Spuds Oxley: There's a specific kind of confusion that — okay, it only happens when two people are both right about two completely different things, but everyone watching assumes they must be fighting about the same one.
Spuds Oxley: That's what happened in Tokyo.
Spuds Oxley: Jensen Huang, July 15th, 2026, on stage at a developer event in Japan — 'Vera Rubin is already in production. Giant amounts of production incoming.' He called SemiAnalysis's delay reports 'not true.' Direct, unhesitating, public.
Spuds Oxley: He's not wrong. NVIDIA confirmed Vera Rubin in full production May 31st of this year. 350-plus factories. 30 countries. The NVL72 rack, Rubin GPUs fabbed on TSMC's N3-class process, a platform NVIDIA walked onto the CES 2026 stage to formally unveil — that system is shipping.
Spuds Oxley: But SemiAnalysis wasn't writing about Vera Rubin.
Spuds Oxley: They were writing about the Kyber NVL144 — a rack system for Rubin Ultra, the next chip generation after the current platform, designed for 2027 deployment, and now reportedly delayed more than 12 months to 2028 because the specialized circuit board it requires kept failing in manufacturing.
Spuds Oxley: Vera Rubin. Kyber NVL144. Not the same thing.
Spuds Oxley: One is the NVL72 — in production right now. The other is one tier up, a larger rack for the Rubin Ultra generation, not due until 2027 in the first place.
Spuds Oxley: Huang denied delays to a product that isn't delayed. SemiAnalysis reported delays to a product Huang never mentioned. And somewhere in the gap between those two things — that's where most of the coverage went.
Spuds Oxley: I think we got the whole frame wrong. And I want to slow down at that hinge point — because the gap matters more than either side of it.
Spuds Oxley: Now here's what Noetra Corp. is actually buying.
Spuds Oxley: July 16th, the day after Huang's Tokyo remarks — NVIDIA announces a partnership with Noetra Corp. to build a sovereign AI factory in Japan. Thirteen thousand seven hundred and fifty Vera CPUs. Twenty-seven thousand five hundred Rubin GPUs. That is not a pilot program. That is infrastructure.
Spuds Oxley: And it's built on Vera Rubin. Specifically on the NVL72 rack — the AI Factory concept NVIDIA has been describing as the unit of industrial AI compute. Not individual chips. The whole system: GPUs, CPUs, networking, memory, all integrated at rack scale.
Spuds Oxley: One NVL72 rack pools 20.7 terabytes of HBM4 across 72 GPUs.
Spuds Oxley: That number is the design philosophy made physical. NVIDIA isn't chasing raw compute here — they're chasing memory bandwidth. Because the workloads that matter at sovereign scale, the large inference runs, the model serving, they are constrained by how fast you can move data, not just how fast you can crunch it.
Spuds Oxley: NVIDIA claims Vera Rubin delivers ten times the agent throughput over the Blackwell generation.
Spuds Oxley: Ten times. That's the promise Japan is buying.
Spuds Oxley: And this is where sovereign AI stops being an abstract geopolitical term and becomes a deadline. Japan isn't leasing compute from a foreign cloud — they are building domestically controlled AI infrastructure, data, models, compute, the whole stack inside their own borders. That is the sovereign AI concept. And it runs on a timeline. It runs on delivery dates.
Spuds Oxley: Think about what that actually means for Noetra Corp. right now.
Spuds Oxley: They are planning against a roadmap. The Vera Rubin NVL72 — partner shipments targeted for fall 2026. That part looks solid. But the next tier up, the Kyber NVL144, the rack designed for Rubin Ultra, the system SemiAnalysis says is delayed more than twelve months to 2028 — that is the tier customers need if they want to scale beyond what the NVL72 can do.
Spuds Oxley: The gap is quiet. That's the problem.
Spuds Oxley: Blackwell systems are still shipping right now — massive demand, deep into Q3 — and honestly, many of those units haven't even been deployed yet by the customers holding them. So there's absorption capacity. There's cushion. For a while.
