A partnership announcement with no number in it pulls every reader toward the same question — what is it worth? — and this morning that question has no answer, because the two companies did not put one in the document. Asking it anyway is how a press release turns into a valuation.
The answerable question is narrower. Before a deal can be worth anything it has to land somewhere inside a business, and the two ends of this one land in very different places. One end sits on a seller that says it is already turning demand away. The other sits on the only part of a business that publishes the unit the deal would move. Only one of those ends can be measured at all, and it is not the one the news is written around.
So: what the document actually says, why the obvious instrument does not fit the seller, where the deal lands on the buyer, and the one test a reader can run without a model.
The release names a relationship and prices nothing
Palantir and Nebius announced a strategic partnership at 06:59 New York time this morning, datelined Miami and Amsterdam. Palantir "has named Nebius its preferred sovereign AI infrastructure partner." Following an integration period, Nebius compute and inference endpoints come "inside the Palantir enterprise perimeter," so that eligible Palantir customers can deploy open models, adapt them on their own data, and keep control of the compute, the data and the models. Eligible, not all — and the release does not define the test. The two will also "work together to accelerate the deployment of new compute capacity," including "modular data-center deployments at sites where power is already available."
That is the whole commercial content. There is no dollar value, no annual or total contract value, no megawatts, no GPU count, no named sites, no length for the integration period, no minimum, no take-or-pay, and no term. The body is commercial-only; government is not mentioned on either side.
One number is already travelling with this story that does not belong to it. At least two write-ups pair the partnership with Nebius's more than $40 billion of customer commitments. That figure is a June-quarter disclosure, reported on 12 August. It was true in August, it is not a term of this deal, and nothing announced this morning added a dollar to it.
"Preferred" is not "exclusive," and nothing in the release says otherwise. The word appears once, in that clause, and the release is silent on whether Nebius sits alongside Palantir's existing hyperscaler relationships or in place of any of them. Anyone reading a displacement into this is supplying it themselves.
The only in-document evidence that a contract exists at all is each company's own risk language. Palantir's forward-looking block refers to its expectations regarding "the amount and the terms of the contract," and to "customers' ability to modify or terminate the contract." Nebius's names the risk to "our ability to expand our client base or revenues through our status as a preferred sovereign AI infrastructure partner to Palantir." Boilerplate that presumes a contract, on both sides, with the amount and the terms withheld. As of this writing neither company has filed anything on it — no Palantir 8-K, no Nebius 6-K — so this is an investor-relations and wire event, not a filed one.
The modular data-centre clause is the priceable sentence, and it has no number attached
One clause in the release could carry real money: the two companies accelerating new capacity "through modular data-center deployments at sites where power is already available."
We have published the discipline for pricing exactly that. Nebius disclosed in its June quarter that its four landmark deals carried annual contract value of $20–25 million per megawatt, and we used it three weeks ago to take apart a per-megawatt league table that mixed annual and cumulative rows. Apply it here: 100 MW at that rate is $2.0–2.5 billion a year, against a Nebius annualised revenue run-rate of $3.0 billion at the end of June. Two-thirds to five-sixths of the existing book, from one line in a press release.
Which is precisely why the absence of a megawatt figure matters rather than being a detail to work around. The conversion is instant once a number exists. This is the point where most write-ups supply one anyway. We are not going to, and the sentence is worth nothing until Nebius or Palantir attaches a figure to it.
The ceiling is at least bounded. Nebius has guided year-end connected power to 800 MW to 1 GW, so that is the whole fleet this clause could draw on. What share of it, if any, this arrangement touches is not stated, which leaves the plausible range running from a rounding error to something larger than the existing book.
For Nebius the usual instrument does not fit, because capacity binds
The reflex reading of a marquee customer channel is that it lifts the revenue curve — a step up in the base, compounding from there. Our Nebius model carries a field for exactly that, and it is tempting to reach for it: shift the base by 20% and our fair value moves from $136.85 to $164.79, up $27.94. Every figure in this paragraph is our estimate on our own published assumptions, and this is the only place they appear.
