Nvidia's Rubin Platform Is in Full Production. Its Successor Just Hit a Wall.

Two things happened to Nvidia's chip roadmap in the same week, and they sound contradictory until you realize they're about two different products. Rubin, the platform shipping this year, is in full production. Kyber, the rack built for Rubin's more powerful successor, just slipped by more than a year.

What Rubin actually is

Rubin isn't one chip. It's six chips designed together: the Vera CPU, the Rubin GPU, an NVLink 6 switch, a ConnectX-9 SuperNIC, a BlueField-4 DPU, and a Spectrum-6 Ethernet switch, all co-designed so the bottleneck isn't wherever the weakest chip sits.

Packaged into the Vera Rubin NVL72 rack or the smaller HGX Rubin NVL8 system, Nvidia's own numbers claim up to a 10x reduction in inference token cost and 4x fewer GPUs needed to train mixture-of-experts models, compared with Blackwell.

That 10x figure matters more than it sounds. It's the same direction every hyperscaler custom-silicon program is pushing, including Meta's Iris chip entering production in September. Nvidia has to keep cutting its own per-token cost or the economics start favoring in-house silicon by default.

Rollout

Partners get Rubin-based instances in the second half of 2026. AWS, Google Cloud, Microsoft, and OCI are named as first deployers, alongside CoreWeave, Lambda, Nebius, and Nscale. All six chips are back from fabrication and currently running validation workloads.

Then Kyber ran into a PCB problem

On July 6, SemiAnalysis reported that Kyber NVL144, the rack built for the more powerful Rubin Ultra chip, is delayed by more than 12 months, pushing it to 2028. The reported cause: manufacturing yield on Kyber's midplane, a 78-layer orthogonal backplane PCB using exotic materials that isn't behaving in production the way it did in the demo Jensen Huang gave three months earlier.

The report adds two more details worth noting: the four-die Rubin Ultra GPU design has reportedly been dropped for a simpler two-die package, and a stopgap rack architecture, NVL72x2, was scrapped after cloud customers pushed back on it.

[!WARNING] A 78-layer PCB midplane sounds like an obscure manufacturing detail until you remember it's the physical layer that determines whether a $3 million rack actually boots reliably at scale. Yield problems at that layer don't show up in a demo. They show up six months into a production ramp, which is exactly when SemiAnalysis says this one did.

Nvidia pushed back the same day, telling Bloomberg its roadmap is "intact" and pointing to standard Rubin deliveries reaching eight cloud customers this fall as proof the near-annual cadence is holding.

Both statements can be true

Rubin, the platform arriving this fall, is unaffected. Kyber is the rack for Rubin Ultra, the higher-end successor still a year-plus out. Nvidia denying a "delay" while a specific successor product slips isn't spin so much as two different products getting conflated in the same headline.

The part worth tracking isn't who's right this week. It's whether Nvidia can hold a genuinely annual cadence at this level of physical complexity. Six co-designed chips landing on schedule is one thing. A 78-layer backplane behaving in volume production is another, and that's the one that just missed.