Nvidia says the roadmap is intact. It won’t say which roadmap.

Jensen Huang spent Wednesday in Japan killing a rumor. The reports about Vera Rubin delays are “not true,” the Nvidia CEO told reporters, in comments reported by Bloomberg and picked up by Tom’s Hardware. “Vera Rubin is already in production. Giant amounts of production incoming.”

On Thursday the company produced a receipt: a 140-megawatt AI factory in Japan, built with Noetra Corp., a new consortium founded by SoftBank, Sony, NEC and Honda under a METI-backed national program. It will house 27,500 Rubin GPUs and 13,750 Vera CPUs across 382 Vera Rubin NVL72 racks, with up to ¥1 trillion (about $6.1 billion) committed over five years.

Neither the denial nor the design win touches the report that moved the supply chain this month. That one came from SemiAnalysis on July 5. Kyber NVL144, the rack-scale system that turns Rubin Ultra silicon into the machines hyperscalers actually buy, has slipped from 2027 to 2028, the analyst firm wrote on X:

MASSIVE DELAY: Just 3 months after Jensen demoed Kyber NVL144 at GTC, it has faced major setbacks and has been delayed by more than 12 months, pushing it back to 2028.

The reported culprit is Kyber’s PCB midplane, the rigid board that replaces cable harnesses and carries the copper NVLink fabric connecting every GPU in the cabinet. Per Tom’s Hardware’s breakdown, that board is a roughly 78-layer stack spanning nearly a square meter, with trace spacing at or below 25 micrometers and impedance tolerances of 5 percent. Signal integrity gets harder as the layers pile up. So does yield.

Suppliers felt it immediately. Ibiden fell as much as 10 percent, Kingboard Laminates 18 percent, Samsung Electro-Mechanics 11 percent and Elite Material 10 percent in the sessions after the report, The Decoder noted, sharp reversals for stocks that had ridden Nvidia’s rack ambitions up several hundred percent this year.

Nvidia’s answer, delivered to Tom’s Hardware and other outlets: “Our roadmap is intact.”

Four words, no dates.

The chip is not the rack

Read closely and the two statements deny different things. Huang’s denial covers Vera Rubin, the GPU-and-CPU platform, and on that point the record backs him: Rubin entered production in January and began sampling to customers in February, and 27,500 of the chips now have a Japanese address. “Roadmap is intact” is the only phrase that reaches Kyber, and it is compatible with almost any Kyber timeline. If the rack was always a 2028 product internally, the statement is true. If the schedule slipped a year and the revised roadmap now says 2028, the statement is still true. Roadmaps, once revised, are also intact.

The Noetra deal, for all its scale, is evidence about the chip, not the rack. The Japan factory uses Vera Rubin NVL72 racks: the current 72-GPU, Oberon-generation architecture carried forward to new silicon, due to hit volume production in the second half of this year. Kyber is a different object entirely, cramming 144 GPU packages into one cabinet, double today’s density, all hanging off that meter-scale midplane. A 27,500-GPU order proves Nvidia can supply Rubin chips in volume. It says nothing about when the next rack ships.

The date matters because the fallbacks are gone. SemiAnalysis also reported, and Tom’s Hardware detailed, that Nvidia axed NVL72x2, a stopgap that would have bolted two Oberon cabinets together, after customers pushed back.

Our read: if Kyber really is a 2028 machine, Rubin Ultra spends 2027 in racks designed for half its density, the window in which AMD’s and Google’s own large-scale systems are expected to land.

Huang held Kyber’s gray backplane up onstage at GTC in March. Four months later, the most specific thing his company will say about it is four words that fit any calendar.