SemiAnalysis says Nvidia's Kyber rack has slipped to 2028 over a circuit board it can't yet build in volume

The research firm SemiAnalysis says Nvidia has pushed its Kyber rack system from 2027 to 2028, a slip of more than a year, and blames a single component: the printed circuit board that ties the whole cabinet together. Nvidia's answer to a request for comment was four words. "Our roadmap is intact."

Kyber is not a chip. It is the cabinet Nvidia designed to hold the chips. Announced at GTC in 2025, a Kyber rack packs 144 Rubin Ultra GPUs, wires them together over the NVLink 7 switch, and stands the compute trays up vertically instead of laying them flat, a layout Nvidia has compared to books on a shelf. It is liquid-cooled, draws around 600 kilowatts, and was pitched as roughly four times the performance of the current Blackwell NVL72 generation, known internally as Oberon. The point of a rack like this is scale-up: getting a large number of GPUs to behave as one tightly coupled machine, with every chip able to talk to every other chip at very low latency.

The part that broke is the boring part

The component SemiAnalysis fingers is the midplane. In a rack this dense, the midplane is the backbone every compute and switch tray plugs into, the board that actually carries the high-speed signals between GPUs. Reporting on the SemiAnalysis note puts it at 78 layers, which would make it one of the most complex PCBs ever built for a commercial product.

Layer count matters here for a concrete reason. Every extra layer is another sheet of copper and dielectric that has to be aligned to the ones above and below it within tolerances that shrink as signaling speeds climb. At NVLink 7 data rates, small imperfections in that stack turn into signal loss and crosstalk, and the share of boards that come out usable drops. This is not a marketing problem or a software slip. It is a manufacturing yield problem on a piece of hardware that has to be close to perfect, and those are the kind that move release dates by quarters, not weeks.

"Kyber NVL144 rack architecture has been delayed to 2028 as the PCB midplane remains challenging from a manufacturability standpoint," SemiAnalysis wrote.

The backup plan already fell over

Nvidia's hedge, per the same report, was to bolt two Oberon racks back to back and reach a similar power envelope that way. That plan is dead too. SemiAnalysis says cloud operators and hyperscalers rejected it over its "odd design and heavy operational burden," which is a polite way of saying the people who would have to run these things in production did not want to.

The knock-on hits the larger NVL576 system, which links eight Oberon racks together over co-packaged optics. SemiAnalysis expects that one to be delayed as well, or shipped in small volumes. Put together, the firm's conclusion is that Nvidia has "no proven solution to expand the scale-up world size for Rubin Ultra." That, not the calendar, is the actual story. A GPU is only as useful as the fabric you can wire it into, and if the biggest coherent domain you can build shrinks, so does the ceiling on the largest models these racks were meant to train and serve.

This fits a pattern from the same week

A week before the Kyber note, SemiAnalysis reported that Nvidia had scrapped the four-die version of Rubin Ultra in favor of a two-chiplet design, again citing manufacturing execution. The two-die part lands at roughly half the compute of the four-die version it replaced. Two reports, two retreats from the most aggressive version of the same product, both pinned on getting hard things to come off a line at acceptable yield.

What to actually take from this

Everything above is one research firm's account, and Nvidia disputes the conclusion. SemiAnalysis has a strong track record on supply-chain reporting, but "our roadmap is intact" is a real denial from the company that would know, and it is worth holding both in mind until parts either ship or don't. This also does not touch the Rubin generation already in the pipeline, which is a separate product from the Rubin Ultra racks Kyber was built for.

What is hard to wave away is the theme. The constraint on frontier AI hardware right now is not whether Nvidia can design a faster chip. It is whether the packaging, the boards, and the cooling that surround the chip can be built in volume without the yield falling through the floor. A 78-layer midplane that a factory can't reliably produce is a more honest picture of where the industry is stuck than any performance-per-watt slide.