Miles Ashworth: You look like someone who's been staring at a press release for too long.
Megan Skiendel: Palantir Technologies dropped five sovereign AI customer announcements in a single week. McCarthy Building Companies, L3Harris, NVIDIA, Ondas, Tampa General Hospital — all at once, Q3 earnings landing around November second. And the McCarthy number is the one I can't shake.
Miles Ashworth: Thirty thousand hours. Yes.
Megan Skiendel: Thirty thousand annual hours on contract workflows, using Palantir Foundry with NVIDIA Nemotron models fine-tuned on over a hundred thousand production RFIs from actual McCarthy contracts. The training data is specific. The output number — no methodology, no verification — is suspiciously clean.
Miles Ashworth: Look, the training data I'd accept — construction companies track labor by the minute, a hundred thousand RFIs is a plausible corpus. But thirty thousand hours is exactly the kind of figure that lives between 'we measured this' and 'we need a number for the deck.'
Megan Skiendel: Which is what makes the timing so loaded. Five customers, one week, one earnings date.
Miles Ashworth: Well — is this what genuine enterprise AI adoption looks like when it finally arrives, or is this the version of adoption that Palantir needs analysts to believe in right now?
Megan Skiendel: That's what I can't resolve yet. And I don't think the thirty thousand hours is where you find the answer.
Miles Ashworth: Where you do find it, though — actually, let me try to make this concrete. What McCarthy did is not complicated. You take a hundred thousand real contract requests, RFIs that their own teams wrote and approved and litigated over, you feed those into an NVIDIA Nemotron open-weight model running on Palantir Foundry, and you get something that has read every deal McCarthy ever signed before it touches its first new one. It's a paralegal who spent a decade in your firm's own archive. Not a general paralegal — yours. That's the click.
Megan Skiendel: And a hundred thousand production RFIs is not a pilot. That's operational depth.
Miles Ashworth: Which is why I half-believe the training story. The thirty thousand hours, though — that's slippery.
Megan Skiendel: Picture a contracts manager at McCarthy, Tuesday morning, forty-million-dollar subcontractor agreement in front of them. Pre-deployment, that review is four hours. Post-deployment, ninety minutes. McCarthy calls that saved time. But saved against what outcome exactly?
Miles Ashworth: That's — yes. Because thirty thousand hours at, say, a hundred and fifty dollars fully-loaded cost per hour is a four-and-a-half million dollar efficiency claim. Which is a very tidy number for an earnings deck.
Megan Skiendel: And Palantir's press release does not tell you whether McCarthy's legal headcount shrank by a single person.
Miles Ashworth: No. And those are genuinely different things — one is a workforce story, the other is a productivity story — and Palantir is quite carefully selling the number without specifying which story it is.
Megan Skiendel: Which means the thirty thousand hours is real evidence of something. We just don't know of what yet.
Miles Ashworth: Which points us straight at the stack itself — because the McCarthy story, the L3Harris story, Tampa General, Ondas, even NVIDIA as a customer — they all land on exactly the same two names. Palantir AIP or Palantir Foundry, NVIDIA Nemotron, NVIDIA hardware. Every single one.
Megan Skiendel: Every one. And that's the thing about the sovereignty pitch — it's supposed to mean your data stays on your infrastructure, your model is yours, nothing routes through someone else's cloud. Tampa General is the clearest case. Patient data, HIPAA, they've been a Palantir customer since 2021, so this isn't a new logo — they fine-tuned Nemotron locally, classified their own LLM traffic against their own privacy standards. That 7% model performance improvement is actually — honestly, it's evidence it worked.
Miles Ashworth: Seven percent, though. In a hospital's own sandbox. That's not a number you release if you have a bigger one.
Megan Skiendel: Which might mean it's honest. NVIDIA is on the partner list and the customer list simultaneously — they're a named sovereign AI customer of Palantir. That's structurally weird.
Miles Ashworth: Good lord. And nobody's explaining that dual role.