We are not going to use it, because our own published thesis says it does not apply. That model, struck on 27 August, describes a GPU cloud whose revenue is the delivery of connected, GPU-equipped power, where capacity is the stated constraint — the company says it could sell its whole 2027 book today and is choosing not to. A seller in that position does not convert a new demand channel into new revenue. It converts it into a different buyer at the front of the same queue, and possibly a different price.
This is an argument we already published in the other direction. When we priced the labs inside NVIDIA's revenue on the same day, the point was that when supply binds, a customer leaving the queue moves a sale rather than losing one. Run it backwards and a customer joining the queue does not create one. Same mechanism, same company constraint, opposite sign.
So the honest position on the Nebius side is that there is no reading here that is not already inside the growth assumption — and that assumption is doing almost all of the work. Move the single quarterly growth rate our model turns on — it assumes 57.5% sequential — from 45% to 75% and our fair value runs from $46.60 to $401.30, a factor of 8.6. On a business compounding at that pace a point estimate is a statement about the growth rate and very little else.
The test that needs no model is Nebius's own run-rate ladder
Nebius publishes the one series that will settle this without anybody's spreadsheet. Annualised run-rate revenue was $1.9 billion at the end of March and $3.0 billion at the end of June, and the company guided to exit 2026 between $7 billion and $9 billion. All three figures are the company's own.
That gap — $4.0 billion at the low end of the guide, $6.0 billion at the high end — is where a partnership either shows up or does not. If this deal is a real demand event, the guide moves or the exit number lands in the upper half of it. If the guide is unchanged at the next report and the exit lands where it was already going, the announcement was a relationship rather than a sale. A reader can check that from Nebius's own disclosure and nothing else.
Note the basis, because two numbers are circulating together. The partnership will be talked about against the run-rate figure and reported against quarterly revenue, which was $582.3 million in the June quarter. Those are not the same quantity, and the June print has both.
Palantir is the side with a published volume unit
Now the other end. Palantir's US commercial business is the one part of the company that discloses the thing a deal like this moves.
From the June-quarter call, all of it the company's own: 653 US commercial customers at 30 June, a count up 35% year over year and 6% sequentially, against revenue of $764 million, up 149% year over year and 28% sequentially. That is about $1.17 million per customer in the quarter — the division is ours, both inputs are Palantir's. Net dollar retention was 157%, up 700 basis points in a quarter. US commercial bookings on a total-contract-value basis were $2.132 billion, up 153%.
The gap between a 35% customer count and 149% revenue is the entire business. Palantir is not growing this vertical by adding logos; it is growing it by expanding accounts it already has. And what this deal does — let existing customers run open models on their own data inside their own perimeter — is an expansion-and-retention mechanic before it is an acquisition one. It lands on the number that is already doing the work.
And it lands there heavily, because that single vertical is 70% of our whole Palantir base case. Our model values the company at $166.13 a share; remove the US commercial vertical entirely and $50.49 is left. It is also the only one of the four boxes carrying a real driver, for the same reason the deal lands there — it is the only one with a published unit to drive.
The retention lever is the bigger one, and it is the one this deal touches
Inside that vertical there are two ways to move the number, and they are not the same size. Both of these are our assumptions, moved on our own published model:
- Add accounts faster — from 37 net new a quarter to 45 — and our fair value goes from $166.13 to $180.39, +$14.26; at 55 a quarter, $198.20.
- Expand existing accounts faster — revenue-per-customer drift from 16.5% a quarter to 19% — and it goes to $187.29, +$21.16; at 22%, $217.41.
- The same lever runs the other way. Drift at 14% a quarter instead of 16.5% takes it to $148.06, −$18.08.
