Nvidia backs Marvell with $2b to widen custom AI stack

Nvidia has invested US$2 billion in Marvell Technology to integrated its custom AI chips with Nvidia networking gear and CPUs stack.

Marvell shares rose about 7% after the announcement.

The companies will work on networking for AI systems, including optical interconnects and silicon photonics, with Marvell supplying chips and networking products compatible with Nvidia’s NVLink Fusion, and Nvidia providing CPUs, and network interface cards.

An eMarketer analyst said the deal gives Nvidia access to Marvell’s semi-custom silicon and optical interconnect technology to address bandwidth and power constraints in AI data centers.

Alphabet and Meta are expected to spend at least US$630 billion on AI infrastructure this year, while Marvell has said it expects revenue to rise.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Nvidia’s investment fits into a wider push for optical interconnects

The $2 billion investment follows a broader effort to lock down next-generation data center parts 1.
Recent $2 billion investments in Lumentum and Coherent add laser suppliers that support co-packaged optics, which places optical parts closer to chips for faster switches 1.
That approach overlaps with Marvell’s optical work, including optical digital signal processor (DSP) platforms such as 1.6T optical DSPs, plus its 3D silicon photonics (SiPho) Engine for co-packaged optics 2.
Marvell also planned to buy Celestial AI, which builds “Photonic Fabric” technology to speed data movement inside AI servers, though the deal had not closed as of Marvell’s Q3 FY 2026 update 3.

Nvidia is pulling rivals into its closed ecosystem through partnerships

Marvell designs custom chips for Amazon Web Services (AWS), whose Trainium AI accelerators give hyperscalers (very large cloud providers) another path besides Nvidia GPUs 4.
The partnership ties Marvell into Nvidia’s AI factory and AI-RAN (AI for radio access networks) efforts through Nvidia NVLink Fusion, a rack-scale platform for semi-custom AI infrastructure built on the NVLink ecosystem 5.
NVLink Fusion systems must include at least one Nvidia product, which keeps Nvidia revenue in the stack even when an Nvidia GPU is not the main processor 4.
This direction puts Nvidia up against UALink, an open interconnect backed by AMD, Intel, and Broadcom 4.