NVIDIA Kyber Architecture Delayed to 2028 as CPO Challenges Drag on Scaling Roadmap
0xBroomberg
Nvidia's next-generation Kyber rack architecture has been delayed by over 12 months to 2028 due to midplane manufacturing hurdles, while the four-die Rubin Ultra has been cancelled — opening a window for AMD and Google to close the scale-up gap.
Why is Kyber delayed by more than a year?
Semiconductor research firm SemiA reports that the Kyber NVL144 rack architecture is delayed by over 12 months, now expected in 2028 — just three months after CEO Jensen Huang demonstrated it at GTC.
The bottleneck is the PCB midplane — the large circuit board that connects every chip module inside the rack. SemiA says it remains unfeasible to manufacture in the near term.
This means → the delay is not a design problem but an engineering-production problem, and there is no quick fix on the horizon.
Why have the backup plans also fallen through?
Nvidia had developed NVL72x2 — two racks placed back-to-back, linked by copper NVLink to extend scale-up range. Cloud providers and hyperscalers pushed back hard on complexity and operational burden, and the design has been cancelled.
The other path was NVL576, which uses CPO (co-packaged optics — embedding optical transceivers right next to the chip so data travels as light instead of electricity) to connect eight racks through NVSwitches. But CPO technology itself is not ready; this plan will likely be delayed or limited to small volumes.
SemiA says a CPO-equipped NVSwitch will not arrive until the Feynman generation. In plain terms = the optical-interconnect path is at least one full product generation away.
Why has Rubin Ultra been scaled back?
The original four-die Rubin Ultra has been cancelled. Only a smaller two-die version remains, delivering roughly half the performance of the original design.
Nvidia plans to compensate by selling more racks — pushing Oberon Rubin and Oberon Rubin Ultra racks in higher volume to fill the performance gap.
This means → the per-rack ceiling is lower, and Nvidia is betting on volume over density — a trade-off that raises deployment costs and data-centre space requirements for customers.
What kind of window do competitors now have?
Kyber delayed plus Rubin Ultra halved — the combined effect is that Nvidia currently has no validated path to scale Rubin Ultra beyond a single rack.
SemiA sees an opening for AMD's MI500X and Google's TPUv8i Broadfly to surpass Rubin Ultra on the scale-up dimension.
This reflects a deeper signal: Nvidia still leads on single-chip performance, but connecting chips into massive clusters is becoming the industry's shared bottleneck — whoever solves the interconnect problem first captures the next round of advantage.
Content is for reference only, not financial advice.