OCP Global Summit 2025 | NVIDIA | Shaping the Future of Open Infrastructure for AI

3 days ago

3 min read

Introduction

At OCP Global Summit 2025, NVIDIA shared how it is driving the future of AI through open infrastructure. The talk stressed that AI has entered the gigascale era: data centers around the world are being built at the gigawatt (GW) scale, and everything from single chips and servers to cross-data-center superclusters needs new architectures and innovation. In this process NVIDIA is not just a GPU supplier but a driver of the entire AI infrastructure ecosystem.

Content

The Gigascale AI Era

NVIDIA noted that we are in an unprecedented period of AI expansion. Data centers routinely commit GW-scale power, and build density and technical complexity are at record highs. This infrastructure is no longer just a compute center but an asset that "gets smarter on its own," with performance and value rising continuously as software and models iterate.

GPU and Data Center Performance Gains

Using B200 and GB200 data centers as examples, NVIDIA showed the leap in performance:

In just two months, a B200 data center raised the performance of OpenAI's GPTOSS model by 5x, cutting token cost from 11 cents to 2 cents per million tokens.
In another case, GB200 with the NVL72 architecture delivered 15x the performance of H200 on the DeepSeek R1 model, translating directly into revenue.

These examples show that hardware performance gains and network optimization are not just technical breakthroughs but drivers of the business model.

Next-Generation Architecture: Vera Rubin

NVIDIA unveiled the Vera Rubin architecture, arriving in the second half of 2026:

A two-chip design: the CPX processor handles context processing, paired with the Rubin GPU to deliver more than 8 exaflops of inference performance.
Networking enters the 400G scale-up era, with total bandwidth of 260 TB/s, twice that of GB200.
100% liquid-cooled, OCP MGX-compatible design that keeps the existing 45°C inlet temperature spec, lowering cooling infrastructure costs.
A new 500A liquid-cooled busbar improves energy efficiency and provides greater resilience and scalability.

NVLink Fusion and Broader Partnerships

NVIDIA introduced NVLink Fusion, enabling heterogeneous processors to integrate deeply with GPUs:

The Fujitsu Monaka CPU already connects to GPUs through NVLink Fusion.
Intel will also launch CPUs supporting NVLink Fusion, further expanding the ecosystem.
Other partners include Alchip, Samsung and MediaTek, jointly building an open heterogeneous computing environment.

SpectrumX and Cross-Data-Center Interconnect

As AI clusters scale to the million-GPU level, Ethernet and SpectrumX become critical:

Microsoft's Fairwater supercomputer already uses SpectrumX on the OCP SONiC architecture.
Oracle's Stargate data center uses OCP and SpectrumX to build an interconnect fabric with 95% effective bandwidth.
NVIDIA announced Spectrum XGX, supporting AI training across data centers and geographic locations.

Looking Ahead: The Kyber Architecture

In the second half of 2027, NVIDIA will launch Kyber:

A single rack integrates 576 GPUs.
Combined with 800V DC power infrastructure to handle extreme power density.
This will push AI training further into hyperscale.

Conclusion

NVIDIA delivered a clear message in this talk: the evolution of AI infrastructure is not just stacking GPU performance but full-stack innovation across chips and data centers. From Vera Rubin to Kyber, NVIDIA is working with the industry through OCP to lay the groundwork for gigascale AI infrastructure.

Further Perspectives

Technical impact: Vera Rubin's two-chip design and 400G scale-up network show that future AI architectures will rely heavily on silicon photonics interconnect to break through copper and power bottlenecks.
Supply chain observations: opening up NVLink Fusion will reshape the CPU–GPU relationship, creating opportunities for Intel, AMD and even the Arm ecosystem. Optical module and SiPh vendors (such as Broadcom, Marvell and Innolight) will benefit from SpectrumX deployments.
Market trends: as Kyber doubles the GPU count per rack, AI data center CAPEX and power demand will keep climbing. This will push liquid cooling, 800V DC power distribution and silicon photonics optical interconnect into the mainstream and drive a new wave of industry investment.