Nvidia Invests $2 Billion in Marvell, Launches NVLink Fusion to Co-opt Custom Chip Rivals

Nvidia announced on March 31 a strategic $2 billion equity investment in Marvell Technology, paired with a multi-year technical partnership centered on a new platform called NVLink Fusion. Marvell shares surged as much as 11% on the news. The deal marks a fundamental shift in how Nvidia positions itself in the AI infrastructure market.

What NVLink Fusion Is

NVLink Fusion is a platform that allows Marvell’s semi-custom AI accelerators (XPUs) to connect directly into Nvidia’s NVLink high-speed interconnect fabric. NVLink has historically been exclusive to Nvidia GPUs — it is the high-bandwidth interconnect that links GPUs inside an H100 or H200 cluster. Opening it to third-party silicon is a significant architectural change.

Under the partnership:

• Marvell provides: custom XPUs and NVLink Fusion-compatible scale-up networking silicon
• Nvidia provides: Vera CPU, ConnectX NICs, BlueField DPUs, NVLink interconnect, Spectrum-X switches, and rack-scale AI compute integration

The result is a heterogeneous AI factory where hyperscalers building custom silicon can still plug into the Nvidia ecosystem rather than building fully independent infrastructure.

Silicon Photonics Collaboration

The two companies will also collaborate on silicon photonics, the technology that moves data using light rather than copper inside AI clusters. Marvell acquired photonic fabric specialists in 2025, making it one of the leaders in optical DSPs for data center interconnect. Nvidia’s backing accelerates the commercial deployment of optical interconnects at AI cluster scale.

Jensen Huang’s statement frames the urgency: “The inference inflection has arrived. Token generation demand is surging.” Higher token throughput demands more interconnect bandwidth — silicon photonics is the path to getting there without the heat and power constraints of copper at scale.

Strategic Logic: Co-opting Custom Silicon

The move is a direct response to the growing trend of hyperscalers (Google, Amazon, Meta, Microsoft) building custom AI chips to reduce Nvidia GPU dependence. Rather than fighting this trend, Nvidia is co-opting it.

By offering NVLink Fusion compatibility, Nvidia gives hyperscalers a reason to keep their custom silicon within the Nvidia ecosystem. A Google TPU or an Amazon Trainium chip that connects via NVLink can still run Nvidia’s CUDA software stack, management tools, and networking fabric. Nvidia keeps software revenue and interconnect revenue even when it loses the GPU sale.

This is a mature market move — the kind a platform company makes when it realizes fighting commoditization directly is less profitable than owning the platform layer that everything plugs into.

Market Context

The $2 billion investment follows Nvidia’s acquisition of Mellanox in 2020, which gave the company its networking position. The Marvell deal deepens that networking advantage while extending it into the custom silicon tier — exactly where Nvidia has seen its greatest competitive threats in 2024 and 2025.

Marvell’s market position is notable: the company already supplies custom AI chips to AWS, Google, and Microsoft. By formally partnering with Nvidia, both sides get something: Marvell gains access to NVLink’s performance advantages, and Nvidia gains visibility and integration points with the largest hyperscaler AI compute programs in the world.

What Developers Should Know

If you work with AI infrastructure at scale, NVLink Fusion means the gap between “Nvidia GPU cluster” and “custom silicon cluster” is narrowing in terms of interoperability. Future hybrid AI clusters — Nvidia GPUs for training, custom XPUs for inference — running on a common NVLink fabric become technically feasible. That changes the economics of AI infrastructure planning significantly.