Megan Skiendel: David, long week — but this one landed and I couldn't put it down.
David Sterling: The Nvidia-NPS thing. I know. Let me just state the facts cold, because they're strange enough without editorializing — July 22nd, 2026, Jensen Huang personally attends a commissioning ceremony at Naval Postgraduate School. Nvidia donates a DGX GB300. NPS calls it, quote, the most powerful supercomputer in the Pentagon's inventory.
Megan Skiendel: The most powerful. In the Pentagon's inventory.
David Sterling: That's what they said. Seventy-two Blackwell Ultra GPUs in a single liquid-cooled rack. And the estimated value is fifteen million dollars — a number that, frankly, neither Nvidia nor NPS will actually confirm.
Megan Skiendel: Both sides staying quiet on the number — that's not an accident. Honestly, both institutions benefit from the ambiguity. Nvidia gets the relationship optics without the scrutiny a confirmed price triggers. NPS gets to say they have a fifteen-million-dollar asset without anyone asking hard questions about why it came through the NPS Foundation instead of a procurement contract.
David Sterling: The nonprofit routing is the structural tell. Standard defense procurement has IP carve-outs, audit visibility, Congressional notification thresholds. A foundation donation has none of that.
Megan Skiendel: And Admiral Samuel Paparo was standing right there at the ribbon-cutting — INDOPACOM's commander. That's not a photo op.
David Sterling: So the question is whether this is a donation or a procurement that needed a different door.
Megan Skiendel: But here's what the 'different door' framing still misses — the CRADA was announced in December 2024. The relationship was already legally structured six months before anyone cut a ribbon. The hardware is almost incidental at that point.
David Sterling: Wait — the Cooperative Research and Development Agreement predates the donation by how long?
Megan Skiendel: Seven months, roughly. The CRADA is the actual legal spine. The DGX GB300 and the ceremony and Huang on stage — that's the visible layer. The agreement is what locks in joint research rights, access terms, all of it. Before a single GPU shipped.
David Sterling: And the NPS Foundation as legal recipient means the IP ownership question — who owns what comes out of joint research — that doesn't get resolved by the CRADA automatically. That's a gap.
Megan Skiendel: It's not a gap accidentally. Look — think of it like a tech company donating its entire software suite to a law school. The students graduate, they've only ever worked inside that system, and they spend the next thirty years buying that company's products because it's the baseline they know. That's the mechanic here. Nvidia Mission Control runs on the GB300 — the software dependency isn't separable from the hardware.
David Sterling: The Pentagon has broken acquisition timelines — I'll give you that. Foundation structures move faster. Couldn't this just be pragmatic workaround with known tradeoffs?
Megan Skiendel: Honestly, yes — the Pentagon has quietly encouraged these foundation routes precisely because DFARS procurement can take years. I'm not disputing the speed rationale. But the ambiguity on congressional visibility, IP carve-outs — that's not a side effect of moving fast. Both sides chose not to resolve it. That's a feature.
David Sterling: So the headline is 'donation.' The actual mechanism is a CRADA wrapped in a nonprofit transfer that does specific legal work a procurement contract wouldn't allow. Those are different things.
Megan Skiendel: That's the one clean sentence. The donation is real. The structure around it is doing a lot more work than the word 'donation' suggests.
David Sterling: And that structure is doing its work right now — but everyone's writing this up as a Huang PR move. A branding exercise for Nvidia's national security positioning. That's underselling it by a lot.
Megan Skiendel: Because Paparo was there.
David Sterling: INDOPACOM's commander. That's not — look, Paparo doesn't attend hardware ribbon-cuttings. That is not what the commander of U.S. Pacific Command does with his calendar. His presence is a deliberate signal to the defense procurement ecosystem that this relationship has command-level endorsement. That's a different message than 'Nvidia supports the troops.'
Megan Skiendel: It's telling the acquisition officers in that room who their chain of command thinks they should be working with.