Megan Skiendel: No. And Jensen Huang is out there publicly saying every country should build its own AI infrastructure — which, look, that's a sovereignty argument that also happens to be a hardware sales pitch. So the executive-level air cover for this entire narrative comes from the man who sells the silicon.
Miles Ashworth: You've swapped Azure for a two-vendor dependency and called it independence. Ondas launching Regna AI on Palantir AIP with Nemotron for 'mission-critical government operations where data control is non-negotiable' — that's the pitch word-for-word. Non-negotiable. And yet the model weights are NVIDIA's, the platform is Palantir's, and there's no independent third party verifying any of it actually constitutes sovereignty.
Megan Skiendel: Tighter lock-in dressed as liberation.
Miles Ashworth: Quite. And the question — the one we don't have the answer to yet — is whether Q3 earnings on November second actually convert this into numbers that justify where Palantir's stock is sitting. Because that Tampa General 7% figure might be telling us something very specific about the ceiling in regulated industries, and I don't think we've pulled that thread fully.
Megan Skiendel: The real ceiling question. Whether sovereignty — as Palantir is packaging it — is structurally constrained to incremental gains in any regulated environment. Yeah. That's the one.
Miles Ashworth: And that ceiling is precisely what the November second print has to answer — because the stock sitting at a hundred and eighty-six dollars, down seven percent from mid-October, roughly twelve-and-a-half percent off the fifty-two-week high of two-twelve ninety-four — that's not a company being punished. That's a market saying: prove it converts.
Megan Skiendel: And the conversion bar is high. Q2 revenue was one-point-nine-four billion. Palantir's own Q3 guidance is two-point-one-six billion.
Miles Ashworth: Which implies — what, eleven percent sequential growth? In one quarter?
Megan Skiendel: And analysts are expecting EPS around thirty-four cents, which is roughly eighty-nine percent year-over-year growth. Palantir has beaten Wall Street estimates four quarters running, so there's a track record. But Goldman at two-thirty, Baird at two-fifty, Mizuho at two-forty-five — those targets all assume the growth rate actually holds.
Miles Ashworth: Which brings us back to Tampa General's seven percent. Because — frankly, no methodology disclosed, no benchmark definition — that number is either the most honest thing Palantir has released this cycle, or it is the ceiling of what regulated-industry sovereign AI can actually do inside a compliance sandbox right now.
Megan Skiendel: Those are actually two different problems. Deployment depth — Tampa General fine-tuning Nemotron on live patient data locally, that's real — and commercial velocity are not the same thing. One proves the use case works. The other is what Goldman's two-thirty target requires.
Miles Ashworth: Right — but the seven percent is the only independently-described outcome number from any of these five customers. McCarthy's thirty thousand hours: no verification. L3Harris: announced, nothing disclosed. Ondas, Regna AI: barely launched. Seven percent from Tampa General is literally all the performance data we have.
Megan Skiendel: Which means November second isn't testing whether sovereignty works technically. It's testing whether a market priced for eighty-nine percent EPS growth believes five announcements and one seven-percent data point constitute a pipeline.
Miles Ashworth: And the answer — well, we don't have it yet. But the question is now much more specific than it was a week ago.
Megan Skiendel: The thing I keep sitting with — and I don't have a clean answer — is that the customer stories are specific enough to be credible. McCarthy's hundred thousand RFIs, Tampa General's privacy classifier running locally, Ondas launching Regna AI on Palantir AIP, L3Harris on the platform. That's not vague. But specificity of deployment and acceleration of revenue are not the same variable, and I'm not sure the market has fully separated those two things yet.
Miles Ashworth: No. And November second doesn't answer whether sovereign AI is real. It answers whether it's fast. Those are genuinely different questions, and I suspect Palantir would quite prefer everyone conflate them for another quarter.
Megan Skiendel: Which is an uncomfortable place to land, honestly. Because I think the use case works. I think Tampa General's classifier works, I think McCarthy's contract review works. I'm just not certain it compounds at the rate Goldman's two-thirty target requires.
