framing has travelled a great deal further than that — it is circulating as a "Neocloud Efficiency Ledger," and X's news module is carrying it as "IREN Tops AI Infrastructure Efficiency Ranking by Revenue per Megawatt" against thousands of posts.
The table is worth checking because we hold the underlying disclosures for three of its rows. The check fails, and it fails in a specific and fixable way: the rows are not the same metric. Some divide cumulative multi-year contract value by megawatts. At least one divides an annual figure by megawatts. Nothing in the table divides by years.
The summary sits at the top of this page. The rest of this piece is the arithmetic behind it.
The one row that reproduces exactly
Start with CoreWeave, because it is the row we can rebuild from a filing without a single assumption.
Our Q2 2026 capture has revenue backlog of approximately $104 billion at 2026-06-30, up from $99.4B at 2026-03-31, with a footnote that the figure excludes more than $25 billion of net new customer commitments added in early Q3. The same capture has total contracted power of approximately 3.7 GW, against active power that expanded by nearly 500 MW in the quarter to 1.5 GW.
$104B ÷ 3,700 MW = $28.1M per MW.
The table says ~$28M. That is not a coincidence and it is not an estimate — it is the table's own arithmetic, reproduced. So for CoreWeave the metric is unambiguous: total revenue backlog, a cumulative multi-year contract value, divided by contracted megawatts.
Two things follow immediately. First, whatever the "Total Platform: ~5.2 GW" in the CoreWeave bracket is, it is not the denominator — our capture has 3.7 GW contracted and 1.5 GW active, and neither is 5.2. The parenthetical is a third, differently defined number that plays no part in the ratio it appears beside. Second, if CoreWeave's row is backlog-based, every row has to be backlog-based for the ranking to mean anything.
The row that breaks it
Now Nebius, seventh on the list at $20–25M/MW.
Our Q2 2026 capture contains that exact range, and it is explicit about what it is: four landmark deals, average TCV above $1 billion each, ACV $20–25M per MW. Annual contract value. The company also raised contracted-power guidance to more than 5 GW and reported more than $40 billion of customer commitments.
Those two sentences cannot both be inputs to the same league table. If the average landmark deal carries total contract value above $1 billion while yielding $20–25M per megawatt per year, then the cumulative figure per megawatt — the number CoreWeave is being scored on — is several times larger. Nebius is in seventh place because its row is annual and the row above it is cumulative.
It is worth being precise about the scope of that objection: the $20–25M ACV figure describes four specific Q2 deals, not the whole contracted book, so it is not a like-for-like company-level number either. That is a second problem with the row, not a defence of it.
Two contracts where every term is public
The cleanest way to show what the mixing does is to take deals where the dollar value, the megawatts and the term are all disclosed, and compute the metric both ways on the same contract.
IREN and Microsoft. We covered this contract in February, when IREN joined the MSCI USA Index: a five-year, $9.7 billion GPU cloud services agreement signed in November 2025, targeting 200 MW of critical IT load at the Childress, Texas campus across four build phases.
- $9.7B ÷ 200 MW = $48.5M per MW over the term.
- $9.7B ÷ 5 years = $1.94B a year; ÷ 200 MW = $9.7M per MW per year.
Nebius and Microsoft. The Vineland, New Jersey campus, announced in September 2025: a reported $17.4 billion base value over five years — up to roughly $19.4B with expansion options — against a campus designed for about 300 MW.
- $17.4B ÷ 300 MW = $58.0M per MW over the term.
- $17.4B ÷ 5 years = $3.48B a year; ÷ 300 MW = $11.6M per MW per year.
Same customer, same five-year term, both metrics computed on the same disclosure. The gap between the two columns is 5×, and it is 5× for one reason: there are five years in the contract. Nothing about pricing, utilisation, power cost or GPU efficiency produces that spread. The term does.
This is the whole argument in one line. A number that changes by 5× depending on whether you divide by the contract length is not measuring how efficiently a megawatt has been sold. It is measuring how long the lease is.
Reorder the table and it inverts
Put the two ends of the ranking on the same footing and watch what happens.
| Per MW, whole term | Per MW, per year | |
|---|---|---|
| IREN — Microsoft, Childress | $48.5M | $9.7M |
| NBIS — Q2 2026 landmark deals | not stated | $20–25M |
On the ranking's cumulative basis, IREN is first and Nebius is seventh. On an annual basis — the only basis on which "efficiency" is a coherent word — the Nebius deals earn two to two and a half times as much per megawatt per year as the IREN contract that anchors first place.
