Maya Chen: Nathan, hey — I have been sitting with this thing all morning and I cannot shake it. I was reading about Japan's new government AI push, this entity called Noetra, and they've been allocated ¥387.3 billion to build what they're calling a homegrown foundational AI model — like, sovereign, indigenous, theirs. And the first move they make is ordering 27,500 Nvidia chips.
Dr. Nathan Hayes: Rubin chips, specifically — next-generation Nvidia Rubin. Yes.
Maya Chen: Right, and that's — I mean, that's the whole knot we're pulling on today. Because Counterpoint Research published their Sovereign AI LLM Index on August 5, 2026, and the number that stopped me was 92.4%. That's the share of sovereign AI large language models that are trained or running on Nvidia semiconductors. 92.4. Of sovereign AI.
Dr. Nathan Hayes: And Marc Einstein at Counterpoint put it cleanly — LLM hosting is increasingly sovereign, the underlying chip source is not.
Maya Chen: Which is sort of — I don't know, it's almost a category error? Like, is the thing you're making independent if every single layer underneath it belongs to someone else? Saudi Arabia, the UAE, they're writing tens-of-billions-of-dollar checks for AI infrastructure they cannot actually produce or substitute on their own. So what I keep asking is — is this independence, or is it just... geography?
Dr. Nathan Hayes: That's the mechanistic tension, yes. And it gets sharper when you look at Nvidia's revenue from sovereign AI customers — it exceeded $30 billion in fiscal year 2026. More than triple the prior year. So the dependency isn't static. It's accelerating.
Maya Chen: Wait — triple?
Dr. Nathan Hayes: More than triple. Which means the nations declaring independence are, simultaneously, the customers funding the very dominance they say they're trying to escape.
Maya Chen: And here's what that triple number actually makes me think — it's not that nations are failing to be independent. It's that they're maybe buying a different *kind* of independence than the one they're announcing.
Dr. Nathan Hayes: Right — but the part that doesn't fit is, what are they actually getting? So, imagine you want a homemade dinner. You buy all local ingredients, local produce, everything from your region. But the stove, the oven, every recipe you're using — all from one foreign manufacturer, and there's no other manufacturer. That's it. That's sovereign AI right now.
Maya Chen: The food is yours but the kitchen isn't.
Dr. Nathan Hayes: Exactly that. And now here's the complication — Scrutica published their Sovereign AI Program Index, version 2026.3, July 17th, and they tracked roughly 30 jurisdictions, and what they flagged explicitly is that even the ambition is often inflated. Announced spending, committed spending, disbursed spending — those are three completely different things. They called conflating them the single most common source of inflation in sovereign AI reporting.
Maya Chen: So the headline number might not even be real spending.
Dr. Nathan Hayes: Potentially, yes. And then Counterpoint layered on something else — 56% of sovereign LLMs globally are adapted models. Built on foreign open-source bases. Not trained from scratch. So the ingredient metaphor actually needs revising — it's not even fully local produce. More like, you bought someone else's pre-made sauce and added your own herbs.
Maya Chen: Which means Noetra — Japan's thing, backed by Sony, SoftBank, Preferred Networks, NEC — the first sovereign act of that entire consortium is a purchase order. To Nvidia. For 27,500 Rubin chips.
Dr. Nathan Hayes: The sovereignty begins with a procurement form sent to the company you're declaring independence from. And the data center doesn't even come online until June 2028. So what governments are actually buying right now is — I mean, it's placement. Control over where the model lives, who can query it. Not control over the fundamental stack.
Maya Chen: But that's — placement, control over where the model lives — I keep turning that over, because what it hides is the deeper lock. Like, Japan gets its data center in June 2028, fine. But the engineer sitting down to actually optimize the training run for Noetra's robotics model? Every library they reach for, every debugging tool, every research paper with working code attached — it's all written for CUDA.
Dr. Nathan Hayes: That's the mechanism exactly. CUDA isn't a chip spec — it's a two-decade-old software stack. Parallel-computing platform, yes, but more precisely: it's baked into the optimization libraries, the training workflows, the entire talent pipeline. Rewriting for an alternative accelerator isn't a hardware swap. You're rebuilding the training pipeline from the ground up.
