Bernstein puts Nebius’s annualized revenue per active megawatt at $11.1 million in the second quarter of 2026 and Riot’s at $1.3 million, according to an exhibit from its note on OpenAI’s compute commitments, as relayed on X on 4 November by the X account of Exencial Research Partners. We have not read the note.
By our arithmetic from the exhibit’s printed figures, the three neoclouds it lists average about $9.9 million per megawatt, about 6.3 times the roughly $1.6 million average of the four companies it groups as emerging AI infrastructure.
Business Pill · Take-or-Pay
An evergreen Business Engineer explainer on take-or-pay contracts. It teaches the general idea only and does not discuss Bernstein, OpenAI or the companies in this piece.
The key insight: On a Bernstein exhibit relayed on X, the three neoclouds it lists average about $9.9 million of annualized revenue per active megawatt, about 6.3 times the roughly $1.6 million of the four emerging AI hosts, by our arithmetic; the same post puts Bernstein’s model of OpenAI’s infrastructure payments through 2030 at about $614 billion.
The exhibit is titled “Annualized Revenue per Active MW, 2Q26 ($M)”. It lists CoreWeave at $8.2 million, Nebius at $11.1 million and IREN at $10.4 million as neoclouds; Digital Realty at $3.1 million, Equinix at $4.4 million and CoreSite at $3.5 million under colocation; and Cipher at $1.8 million, TeraWulf at $1.8 million, Core Scientific at $1.4 million and Riot at $1.3 million as emerging AI infrastructure.
A footnote says “emerging AI & IREN figures are calculated on a per-contract basis”, so those bars are not measured the same way as the others. The source line reads “Company filings, Bernstein analysis and estimates”.
Another exhibit in the same image is headed “Spreading Oracle’s revenue from OpenAI we get ~$300B in revenue of which 57% is captured through FY30”. By our arithmetic, that is about $171 billion through Oracle’s fiscal 2030 and about $129 billion after it.
Annualized revenue per active megawatt in the second quarter of 2026 for ten AI infrastructure companies, as printed on a Bernstein exhibit relayed on X by Exencial Research Partners; not independently verified. IREN and the emerging-AI figures are per contract, per the exhibit’s footnote. The group averages and the ~6.3x ratio are our arithmetic.
What Was Posted on X
🚨 BERNSTEIN JUST DROPPED A MAJOR REPORT ON THE $750 BILLION AI INFRASTRUCTURE BOOM.
After Nomura argued that AI data centers are far more profitable than investors think, Bernstein tackles an even bigger question:
Can OpenAI actually afford its massive compute commitments and who ultimately captures the profits
The findings are fascinating:
• $750B vs. $614B: Bernstein models ~$614B in OpenAI infrastructure payments through 2030, with annual spending potentially peaking at $184B in 2031. Not every announced commitment is binding, and actual spending could be lower.
• OpenAI’s revenue is catching up: Bernstein cites a ~$50B annualized revenue run rate in September 2026, growing ~77% YoY. If growth remains strong, revenue could scale faster than infrastructure obligations.
• Oracle’s $300B opportunity: Capacity comes online in stages, with Bernstein estimating only ~57% of modeled OpenAI contract revenue recognized through Oracle FY2030. The backlog represents a multiyear revenueopportunity, not immediate cash generation.
• Microsoft and Amazon have structural advantages: Their integrated cloud platforms, enterprise ecosystems and AI services offer more durable economics than simply renting GPUs.
• CoreWeave faces a different challenge: Bernstein is comfortable with existing contracts but questions renewal economics once compute becomes less scarce.
• The biggest risk may be oversupply, not OpenAI defaulting. Bernstein sees potential supply-demand normalization within 2–3 years, putting pressure on GPU rental prices and weaker infrastructure providers.
HERE’S THE REAL INSIGHT:
Nomura explains why AI infrastructure can remain profitable despite negative free cash flow.
Bernstein explains why that profitability won’t necessarily be distributed equally.
The first phase of the AI boom rewarded companies that could secure GPUs, power and data-center capacity.
The next phase may reward companies that control customers, platforms and pricing power.