David Sterling: Exactly. And then — wait, this is the part that hasn't been named yet — Nvidia is under active DOJ export control scrutiny at the same moment they're staging this ceremony. A White House official confirmed Moonshot AI, a Chinese company, accessed Nvidia chips despite the export ban. So the geopolitical AI gap Huang invoked in his remarks at NPS — 'nothing more valuable than information and insight' — that gap is narrower than the ceremony implied.
Megan Skiendel: Hold on — Moonshot AI specifically?
David Sterling: Named by a White House official. So Huang is on stage invoking national security urgency while Nvidia's own supply chain has already leaked to a Chinese firm the controls were meant to stop.
Megan Skiendel: So Huang is positioning Nvidia as the trustworthy domestic option precisely because the export control story is getting messy. The ceremony is — I mean, it's almost inoculation. Stand next to Paparo, say the right words about insight and national security, and the DOJ investigation becomes a footnote.
David Sterling: That's the play. And the dependency piece — we'll get to exactly how hard it becomes to reverse once NPS graduates are writing RFQs benchmarked against Blackwell — structurally uncomfortable is what we're looking at.
Megan Skiendel: And that's the dependency — that's where the lock-in actually bites. Because the use cases at NPS aren't narrow. Weather prediction, cybersecurity, oceanic modeling, naval digital twins, disaster resilience — I mean, that's not one mission area. That's coverage across almost every domain NPS faculty and students touch.
David Sterling: Seventy-two Blackwell Ultra GPUs and thirty-six Grace CPUs. Twenty terabytes of GPU memory at bandwidths up to 576 terabytes per second. That's genuinely powerful. But it's one rack. At a school.
Megan Skiendel: Which has fifteen hundred in-resident students and six hundred faculty. That's the point — you don't need to put it in a carrier group. You put it where the people who will eventually spec carrier group AI infrastructure do their graduate work.
David Sterling: And Nvidia Mission Control is the orchestration layer running on top. So the workflow familiarity isn't just with the hardware architecture — it's with Nvidia's specific software stack.
Megan Skiendel: Right — but the verification problem is actually worse than the lock-in problem, at least in the short term. NPS called this the most powerful supercomputer in the Pentagon's inventory. And we can't check that claim. Classified assets exist. Nobody outside a very short list knows what DoD actually runs.
David Sterling: So the prestige claim is unverifiable by design.
Megan Skiendel: Exactly. And it does real work — a research institution holding the most powerful unclassified system in the building is a recruiting and grant signal, not an operational one. This isn't forward military AI infrastructure. It's a very expensive, very prestigious research tool. Those are different things and the press coverage isn't distinguishing them.
David Sterling: What to watch, then. Whether DoD RFQs start referencing DGX architecture as baseline spec. That's the canary. The moment acquisition language defaults to Blackwell as the benchmark everything else gets evaluated against — that's when this donation converts to revenue.
Megan Skiendel: And whether the CRADA produces classified research. Because if it does — honestly, that's an IP dispute waiting to happen. Who owns a model trained jointly on a donated rack, under a nonprofit transfer, inside a government institution? Nobody has answered that cleanly, and nobody seems eager to.
David Sterling: The IP question won't leave my head. NPS is the first U.S. military command to deploy the GB300. That's an institutional first-mover precedent — and nobody has publicly answered who owns what gets built on that machine. Nvidia, the Foundation, the government. That's not a detail.
Megan Skiendel: Nobody's eager to find out, either. And that's — I mean, that's almost the answer. The ambiguity is load-bearing. The moment you resolve it, someone loses something they currently think they have.
David Sterling: If NPS graduates become the acquisition officers writing the next decade of AI infrastructure RFQs, benchmarked against Blackwell because that's what they trained on — the Pentagon may have built a single-vendor dependency willingly. And then one day someone in acquisition will notice. And the leverage question gets very uncomfortable very fast.
Megan Skiendel: Yeah. I don't have a clean answer to that.
David Sterling: Neither do I. Good conversation.
