Nvidia announced new projects: Rubin Ultra will be released in 2027, and Feynman is planned for 2028.

Updated Nvidia Roadmap for Data Centers

NVIDIA announced a development plan for data center equipment (DC).

The company confirms:

2026 – Vera Rubin Platform
* CPU + GPU – central processors *Vera* and graphics *Rubin*.

* Ecosystem add‑ons:

* LPU (Language Processing Unit) Groq LP30 – inference accelerator.

* DPU BlueField‑4 – network processor.

* NVLink 6 switch and upgraded Ethernet Spectrum‑X network with CPO optics.

* ConnectX 9 1600G SuperNIC network adapter.

> The platform is aimed not only at CPU/GPU but also at integrating the Groq LPU into Nvidia’s lineup – replacing the outdated Rubin CPX.

2027 – Rubin Ultra AI Accelerators
* New chips: four compute dielets + 1 TB HBM4E.

* Advantages: significantly higher than 2027 models, thanks in part to Groq LP35 inference blocks with NVFP4 support (faster).

* Server solution Kyber NVL144: 144 Rubin Ultra accelerators + NVLink 7 switch – at least *four times* faster than the Oberon NVL72 racks on 72 Blackwell accelerators.

2028 – New Architecture and Scaling
NVIDIA plans radical changes:

* Feynman uses multilayer crystal layout, improving scalability.

* Memory: updated C‑HBM4E with higher bandwidth and > 1 TB per chip.

* NVLink with integrated optics will allow up to 576 accelerators in an Oberon rack or 1152 in a Kyber rack.

Conclusion
NVIDIA systematically updates its data center equipment line, introducing new GPUs and CPUs every few years and raising performance annually. Plans for 2026‑2028 include the full Vera Rubin platform, Rubin Ultra accelerators, and finally a new architecture with Rosa processors, Feynman graphics, and Kyber CPO optical infrastructure. This will enable the company to compete with market leaders (AMD, Intel) in development speed and scalability of solutions for large data centers.