Nvidia Delays Its Next AI Rack System to 2028

Nvidia has pushed back the launch of Kyber NVL144, the rack-scale system built to house its most powerful coming chips, by more than a year. The research firm SemiAnalysis reported on July 6 that the product will now arrive in 2028 rather than 2027, held up by a manufacturing problem that has proven harder to solve than the company expected.

Kyber is not a chip. It is the metal-and-copper cabinet that ties hundreds of Nvidia processors into a single machine, the unit cloud providers actually buy and install. Nvidia demonstrated it at its GTC conference only three months ago, in a keynote by chief executive Jensen Huang. The system was designed around Rubin Ultra, the flagship accelerator Nvidia had slated for 2027.

A circuit board that will not cooperate

The holdup sits in an unglamorous place: the midplane, a large printed circuit board that routes signals between the racks of chips. Kyber's midplane runs to 78 layers of wiring stacked into one board, and manufacturing it at volume without defects has become the bottleneck. When a board that complex fails quality checks, the whole rack waits.

Nvidia had a fallback. An alternative design, known internally as NVL72x2, would have bolted two of the current-generation racks back to back to buy time. SemiAnalysis reports that cloud customers rejected it, unwilling to build data centre space around a stopgap. A still larger configuration, NVL576, which lashes eight racks together through optical links, now looks likely to slip as well or ship only in small numbers.

Nvidia did not respond to a request for comment from CNBC, which covered the report. The company has not publicly confirmed the new timeline.

Why a rack delay matters

For most of the past two years the AI hardware story has been one of Nvidia setting the pace and everyone else reacting. A twelve-month slip on its headline system is a rare stumble, and it lands at an awkward moment. Anthropic, OpenAI, Google and Meta are all racing to stand up gigawatts of new compute, and their build-out schedules assume Nvidia ships roughly on time. A rack that arrives a year late forces a choice: wait, or fill the gap with today's chips and rivals' silicon.

That gap is exactly what Nvidia's competitors have been waiting for. Custom accelerators from Google and Broadcom, along with a growing field of challengers, become far more attractive when the market leader's roadmap wobbles. We covered one of those challengers separately today, the Chinese GPU maker Biren, which is raising close to 900 million dollars to press its own case.

None of this dents the demand for AI compute, which remains enormous. It does show that the hardest problems in the industry are no longer only about designing faster chips. Sometimes they are about the far more mundane business of manufacturing a very large circuit board without a single flaw, at scale, on schedule.