We are not claiming this reversal is the true ranking. The IREN row here is one contract, not IREN's book; the Nebius row is four deals, not Nebius's book. What it establishes is narrower and sufficient: the ordering of this table is not robust to the choice of numerator, and the choice of numerator is currently being made row by row.
There is a further wrinkle that cuts the same way. The Vineland contract, at $11.6M per MW per year, is a 2025 deal, and it sits well below the $20–25M ACV Nebius reported on its Q2 2026 cohort — and below the $40–50M per MW the company quotes for its short-term capacity and auction pilot against a $12M 2026 base. Nebius's own numbers show annual pricing per megawatt moving by a factor of four in a year. A ranking that reports one figure per company, undated and undivided by term, cannot see any of that.
What is wrong with the denominators too
The numerator is the load-bearing problem, but the denominators are not clean either.
The bracketed "Total Platform" gigawatts are not what the ratios are computed from — CoreWeave proves that, at 5.2 GW in brackets against the 3.7 GW that reproduces the number. Across the rest of the table these figures appear to mix contracted, secured and pipeline power without labelling which is which. IREN's ~5.0 GW bracket, for instance, sits against a company whose disclosed AI contracts cover a few hundred megawatts of critical IT load; our own coverage recorded secured power passing 4.5 GW in February, including a planned 1.6 GW Oklahoma campus that was then an announcement, not a rack.
A megawatt of signed contract, a megawatt of secured interconnect and a megawatt of energised IT load are three different assets with three different costs of capital. A per-MW ratio that does not say which one is in the denominator can be moved by a factor of five without anybody typing a wrong number.
The rows we are not going to pretend to check
Six of the ten names — GLXY, HUT, APLD, WULF, CORZ and CIFR — sit outside the set we hold filings for. We are not going to reverse-engineer them from the ranking itself, because that is circular, and we are not going to estimate them, because the entire objection here is to unsourced arithmetic.
So this piece audits three rows and drops seven. That is the honest shape of it, and it is also the point: a four-row table that can be traced to disclosures is worth more than a ten-row table that cannot. If the ledger has a published methodology page naming its numerators, terms and power definitions, the audit gets easier and we will happily run it. We could not find one.
What the metric should be
The fix is not complicated, and it is three columns rather than one.
- ACV per MW — annual contract value divided by contracted megawatts. This is the closest thing to a yield on a megawatt, and it is comparable across companies and across vintages.
- Weighted average remaining term — because a $20M/MW/year contract with seven years to run is a different asset from the same rate with eighteen months to run.
- The denominator's definition — contracted, energised or secured, stated explicitly.
TCV per MW is the product of the first two, and it is a perfectly good number to publish alongside them. It is only misleading when it is published on its own and labelled efficiency, because then a company that signs ten-year deals at a modest annual rate outranks one signing three-year deals at a rich annual rate, and the table calls the first more efficient.
The missing variable is duration, and it is frequently undisclosed. Where it is disclosed — the two Microsoft contracts above — it does all the work. Where it is not, the honest conclusion is not a lower ranking but no ranking: this table cannot be computed consistently from public data, and the version circulating has not tried.
We have used per-megawatt economics twice in the last fortnight ourselves, on Nebius's Q2 print and on CoreWeave's. This is the correction we are applying to our own future use of it: revenue per megawatt is not an efficiency metric unless you divide by years.
CoreWeave backlog (≈$104B at 2026-06-30, excluding >$25B of early-Q3 commitments) and contracted power (≈3.7 GW, active 1.5 GW) are from our Q2 2026 release capture. Nebius ACV of $20–25M per MW, average landmark-deal TCV above $1B, contracted power above 5 GW, customer commitments above $40B, and the $40–50M per MW auction figure against a $12M 2026 base are from our Q2 2026 capture. The IREN–Microsoft contract terms — five years, $9.7B, 200 MW of critical IT load at Childress — and IREN secured power above 4.5 GW are from our February 2026 coverage. The Nebius Vineland figures ($17.4B base over five years, campus designed for ~300 MW) reach us through press reports of the September 2025 announcement rather than a filing, and some later commentary puts the campus at 300–350 MW; at 350 MW the same arithmetic gives $49.7M per MW over the term and $9.9M per MW per year, which does not change any conclusion here. Per-year figures for both Microsoft contracts are the headline value divided by the stated five-year term, not a company-disclosed ACV. The ranking's own row values are quoted as posted; we have not been able to source the numerator or denominator behind any row other than CoreWeave's.
