NVIDIA's $2B Marvell Bet
NVIDIA is investing $2 billion in Marvell to deepen an NVLink Fusion partnership and build a more tightly integrated AI chip ecosystem, a move designed to harden NVIDIA's infrastructure position against open alternatives. The deal signals more vertical integration between networking and GPU vendors — the kind of stack consolidation that shapes future edge and datacenter platforms used by robotics and AV systems. (techfundingnews.com)
NVIDIA announced on March 31, 2026 that it will invest $2 billion in Marvell and expand a partnership that formally links Marvell’s custom accelerator and networking products into NVIDIA’s data‑center platform, and the companies said they will jointly develop light‑based interconnect components used to move large volumes of data inside and between racks. (nvidianews.nvidia.com) Marvell’s shares jumped in intraday trading after the announcement, and analysts noted this follows a recent pattern of NVIDIA making multiple $2 billion strategic investments in companies that sit around its AI stack. (cnbc.com) NVLink Fusion is NVIDIA’s name for a rack‑scale platform that connects processors, accelerators, networking and storage so customers can assemble semi‑custom AI systems — “semi‑custom” meaning hardware that is partly tailored to a particular customer or workload rather than fully off‑the‑shelf. The press release says Marvell will supply “XPUs” (custom accelerator chips designed for a customer’s AI workloads) and scale‑up networking that is explicitly compatible with NVLink Fusion. (nvidianews.nvidia.com) The announcement lists specific pieces that will be made to interoperate and what those names mean: an XPU is a custom AI accelerator chip built for specific workloads, silicon photonics refers to on‑chip and co‑packaged light transmitters/receivers used to move data at high bandwidth with lower power, BlueField DPUs are data‑processing units that offload networking and storage tasks from general‑purpose CPUs, ConnectX NICs are high‑speed network interface cards, and Spectrum‑X are NVIDIA’s high‑performance switches for rack‑scale fabrics. (nvidianews.nvidia.com) (marvell.com) The companies also said they will push into telecom infrastructure by integrating NVIDIA Aerial (their software/hardware stack for running AI in radio access networks) with Marvell’s optical and networking technology to target 5G and future 6G deployments — in other words, the deal explicitly covers using carrier networks as parts of AI compute fabrics. (nvidianews.nvidia.com) Marvell has been building a pipeline of custom XPU projects and says it is engaged in multiple multigenerational custom design wins with large cloud customers; the company has documented a program of XPU and XPU‑attach devices for hyperscalers, and NVIDIA’s release frames Marvell’s designs as being supported inside the NVLink Fusion ecosystem. (marvell.com) (nextplatform.com)
