Update, September 17, 2026: the CoreWeave paragraph below is wrong. Besides its $40.00 listing, CoreWeave had already published four per-share marks: conversion prices of $107.80 and $119.60 on its 2031 and 2032 notes, NVIDIA's $87.20 a share in January and Jane Street's $109.00 in April. The Jane Street sale is the $997 million June-quarter placement described here as carrying no price, so the second item to watch had already been answered. We corrected it when CoreWeave launched its third convertible.
A remark travelled on 27 August. Nikesh Arora, chairman and chief executive of Palo Alto Networks, was reported to have said that in two years from now you will be able to buy a neocloud for less than it raises at today — and to have set Nebius apart as a different tier of operator that will actually reach scale.
That is the weakest kind of provenance this site handles, and it should be said before anything else. The remark reaches us through two accounts on X, hours apart: the first a summary of an image with no text in it, the second a sentence inside quotation marks. Neither names the venue. No transcript, filing or company statement carries it. And it appears nowhere in Palo Alto's own fourth-quarter earnings call five days later, on 1 September, where the words Nebius, CoreWeave and neocloud are all absent. Nobody outside the room can verify what Arora said, and nobody outside the companies he was talking about can verify what he meant by "raise at". We are not going to build an argument on it.
He is also not a neutral party. On Palo Alto's third-quarter call on 2 June he tied his own hardware demand to "the explosion of data centers being built, whether it's by hyperscalers, frontier labs, or neoclouds out there", and warned that when "data center growth starts to taper off or plateau, this thing is going to come to roost." A two-year call on neocloud valuations is a call on his own product cycle.
But the thing the remark is about is checkable today, and it does not need him. These companies publish the prices at which their capital is struck. The marks are in the filings. So rather than wait two years, put the marks next to the tape — which is the table at the top of this piece.
Four things get called a mark, and only one of them is a raise price
- Shares sold for cash. Nebius sold 12.7 million Class A shares at a weighted average of $223.60 through 30 June, roughly $2.8 billion gross, the first use of its at-the-market programme, with 12.3 million shares of capacity left. This is the only figure in this article that is literally a price a neocloud raised at.
- A negotiated strategic block. On 11 March 2026 NVIDIA bought pre-funded warrants over 21,065,936 Nebius Class A shares for $2.0 billion gross. That works out at $94.94 a share; the division is ours, the two inputs are Nebius's.
- A convertible's conversion price. Nebius priced $5.75 billion of notes on 19 August and closed them on 24 August — $3.45 billion of 0.50% notes due 2030 converting at about $313.46, and $2.30 billion of 4.50% notes due 2034 converting at about $324.65. IREN sold $3.0 billion of 1.00% notes due 2033 in May. Both companies raised at par. A conversion price is not a raise price; it is struck above the market on purpose.
- A warrant or right held by a counterparty. NVIDIA holds rights to buy 30 million IREN ordinary shares at $70.00, exercisable in tranches that vest on deliveries of up to 600,000 GPUs. Also not a raise price — a forward option on one.
Confuse the first with the third and the sector looks permanently expensive, because a conversion price is designed to sit above the share price on the day it is set. That is the trap in the claim, and it is the reason to do this with the marks separated rather than pooled.
Nebius publishes all four, and its ladder crossed the tape in August
No other company in this group discloses as much. Set Nebius's marks out in the order they were struck, against $226.39 at the 4 September close:
June 2025 converts at $51.45. September 2025 converts at $138.75. NVIDIA's block in March at $94.94. The March 2026 converts at $183.22 and $180.31. The June-quarter share sales at $223.60. Then August's pair at $313.46 and $324.65.
The ladder rises 6.3x in fourteen months — we set out the coupons and accretion behind it when the August deal was announced but not yet priced — and the crossover is the interesting part: every mark struck before August sits below today's price, and August's pair sits 27.8% and 30.3% above it. It is the first mark in the ladder the share price has not reached.
Before reading that as evidence of anything, look at what happened last time. The March notes were struck at premiums of 57.5% and 55.0% over a $116.33 close on 17 March. The share price passed both conversion prices inside five months. A strike above the tape is the structure of the instrument, not a verdict on the business, and anyone quoting August's $313.46 as proof that Nebius is 28% overvalued is quoting a premium as though it were a valuation.
