Marvell Technology

Marvell Technology, Inc. is an American semiconductor company that designs chips for data infrastructure — storage, networking and, increasingly, custom AI accelerators. Marvell has become one of the principal beneficiaries of the custom silicon trend, in which cloud providers commission purpose-built AI chips rather than buying only off-the-shelf parts from NVIDIA.^[1][2]

Background

Marvell was founded in 1995 in Sunnyvale, California by Sehat Sutardja, his wife Weili Dai and Sehat's brother Pantas Sutardja, and began as a storage-chip company making read-channel controllers for hard disk drives. The company is headquartered in Santa Clara, holds over 10,000 patents worldwide, and reported annual revenue of US$8.2 billion for fiscal 2026 with roughly 7,500 employees.^[1]

Two acquisitions defined its modern AI position. The 2021 purchase of Inphi for about US$10 billion brought high-speed electro-optical interconnect technology and, critically, direct relationships with hyperscale cloud operators. Earlier acquisitions of Cavium (Arm server processors) and Innovium (Ethernet switch ASICs) completed a portfolio spanning compute, switching and connectivity — the three pieces of a data-centre rack.^[2][3]

Custom XPU Business

A custom XPU (a term covering any AI accelerator that is not a GPU) is a chip Marvell designs to a single customer's specification — for example, accelerating that cloud's own models and workloads. Marvell is widely reported to be the manufacturing partner behind Microsoft's Maia AI accelerators and previously helped bring Groq's Tensor Streaming Processor to market.^[3]

By fiscal 2026 the company reported 18 custom chip sockets under development across multiple customers, and its custom XPU sales reached roughly US$300 million per quarter — more than double the prior year — while AI-driven electro-optics reached US$709 million in the same quarter.^[3] Marvell frames the opportunity as two halves: the XPU itself, and XPU attach — the networking, memory and connectivity silicon surrounding each accelerator — projecting roughly US$30 billion in annual custom business by fiscal 2031, split between the two categories. In March 2026 Marvell announced a partnership around NVIDIA's NVLink Fusion platform, letting its custom XPUs interconnect with NVIDIA GPUs and Grace/Vera CPUs.^[4][5]

Why It Matters for AI

The custom silicon movement is a structural challenge to NVIDIA's dominance: every hyperscaler wants pricing leverage and workload-specific optimisation, and Marvell sells the design services and IP to achieve it. Analysts note that XPU attach — the connectivity layer — may grow faster than the accelerators themselves, because networking and optical chips ship with every AI rack regardless of which company designed the compute.^[3][5]

>See Also

• NVIDIA
• Broadcom
• TSMC
• High Bandwidth Memory
• Penang Semiconductor and AI Ecosystem
• AI Chip Export Controls