Network operators face a pressing challenge: how to monetise heavy 5G capital expenditures while preparing for the 6G era. The traditional revenue model based on simple connectivity yields decreasing returns, pushing operators to pivot toward intelligent platforms, exposed network APIs, and value‑added edge services.
In this context, NVIDIA and Marvell Technology have announced a strategic partnership to integrate Marvell into NVIDIA’s AI factory and the AI‑RAN ecosystem via the NVLink Fusion platform. The collaboration offers customers building on NVIDIA architectures more choice and flexibility when designing next‑generation hardware stacks.
Supporting the technical partnership is a substantial financial commitment: NVIDIA has invested $2 billion in Marvell. Analysts often look for meaningful monetary backing to validate hardware alliances, and this investment positions Marvell to scale production of custom silicon and high‑performance networking components.
Beyond capital, the companies will collaborate on silicon photonics and optical interconnects—key technologies to meet the demand for high‑speed, low‑latency networking required to support demanding AI workloads at the edge.
Navigating heterogeneous compute environments
The partnership is built around NVIDIA NVLink Fusion, a rack‑scale architecture that enables semi‑custom AI configurations while maintaining compatibility with the existing NVLink ecosystem. This approach lets operators and cloud builders mix and match specialised components to meet diverse performance and deployment needs.
Under the arrangement, Marvell supplies custom XPUs and scale‑up networking hardware compatible with NVLink Fusion, while NVIDIA provides foundational components such as the Vera CPU, ConnectX NICs, BlueField DPUs, Spectrum‑X switches, the NVLink interconnect, and rack‑scale AI compute. This division produces a heterogeneous infrastructure that supports custom processing engines and ensures interoperability with NVIDIA systems.
For operators, the result is a smoother path to integrate edge deployments with NVIDIA GPUs, LPUs, networking, and storage platforms and to leverage NVIDIA’s global supply chain. BlueField DPUs, for example, let operators offload heavy security and networking functions from main processors, freeing compute cycles for AI workloads that drive revenue.
Matt Murphy, Marvell’s Chairman and CEO, said the expanded partnership highlights the increasing importance of high‑speed connectivity, optical interconnect, and accelerated infrastructure for scaling AI. He emphasised that combining Marvell’s strengths in high‑performance analog, optical DSP, silicon photonics, and custom silicon with NVIDIA’s NVLink Fusion ecosystem enables customers to build scalable, efficient AI infrastructure.
Driving business impact and enterprise revenue
Deploying specialised compute nodes inside the Radio Access Network transforms the economics of cellular sites. Using the NVIDIA Aerial AI‑RAN framework for 5G and 6G, operators can convert the telecom network into distributed AI infrastructure and host enterprise workloads directly at cell towers.
This edge capability opens new revenue streams that are independent of consumer smartphone subscriptions. Enterprises need ultra‑low latency for automated manufacturing, autonomous logistics, and real‑time video analytics. Operators can monetise edge compute by leasing capacity to enterprises, increasing average revenue per user (ARPU) from non‑consumer sources and improving enterprise customer retention.
Private 5G deployments are another direct use case. Marvell’s custom silicon combined with NVIDIA’s rack‑scale compute gives operators the hardware needed to win lucrative private networking contracts while keeping sensitive data on premises to meet strict data sovereignty and compliance requirements.
Jensen Huang, NVIDIA’s Founder and CEO, described the market shift toward inference and token generation demand and said the partnership with Marvell helps customers scale NVIDIA’s AI infrastructure to build specialised AI compute solutions.
However, rolling rack‑scale AI compute into existing mobile switching centres involves non‑trivial integration work. NVLink Fusion deployments must be coordinated with legacy Operations Support Systems (OSS) and Business Support Systems (BSS), which were designed to meter voice minutes and data, not continuous API calls or dynamic edge compute provisioning. Modernising billing engines to support AI‑RAN monetisation is therefore a multi‑year endeavour.
Spectrum and resource management also become more complex. Running multi‑tenant AI workloads alongside high‑priority 5G baseband processing requires strict resource isolation; network architects estimate that maintaining this isolation can consume meaningful edge compute overhead.
Operators must also manage multi‑cloud environments and ensure that containerised network functions interoperate with enterprise AI applications sharing the same physical silicon. Marvell XPUs and NVIDIA Vera CPUs offer the processing diversity needed for these workloads, but the software orchestration layer remains a significant challenge for IT teams.
When operators expose edge capabilities via network APIs, third‑party developers can build applications that integrate directly with radio infrastructure. Creating a developer‑friendly API portal requires considerable investment in software, particularly for authentication, metering, and billing systems that can handle thousands of concurrent enterprise requests. Upgrading legacy backends often involves navigating vendor lock‑in and customised deployments while avoiding disruptions to existing subscriber services.
Vendor context within the telecoms ecosystem
Bringing advanced computing into mobile networks aligns operators more closely with hyperscalers and specialist hardware vendors. The telecoms industry still depends on a few dominant radio vendors, and integrating NVIDIA and Marvell hardware requires those suppliers to adapt proprietary interfaces.
The industry’s push toward Open RAN has created opportunities for third‑party silicon to handle radio processing and enterprise compute on the same platform. Embedding the NVIDIA Aerial framework into physical networks effectively turns operators into geographically distributed extensions of the broader AI ecosystem—a shift that requires careful commercial and operational planning.
Wholesale carriers and operators must protect control over subscriber data and network telemetry rather than serving solely as an unmonetised conduit for third‑party edge services. While NVLink compatibility simplifies physical deployment and supply chain logistics, the commercial agreements that govern multi‑vendor edge environments remain complex and require negotiated terms that preserve operator value.
Operating an AI‑RAN requires a software‑centric mindset similar to DevOps practices used by hyperscale cloud providers. Workforces must learn to manage compute capacity dynamically and treat the network as a programmable platform instead of a static set of cell towers. Only through this operational transformation can operators fully realise the commercial benefits of the sophisticated hardware enabled by the NVIDIA‑Marvell partnership.
Want to learn more about AI and big data from industry leaders? Attend AI & Big Data Expo events held in Amsterdam, California, and London, part of TechEx and co‑located with other technology summits including IoT Tech Expo and Cyber Security & Cloud Expo.
Telecoms content is produced by TechForge Media. Explore other upcoming enterprise technology events and webinars through TechForge Media’s event listings.