Then look at what happened six days earlier. NVIDIA's $94.94 was set on 11 March, and it is 18.4% below that same $116.33 close. So within one week in March, two marks were struck on the same shares — one at an 18% discount to the market, one at a 57.5% premium — a spread of about 76 points between two numbers a league table would list side by side. The discount is the one attached to somebody actually buying a piece of a neocloud.
The one mark that is a raise price is the one that is level
Which leaves the at-the-market sales, and they are the whole test. Nebius sold $2.8 billion of actual shares at $223.60. At the 4 September close of $226.39 the stock is 1.2% above the average price of its own most recent equity issuance.
It has been on the other side. Our Nebius model of 27 August — the day the remark travelled — was built on a reference price of $213.93, which is 4.3% below that same $223.60. Whatever Arora meant, the gap he placed two years out had already opened, and closed again, inside three months, in public, on the same share class.
At IREN every recent mark is above the tape
IREN closed at $44.68 on 4 September. Its May notes were sold at a disclosed 32.5% conversion premium over a $55.15 reference price of 11 May, which puts the conversion price near $73.07 — that arithmetic is ours, the premium and the reference price are the company's. The stock is 38.9% below that strike, 36.2% below NVIDIA's $70.00 rights, and 19.0% below the $55.15 reference price the notes were marketed against four months ago. At 31 March the company had five convertible series outstanding struck between $13.64 and $85.63, so the tape now sits inside its own conversion stack rather than above it. It is not for want of selling stock: the June quarter alone carried $2,112 million of ordinary shares issued.
CoreWeave has not published a mark since it listed
CoreWeave listed on 28 March 2025 at $40.00 and closed at $89.36 on 4 September, 123% above the price it listed at — still the only per-share mark CoreWeave has ever published. Since then it has raised repeatedly, including $997 million of common stock issued in private placements in the June quarter alone, on top of $13.5 billion of gross new debt — and it has disclosed no price for any of it.
So CoreWeave's current mark cannot be compared with anything, in either direction. This is the point where most write-ups supply a number anyway. We are not going to: nobody outside CoreWeave knows what its most recent private placements were struck at, and a claim about whether the company will change hands above or below its marks is, for this name, unfalsifiable.
What our models do and do not say about this
We publish forward models for all three — Nebius, CoreWeave and IREN — and the only line in any of them that answers what would somebody pay for this business is a terminal exit multiple on revenue: 3.5x for Nebius, 3.0x for CoreWeave and IREN. Those are our assumptions, not disclosures, and nothing in a relayed remark supplies a figure that could revise them. Nor is our own fair value the answer to Arora's question: move the single quarterly growth rate the Nebius model turns on by 12.5 points in each direction and its fair value moves more than sixfold. On a business compounding at that pace, a point estimate is a statement about the growth rate and very little else — which is why this piece is built on disclosed marks and not on ours.
What to watch
- Whether Nebius uses the 12.3 million ATM shares it has left, and at what average. That is the next real raise price the sector will publish, and the only one directly comparable to $223.60.
- Whether CoreWeave ever discloses a placement price. Until it does, its marks cannot be checked at all, and every claim about them — bullish or bearish — is unfalsifiable.
- Whether NVIDIA's IREN rights vest. They turn on deliveries of up to 600,000 GPUs, not on the share price, so a $70.00 strike against a $44.68 tape is a statement about delivery schedules first.
- The first Nebius quarter that files the August notes. The carrying-value table will show the effective rate on paper struck at $313.46, and how much of the older 2029 and 2031 notes was exchanged for shares.
Arora's 27 August remark reaches us only through two posts on X — one a summary of an image, one a quoted sentence — with no venue named and nothing from Palo Alto Networks behind it. It is one executive's characterisation of other companies' valuations, and nobody outside those companies can verify it. His 2 June quotations are from Palo Alto's own third-quarter earnings call. Every mark, premium, reference price, conversion price, share count and issuance figure here is disclosed by the company it belongs to — in its filings, its pricing releases, its 2025 prospectus or annual report, and the June-quarter releases behind the release pages linked above. Ours rather than theirs: the $94.94 and $73.07 per-share figures, and every percentage comparing a mark with a price. Prices are the 4 September 2026 close and a live quote will differ. The $213.93 reference price, the exit multiples and the fair values are model assumptions, not company forecasts.