Spuds Oxley: But Noetra Corp., and every other sovereign AI partner who walked out of Tokyo with a 2027 deployment plan, they are working from a roadmap that may have a 2028 hole sitting in the ultra-scale tier — a hole that wasn't named on stage.
Spuds Oxley: Huang wasn't being dishonest. But Tokyo was doing two things at once — announcing product confidence and reassuring partners — and those two jobs can crowd each other out when the precision gets lost.
Spuds Oxley: A billion-dollar deployment decision made against an incomplete roadmap isn't a technical failure.
Spuds Oxley: It's a framing failure. And those are quiet until they aren't.
Spuds Oxley: Now, the thing that stays open — the thread nobody has pulled — is not the timeline. It is the circuit board.
Spuds Oxley: SemiAnalysis did not say NVIDIA was late on a schedule. They named a specific manufacturing failure — a specialized circuit board for the Kyber NVL144 that kept failing in production. That is an engineering claim. A physical problem with a physical component.
Spuds Oxley: NVIDIA has not addressed that.
Spuds Oxley: Huang denied the delay in Tokyo. What he did not do — could not do from a stage — is explain what changed in the board manufacturing process, or whether the problem has actually been solved. That distinction matters more than most of the coverage treated it.
Spuds Oxley: Because the Kyber NVL144 is the rack that houses Rubin Ultra. The 2027 generation. The tier hyperscalers need when the NVL72 is not enough. And the credibility of that whole tier rests entirely on whether one circuit board can be manufactured at scale.
Spuds Oxley: Think about the stack underneath that problem. Vera Rubin is seven new chips — the Rubin GPU, the Vera CPU, Spectrum-X Ethernet Photonics with co-packaged optics, all of it shipping together as a coherent platform. One hundred and fifty supply chain partners in Taiwan alone. Three hundred and fifty factories across thirty countries.
Spuds Oxley: A single board failure does not stay isolated in a system that complex.
Spuds Oxley: So here is what I'm watching. Fall 2026 — that is when partner shipments to cloud providers are supposed to begin for Vera Rubin. That is the first real credibility checkpoint. Not an announcement. Not a stage. Actual hardware in the hands of eight named cloud providers.
Spuds Oxley: And GTC 2026 — that is the moment NVIDIA's own developer conference becomes the place where they either address the Kyber NVL144 engineering question directly, or they don't. Those two dates are the measure.
Spuds Oxley: If Rubin Ultra samples are due in 2027, the circuit board problem has to be solved — visibly, specifically, not rhetorically — before those samples ship. That is not a philosophical point. That is a deadline with a component attached to it.
Spuds Oxley: Noetra Corp. signed that partnership on July 16th, the day after Tokyo. They are building sovereign AI infrastructure. They are planning against a roadmap. Fall 2026, then Rubin Ultra in 2027… and somewhere behind that, a circuit board NVIDIA has not publicly said is fixed. That is the watch. That is the thing still open.
Spuds Oxley: Jensen Huang and SemiAnalysis can both be telling the truth. Completely. At the same time. Vera Rubin is in production. The Kyber NVL144 is delayed. Those two statements do not contradict each other because they are about DIFFERENT products, and the coverage treated them like they were about the same one.
Spuds Oxley: That's the problem. Not that someone lied. Not that NVIDIA is hiding something. The problem is that a stage denial of one product became, in the telling, a defense of an entirely different product — and the customers who need to know the difference, the ones with billions already pointed at 2026 and 2027 Rubin deployments, they are the ones holding a roadmap that may not have been fully defended.
Spuds Oxley: If you are Noetra Corp., or anyone else who walked out of Tokyo with a plan built against Rubin Ultra and the Kyber NVL144, the question worth asking is not whether Jensen Huang was honest. He may well have been. The question is which product he was actually defending when he said 'not true.'