The expansion lever is about 1.5 times the acquisition lever, and expansion is the one a sovereignty deal plausibly touches: it gives an installed account a reason to widen rather than giving a sales team a reason to win. We are not moving either. The point is the size and the direction of the box the news lands in, not a revised number — and our Nebius model carries no equivalent driver at all, which is why the two sides genuinely cannot be read with the same instrument.
The sovereignty language is a quarter old. The compute vendor is the new part
Nothing in Karp's quote this morning is new. On the 3 August call he already framed the entire quarter this way: what enterprises demand is "AI sovereignty, owning the operational definition of the data, logic, actions and security of their enterprise", and companies renting token-metered frontier models are "paying to give away their most important secrets." The CFO raised full-year US commercial guidance to in excess of $3.424 billion "on the back of this exceptional continued strength in the U.S. and accelerating demand for our sovereign AI capabilities." Asked on that same call what a customer does when the model it built on gets pulled, Karp described the mechanic this partnership serves: swapping one model for another "presupposes an open model and sovereignty."
So the framing was already load-bearing in the guide a month ago. What changed today is that Palantir named a compute vendor for it.
No one has traded this yet
Worth stating plainly, because every ticker card carrying a price this morning is showing a stale one. US Labor Day 2026 fell on Monday 7 September — the first Monday of the month — and the exchanges were closed. The last regular cash session was Friday 4 September, four calendar days before this announcement existed. At that close Nebius was $226.39 and Palantir $174.33. This piece was finished before the US open, so those are the last prices either side of this deal has printed, and neither of them contains a single share traded on the news.
For orientation only: our Nebius model was struck on 27 August against a reference price of $213.93, and the 4 September close is 5.8% above it, eight days later. That same Friday close on Palantir is 4.9% above our $166.13 base case. Neither gap is an argument; both are the distance between a dated model and a dated tape.
One live setup, without a prediction attached. Nebius has 12.3 million shares of at-the-market capacity left after selling $2.8 billion at a weighted average of $223.60 through 30 June, and we flagged that open question as the first thing to watch. A marquee partner announced into the first session after a holiday weekend, with the stock above the last raise price, is the setup that question named. No filing appeared today. That is a setup, not a forecast.
What to watch
- A Palantir 8-K or a Nebius 6-K with economics. Palantir's own risk language refers to "the amount and the terms of the contract," which is the reason to look. Nothing has been filed. If 48 hours pass with no filing, this was an investor-relations event and it should be read as one.
- Any megawatt or dollar figure on the modular data-centre clause. At Nebius's disclosed $20–25 million of annual contract value per megawatt, 100 MW converts to $2.0–2.5 billion a year. Until a capacity number exists the clause prices at nothing.
- The exit-2026 run-rate guidance of $7–9 billion, against $3.0 billion at 30 June. Moved or reaffirmed at the next Nebius report, this is the whole test and it needs no model.
- The length of the integration period, and which Palantir products beyond the sovereign stack are in scope. Both undisclosed, and both bound how soon any of this touches revenue.
- Palantir's US commercial net dollar retention next quarter, from 157%. That is the number this deal would move first if it moves anything, and it is reported every quarter.
The partnership terms, quotes and risk language are from the joint Palantir–Nebius release of 8 September 2026, carried on both companies' investor-relations sites; no filing on the transaction had appeared at either company at the time of writing. Palantir's customer count, revenue, bookings, net dollar retention and guidance are as reported for the June quarter on its 3 August call. Nebius's run-rate revenue, exit-2026 guidance, quarterly revenue, per-megawatt annual contract value, year-end connected-power guidance, the more than $40 billion of customer commitments, at-the-market share sales and remaining capacity are its own June-quarter disclosures. Revenue per customer of about $1.17 million is our arithmetic on Palantir's two disclosed figures. Every fair value, sensitivity and share of the base case here is our estimate on our own published assumptions — for Palantir struck 24 August, for Nebius 27 August — and none of them is a forecast of this deal. Prices are the 4 September close, the last regular US session before this piece.