AI demand can keep exploding while returns on undifferentiated compute infrastructure deteriorate.
Being bullish on AI infrastructure spending is NOT the same as being bullish on every AI infrastructure stock.
That distinction could define the next leg of the AI trade.
$MSFT $ORCL $AMZN $CRWV $NVDA $AMD $MU
— Exencial Research Partners (@exencial_RP) Nov 04 10:11
Exencial Research Partners posted the Bernstein exhibits at 10:11Z on 4 November, quoting its own post on a Nomura report from the day before; by 17:00Z it had about 34,100 views and 311 likes. Its closing lines, that being bullish on AI infrastructure spending is not the same as being bullish on every AI infrastructure stock, are the account’s reading, not Bernstein’s.
The $614 Billion and the $750 Billion
The post says Bernstein “models ~$614B in OpenAI infrastructure payments through 2030”, against the $750 billion in its heading, and that annual spending could peak at $184 billion in 2031. It adds: “Not every announced commitment is binding, and actual spending could be lower”.
The post also says Bernstein “cites a ~$50B annualized revenue run rate in September 2026, growing ~77% YoY” for OpenAI. By our arithmetic, the $184 billion peak year is about 3.7 times that run rate.
On the other layers, the post says Microsoft and Amazon have “structural advantages”, that Bernstein “is comfortable with existing contracts but questions renewal economics once compute becomes less scarce” for CoreWeave, and that Bernstein “sees potential supply-demand normalization within 2–3 years”.
What Happened Since Our First Read
In Bernstein Sees OpenAI Infrastructure Spending Peak at $184B we summed our read of the 2025-2030 bars in Bernstein’s Exhibit 1 to about $610 billion. The relayed figure of about $614 billion is about $4 billion higher, by our arithmetic; that piece’s figures were approximate reads of the chart, and the two figures are of the same order.
As we read the exhibit, Bernstein sorts the companies that host OpenAI-era compute into three layers, and revenue per megawatt falls at each step from neoclouds to colocation to the emerging AI hosts.
As we read the Oracle exhibit, the ~$300 billion is spread over a decade: 57% through fiscal 2030, so about $129 billion falls after it, by our arithmetic.
“Spreading Oracle’s revenue from OpenAI we get ~$300B in revenue of which 57% is captured through FY30”
Three Implications
FOR NEOCLOUDS The exhibit puts CoreWeave, Nebius and IREN at $8.2 million to $11.1 million per active megawatt; the post says Bernstein questions renewal economics for CoreWeave once compute becomes less scarce.
FOR EMERGING AI HOSTS Cipher, TeraWulf, Core Scientific and Riot sit at $1.3 million to $1.8 million per megawatt on the exhibit, measured per contract, per its footnote.
FOR ORACLE The exhibit spreads about $300 billion of OpenAI revenue to fiscal 2034, with 57% through fiscal 2030.
The framework says the AI economy is a layered industrial stack and that, at any given moment, “the binding constraint sits at a different layer than it did six months ago.”
As we read it, Bernstein’s exhibit, as relayed, gives one measure of how revenue differs by layer of the hosting stack. The framework predates the Bernstein note and does not discuss it.
We have not seen the Bernstein note, its analysts, its date or its assumptions. The figures come from one X post and the exhibit image it carries, and we have not verified them independently. The revenue-per-megawatt bars mix per-contract and company-level measures, as the footnote says, and the group averages are ours, not Bernstein’s.
The exhibits are titled as Bernstein estimates; the post’s own commentary, including its closing lines, is the account’s reading, not Bernstein’s. We make no forecast about rental prices, renewals or OpenAI’s payments.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
On Bernstein’s exhibit as relayed on X, Nebius books $11.1 million of annualized revenue per active megawatt and Riot $1.3 million, and the neoclouds average about 6.3 times the emerging AI hosts, by our arithmetic. The same post puts Bernstein’s model of OpenAI’s infrastructure payments through 2030 at about $614 billion.
A note on sourcing. All Bernstein figures come from an X post by Exencial Research Partners on 4 November 2026 and the exhibit image it carries; we have not read the Bernstein note and have not verified the figures independently. Group averages, ratios and the Oracle split are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
