NVIDIA and Marvell announced a collaboration to link Marvell into NVIDIA’s AI factory via NVLink Fusion, a rack-scale interconnect platform designed to support heterogeneous accelerators and networking. The deal, which included a $2 billion NVIDIA investment in Marvell, featured work on silicon photonics and NVLink Fusion-compatible scale-up networking, with Marvell supplying custom XPUs.
NVIDIA committed its Vera CPU, ConnectX NICs, BlueField DPUs, NVLink interconnect and Spectrum-X switches to create a combined stack, while Marvell contributed analog, optical DSP and custom silicon expertise. The partnership also targets Aerial AI-RAN for 5G/6G integration and advanced optical interconnect solutions.
For enterprises and telcos, the tie-up enables more choice when building rack-scale AI infrastructure, easing integration of non‑GPU accelerators and optical links into NVIDIA environments. The move reflects a broader trend toward modular, heterogeneous AI factories that prioritize high-speed connectivity and power-efficient silicon photonics.
NVLink Fusion Partnerships Expand
NVIDIA Teams With Marvell to Integrate NVLink Fusion
Trend Themes
1. Heterogeneous Rack-scale Architectures - The convergence of GPUs, XPUs, DPUs and smart NICs within a unified rack fabric enables novel co-designed hardware stacks that optimize workload placement and latency at rack scale.
2. Silicon Photonics Integration - Advances in optical DSP and silicon photonics point to high-bandwidth, power-efficient interconnects that reduce electrical bottlenecks for multi-accelerator systems.
3. Nvlink Ecosystem Expansion - Broader NVLink-compatible networking and third-party silicon participation support an interoperable ecosystem where diverse accelerator vendors can plug into common AI infrastructure.
Industry Implications
1. Telecommunications 5G/6G RAN - Targeting Aerial AI-RAN use cases, integrated optical and accelerator fabrics enable distributed, low-latency inference and real-time signal processing at the network edge.
2. Cloud Data Centers - Rack-scale NVLink fabrics and mixed-accelerator support create opportunities for differentiated cloud offerings that price and provision heterogeneous compute resources flexibly.
3. Edge AI Infrastructure - Power-efficient photonic links and compact XPUs suggest new hardware form factors for high-performance inference hubs operating outside traditional centralized data centers.