NVIDIA has put $2 billion into Marvell Technology and expanded their technical partnership, bringing a top custom chip designer into NVIDIA’s own interconnect ecosystem. While described as a partnership, this is a strategic move to strengthen NVIDIA’s position in AI data centers as large cloud providers look for custom silicon.
At the center of this strategy is NVLink Fusion, a rack-scale platform NVIDIA launched in May 2025. It lets third-party chips connect directly to NVIDIA’s high-speed interconnect. Samsung Foundry joined in October, and Arm followed in November. Marvell is now the biggest addition, bringing custom XPUs and networking to an ecosystem that already includes NVIDIA’s Vera CPU, ConnectX NICs, BlueField DPUs, and Spectrum-X switches.
Each NVLink Fusion setup needs at least one NVIDIA part, like a CPU, GPU, or switch. This means Marvell’s ASICs using this system still generate revenue for NVIDIA. Customers get more options, but NVIDIA keeps its income steady.
The optical interconnect part of the deal shows why Marvell’s recent acquisitions matter. In early 2026, Marvell bought Celestial AI for $3.25 billion, adding photonic fabric technology to its lineup.
As AI clusters grow larger, optical interconnects become more important because electrical signals can’t reach as far. The silicon photonics work in this partnership builds on Marvell’s existing technology.
This $2 billion investment is part of a trend. Earlier in March, NVIDIA made similar $2 billion investments in Lumentum and Coherent to boost laser production for co-packaged optics in its Quantum-X InfiniBand and Spectrum-X switches. NVIDIA is investing across its supply chain, from computing to networking and optics, to strengthen its position before competitors can offer new standards.
The AI-RAN part, which brings NVLink Fusion-enabled computing into 5G and 6G base stations using NVIDIA’s Aerial platform, probably won’t bring in much revenue soon. Still, it fits with NVIDIA’s bigger plan to make its infrastructure essential across the whole AI computing stack.
Broadcom is NVIDIA’s main competitor here. The company leads UALink, the industry’s open alternative to NVLink, and has expected that large cloud providers would eventually push back against NVIDIA’s closed system. By adding Marvell to the NVLink Fusion ecosystem, NVIDIA makes sure Marvell’s custom XPUs stay compatible with and dependent on NVIDIA’s infrastructure.
Marvell’s stock rose almost 13% on the day of the announcement, which investors saw as a positive sign, according to the report from CNBC. But the bigger point is that NVIDIA is making sure the whole AI infrastructure stack is aligned with NVIDIA before the industry’s main challenge moves from computing to connectivity.
