Nebius: The Demand Is Real. What Reaches Each Share?

Our view

Nebius is a useful test of our philosophy: the business, expectations and the stock can move in different directions. We see credible demand for AI infrastructure. We have less confidence in the value that expansion will deliver to each shareholder, and low confidence in a five-session directional forecast.

Our central hypothesis is that slower frontier-model development need not produce an equal slowdown in customers running useful AI applications. That distinction could matter for infrastructure demand. It does not insulate Nebius from weaker pricing, delayed projects or higher financing costs. We want to follow the path from customer commitments to delivered capacity, collected cash and diluted ownership. A large contract or a sharp share-price decline settles only part of that investigation.

Starting market state

This is a September 14 premarket baseline, with completed-session market data through Friday, September 11. Yahoo showed NBIS at $206.85 at 07:46:31 Eastern, down 7.88% from Friday's $224.55 close. It was third in the Yahoo US trending feed retrieved at 07:42:53 Eastern; that is a vendor attention ranking, not measured sentiment or proof of institutional interest. [1,2]

Using Yahoo daily closes, NBIS was already down 11.95% over the 20 trading-session intervals ending September 11, versus a 2.35% decline for QQQ. Friday volume was about 11.18 million shares, 0.58 times the preceding 20-session average. The starting condition is an active stock that had already weakened, not an untouched growth narrative suddenly encountering its first doubt. [3,4]

What changed

Dario Amodei's September essay advocates coordinated pacing of frontier capability development and external evaluators. This is a policy proposal and an operating-risk argument; it is not a disclosed Nebius customer cancellation. Our earlier weekly view covers the wider debate. We do not assume every AI company, investor or government shares one position. [5,6]

Nebius and Palantir announced their partnership on September 8, before this morning. After integration, eligible Palantir commercial customers are intended to access Nebius compute and inference within Palantir's enterprise environment. The release does not quantify a minimum revenue commitment. We see a plausible distribution channel for application workloads, with adoption still to demonstrate. [7]

The August 12 shareholder letter says Token Factory inference workloads more than tripled in Q2. It also describes short contracts of three to six months, core contracts averaging one to three years, and longer arrangements with investment-grade customers. Different contract durations mean different exposure to a change in demand. These are management's disclosures, not independent verification of future utilization. [8]

Expectations

Q2 group revenue was $582.3 million, up 454% year over year. Adjusted EBITDA was $236.2 million, while GAAP operating loss was $175.9 million. For the first half, operating cash flow was $4.504 billion and purchases of equipment and intangibles were $8.130 billion: a $3.626 billion shortfall before other investing activity. The operating cash-flow statement included a $4.395 billion increase in deferred revenue. Customer funding helps finance construction; it is not all cash earned from services already delivered. [9]

The June filing reports $37.491 billion of remaining performance obligations, with 36% expected to become revenue within 24 months. This measure excludes contracts originally lasting one year or less and includes estimates subject to performance and timing constraints. It is neither annual revenue nor an unconditional cash balance. [10]

August's completed financing added $5.75 billion of convertible notes. A separate exchange retired $800 million of older notes in return for approximately 15.8 million Class A shares. The new notes carry 0.50% and 4.50% coupons and mature in 2030 and 2034; their initial conversion prices were about $313.46 and $324.65. Those conversion prices are contract terms, not price targets. [11,12]

Our illustrative ownership bridge starts with 271.855 million June-end ordinary shares, adds 21.066 million pre-funded warrant shares and approximately 15.8 million exchange shares: about 308.721 million economic shares. The warrants were already included in reported EPS calculations; we add them to issued shares, not to an EPS denominator. At the captured $206.85 price, this basis corresponds to roughly $63.86 billion of equity value. It excludes other subsequent issuance, employee awards and remaining convertible dilution, so it is not a verified current fully diluted market cap. [1,9,10,11]

To make the valuation question concrete, assume eventual annual revenue of $8 billion. At 10%, 20% or 30% sustainable cash margins after interest, tax and all required reinvestment, annual cash for equity would be $0.8 billion, $1.6 billion or $2.4 billion. Against that illustrative equity value, the yields are approximately 1.25%, 2.51% and 3.76%. These are our sensitivities, not guidance, consensus or forecasts. Delay reduces present value, and financing must support the path to that cash generation. We have not verified a current consensus distribution or built a complete DCF.

Observed response

At the same 07:46:31 Eastern timestamp, CoreWeave was $82.35, down 7.46% from Friday. Similar declines are consistent with a shared infrastructure repricing, but do not establish its cause or identify who traded. Monday's regular session has not yet occurred in this baseline. [1,13]

We will separate an opening gap, the closing response and subsequent sessions. A rebound would not itself verify our operating thesis. Continued weakness could reflect deteriorating economics, a higher required return, or both. We cannot label it forced selling from price data alone.

Comparable history and its limits

Nebius's own recent history is informative without supplying reliable odds. It closed at $193.23 before the August 12 earnings release and $259.20 on release day, a 34.14% rise. It closed at $243.88 on September 8, the Palantir announcement date, then $224.55 on September 11, a 7.93% retreat. These are observed event windows with overlapping influences, not isolated causal experiments. [3]

This illustrates the question we want history to sharpen: how much improvement can a price already anticipate? ChartLibrary returned no eligible comparison packet for the explicitly requested September 11 NBIS state. Underlying bars and state rows exist, but that does not supply usable analog results. We therefore report no historical win rate, average forward return or directional inference from matches. [14]

The countercase

The strongest objection to our cautious valuation stance is that attractive contracted economics could make today's funding burden temporary. Management describes stronger deal pricing, customer prepayments and a shorter expected payback period on Q2 contracts. Those estimates include forecast costs and capacity not yet built. If delivery and cash conversion follow, focusing only on current investment outflows would understate the business being created. [8]

The strongest objection to our constructive demand stance is that inference growth may not offset weaker frontier spending. Cheaper computation could expand usage while reducing revenue per unit. Customers might retain bargaining power as capacity becomes available elsewhere. That is an economic scenario to test, not an observed cancellation.

Concentration and commitments matter: three customers represented 59% of Q2 revenue, and the June filing disclosed approximately $12.05 billion of future undiscounted payments on leases not yet commenced. That lease figure is not all current debt or immediately due, but it makes the timing of monetization consequential. We cannot assign anonymous customer labels to specific companies. [10]

What would change our view

We would strengthen the demand view after evidence of capacity accepted on schedule, contracted service commencement and sustained customer usage at economically attractive prices. Quantified adoption through the Palantir partnership would be more informative than another broad partnership announcement.

We would weaken it after documented delivery delays, contract reductions, unfavorable renewals or utilization declines. We would become more constructive on valuation after a credible reconciliation from operating earnings through recurring reinvestment to cash per diluted share. Higher projected revenue accompanied by a larger funding gap or share count would not, on its own, meet that test.

For the next five sessions, our market observation is narrower: does NBIS stabilize relative to QQQ and CoreWeave after this premarket reset? Relative stabilization is a timing observation, not proof that the business or valuation case has improved.

What we will revisit

Append the September 14 closing response once completed data are available. Then revisit at the September 18 close, the fifth session counting September 14, and September 25, the tenth. Use September 11 closes as the preserved pre-week baseline and report NBIS, QQQ and CoreWeave over identical windows. Keep any later company disclosure distinct from the original information set.

At the next financial update, check delivered capacity, revenue quality, customer funding, capex, obligations and diluted ownership together. These are research observation windows, not instructions to enter or exit a position. No simulated position is created by this article.