Maya Chen: So — wait, that means 'build an alternative to Nvidia' and 'break CUDA lock-in' are actually two completely different problems.
Dr. Nathan Hayes: Different problems, yes. And most sovereign programs are only attempting the first.
Maya Chen: Which is — mm, that's kind of staggering when you hold it against those spending numbers. All of that — ¥387.3 billion, the Saudi and UAE checks — none of it actually touches the software moat.
Dr. Nathan Hayes: Now, AMD is the case that makes this concrete. Credible chip designer — not a small player. They've been trying to compete seriously in AI deployments, and they still haven't matched Nvidia's performance and software ecosystem. Not for want of silicon. The CUDA gravity is just — it keeps pulling everything back.
Maya Chen: And if AMD can't crack it, what does that say about Noetra, or any sovereign program, actually expecting to?
Dr. Nathan Hayes: It says the full-stack ambition — chips, supercomputers, cloud, inference, all of it — every single layer still bottoms out at Nvidia hardware and CUDA software. The stack is sovereign in name. The foundation is not. And that's before we even get to what China's actually been doing with Huawei Ascend — because that story complicates this in a direction I didn't expect.
Maya Chen: Yeah — and I want to get there, because I think it's the one case that actually stress-tests whether any of this is breakable at all.
Dr. Nathan Hayes: China is actually the one jurisdiction that tried to answer that question — not with a spending announcement, but with a mandate. Beijing reportedly directed domestic tech companies in mid-2026 to actively limit use of Nvidia's H20 chips. Not 'consider alternatives.' Limit. And Huawei's Ascend accelerators — for the first time — topped China's domestic adoption rates, outpacing Nvidia's China-specific products. That's a genuine milestone.
Maya Chen: Wait — for the first time ever?
Dr. Nathan Hayes: First time. And the supply numbers are real — Huawei planned roughly 600,000 Ascend 910C units for 2026, with overall Ascend die output stretching toward 1.6 million, and successor chips — the 950, 960, 970 — rolling out through 2028. That's not prototype quantities. That's industrial scale.
Maya Chen: So that — I mean, does that break the argument? Like, if China's actually doing it at scale...
Dr. Nathan Hayes: Here's what doesn't fit, though — sources still explicitly describe Nvidia as in high demand inside China. Even as Ascend gains share. So it's not decoupling. It's fragmentation. Two parallel ecosystems, both running, neither actually replacing the other.
Maya Chen: Which is — okay, that's almost stranger than full replacement? Because it means even after all of that — the mandates, the 1.6 million dies — a Chinese researcher still reaches for the CUDA-native reference implementation when a new architecture drops.
Dr. Nathan Hayes: Correct. Training on Ascend is possible. Training as efficiently, with the same tooling depth — that's the unresolved part. Whether Huawei's actual software stack matches frontier training needs is, and I want to be precise here, explicitly unresolved in the sourcing.
Maya Chen: And this is the thing no liberal democracy seems willing to actually say out loud — China's path required years of anticipatory industrial policy, state procurement mandates, and a conscious decision to accept performance penalties. No other sovereign AI nation has demonstrated any appetite for that.
Dr. Nathan Hayes: So Huawei's milestone either proves the moat can break — or proves exactly how hard breaking it is, and what the price actually is.
Maya Chen: And that price point matters. What Counterpoint Research found on August 5th is that 92.4% of sovereign AI runs on Nvidia. And the one country that actually moved that number required something no democracy has said yes to. So what are the rest of them buying? I think... geography. Branding. The right to say the model lives here. And I don't want to call that worthless, because maybe that's genuinely meaningful — jurisdiction, legal access, something. But it's not what they're calling it.
Dr. Nathan Hayes: That's — actually, the framing I'd put on it is: they're buying autonomy over the endpoint, not the stack. Which may be the only autonomy available right now. Whether it's worth the price of permanent dependence on U.S. export permissions — that's not a technical question. That's a political one I genuinely can't answer.
Maya Chen: Yeah. I think that's where I'll sit with it — is permanent dependence on American infrastructure just the cost of having AI at all, right now? And if every nation signing these announcements understood that clearly... would they still use the word sovereign? I don't know. I really don't. Thanks for working through this with me.
Dr. Nathan Hayes: Good question to leave unanswered. Genuinely.
