Nvidia Pushes Its New AI Rack to 2028, Over a Year Late

The Kyber NVL144 rack, unveiled by Jensen Huang at GTC three months ago, slips more than twelve months and lands in late 2028. The root cause is a 78-layer PCB midplane deemed unmanufacturable on the original timeline. The backup plan, a back-to-back twin-rack architecture, was pulled after cloud customers pushed back.

Key Takeaways

• Kyber NVL144 slips over 12 months and lands in 2028 instead of 2027.
• The 78-layer PCB midplane remains unmanufacturable at scale, no industrial fix locked in yet.
• The NVL72x2 backup rack was scrapped after hyperscalers rejected the operational complexity.

The blocker is a 78-layer midplane

The core of the issue is not a chip. It is the rack midplane, the board that connects compute trays and switch trays inside the chassis. Nvidia targeted a 78-layer PCB design. That is among the most complex boards ever specified for a commercial computing product.

At that stack height, every fabrication step introduces impedance variations that break signal integrity at the target frequencies. The PCB suppliers approached by Nvidia have not converged on an acceptable industrial yield. The production line does not exist yet, and no one wants to finance it around a design Nvidia is still tuning.

Jensen Huang had sold Kyber as the rack meant to carry the Rubin Ultra generation, the big successor to Blackwell expected at hyperscalers in 2027. The original schedule assumed PCB production would start by the end of this year. We are in July, and no volume contract is signed.

Nvidia’s recent $20 billion bond raise does not paper over this industrial block. Cash is abundant, calendar space is not.

Rubin Ultra loses its scale

Kyber was meant to give Rubin Ultra its scale-up architecture, the setup that lets a single rack behave like one massive accelerator. Without Kyber, every Rubin Ultra instance stays capped at the previous generation’s format.

Nvidia had prepared a fallback, the NVL72x2, two NVL72 racks placed back to back to approach the target density. The format doubles the floor footprint per logical node, complicates the optical cabling between the two halves, and forces server rooms into a cooling plan that has nothing in common with a monolithic NVL144.

The hyperscalers said no. None of the major operators want to redesign their bays for an interim product that will live twelve to eighteen months. The operational complexity is too high, the payback window too short. Nvidia eventually pulled the NVL72x2 from its official roadmap.

The outcome is that Rubin Ultra 2027 will ship in a plain NVL72 format, with the same per-rack memory density as Blackwell Ultra. The promised scale jump moves to 2028. The next question is whether the 2028 date holds up at all.

Nvidia replied with the minimal line: “Our roadmap is intact.” That answer works as long as no analyst breaks down the gap between the original spec sheet and the 2028 version. A 78-layer design does not compress into three quarters without breaking bandwidth specs.

Also on Horizon:

• Zhipu Ships China’s Free Rival to Claude With ZCode
• Claude Code Ports Command & Conquer iOS in 40 Minutes
• Alpha School Sells $75K AI Schools to Wealthy US Families

What it changes for hyperscalers

In the short term, the order is clear. Hyperscalers will over-provision Blackwell Ultra across 2026 and 2027. Every data center that was queued for Rubin Ultra scale-up now orders extra GB300 racks to carry the training load.

That over-order explains the nervousness across memory suppliers. SK Hynix and Samsung, delivering HBM4 for the current generation, are under stress. Part of the demand that Together AI and similar operators were funneling to hyperscalers now shifts to older-generation compute that is actually available.

On the medium horizon, the strategic opening is more interesting. AMD MI400 lands in volume by 2027. Broadcom is pushing its ASIC accelerators directly inside Meta and Google. If Rubin Ultra misses its scale-up window, the catch-up runway for alternatives extends by twelve precious months.

Chinese players use the breathing room. The DeepSeek push to unlock 85% extra speed on existing GPUs takes a whole new weight when the next generation of US scale-up slips a full year. Squeezing the installed base becomes a growth axis of its own.

The equipment layer is the last variable. Kyber’s delay slows TSMC’s advanced packaging order book tied to the project. The high-end PCB suppliers lose an industrial opportunity that could have reshuffled the sector’s cards.

Follow the story on Horizon.