Marvell Technology Sees AI Connectivity as Next Big Infrastructure Bottleneck

Key Points

• Connectivity is emerging as the next major AI infrastructure bottleneck as systems scale across chips, racks, clusters and data centers. Marvell expects optical networking, interconnects and custom silicon to drive a multiyear expansion.
• Marvell is building its position through acquisitions including Inphi and Celestial AI, combining silicon photonics, near-package optics and co-packaged optics; Murphy expects near-package optics to gain adoption before co-packaged optics.
• Memory expansion and custom silicon are growing opportunities: CXL products could become a billion-dollar to multibillion-dollar business, while Marvell has secured more than a dozen “XPU attach” designs and reports over $2 billion in quarterly AI-related data-center activity.

Marvell Technology NASDAQ: MRVL Chairman and CEO Matt Murphy said the artificial intelligence infrastructure market remains in its early stages, with connectivity emerging as a central constraint as computing, memory and storage systems scale.

Speaking at The Six Five Summit 2026, Murphy said the company’s decade-long shift toward what it calls data infrastructure has positioned it for an AI-driven expansion in networking, optical connectivity, custom silicon and memory expansion. He said data infrastructure represented less than 10% of Marvell’s revenue when the company began that shift, and is expected to account for more than 80% of revenue in the future. Murphy also said Marvell has grown more than fivefold over that period.

“The AI application became kind of the killer app of data infrastructure,” Murphy said. “We are still at the very, very early stages of the deployments,” particularly in developing the technology needed to scale AI systems broadly.

Connectivity moves to the forefront

Murphy said the industry’s initial focus was on accelerators, including GPUs, XPUs and custom application-specific chips, followed by a pronounced cycle in memory and storage. The next major challenge, he said, is connecting those resources within chips, racks, clusters, data centers and eventually across multiple data centers.

“The connectivity that needs to get unleashed and the interconnect” will enable processing and memory resources to move data between chips, racks and network layers, Murphy said.

He pointed to Marvell’s acquisition of Inphi, which closed in 2021, as an important part of its optical networking strategy. Inphi brought digital signal processor technology, silicon-germanium analog components and silicon photonics capabilities to Marvell, according to Murphy. He said the combined operations had shipped millions of silicon photonics units and accumulated 15 billion hours of reliability data over the prior decade.

Marvell also acquired Celestial AI late last year to expand its co-packaged optics capabilities. Murphy said the company combined Celestial AI’s photonic fabric and co-packaged optics technology with Marvell’s existing work in data-center interconnect modules and near-package optics.

Murphy said near-package optics could see larger adoption before co-packaged optics, though both technologies are expected to become part of future AI infrastructure deployments.

Memory expansion and custom silicon

Murphy said Marvell’s CXL memory-expansion products are gaining importance as AI inference systems require significant memory capacity for key-value caching and customers seek ways to use memory more efficiently. CXL, or Compute Express Link, was originally intended to let conventional servers add memory without requiring customers to add processors at the same rate, he said.

Marvell offers CXL memory expanders, switches and retimers, along with custom memory-expansion products. Murphy said customers are accelerating adoption of such technologies as they look for alternatives to purchasing as much memory as previously anticipated.

“It has become a real thing,” Murphy said, describing memory expansion as a potential billion-dollar to multibillion-dollar future business for the company.

On custom silicon, Murphy said Marvell entered the market through its 2019 acquisition of Avera Semiconductor, a former GlobalFoundries business with roots in IBM’s custom chip-design operations. Marvell moved the business onto TSMC’s manufacturing platform and targeted data-center applications, he said.

Murphy said custom AI accelerators are now being deployed for training and inference workloads, although he does not expect custom chips to replace merchant silicon across the entire market. Instead, he said, the two approaches will coexist as hyperscale customers tailor systems to their specific workloads.

Beyond accelerators, Murphy highlighted an “XPU attach” opportunity that includes customized networking interfaces, memory expansion and security components surrounding AI processors. He said Marvell has won more than a dozen such designs across major hyperscalers and other customers.

Scaling clusters across racks and data centers

Murphy said the next phase of AI infrastructure will involve more extensive “scale-up” configurations inside racks and across groups of racks, as well as “scale-across” systems that coherently connect clusters in separate data centers.

He said such deployments have not yet reached mass scale and could unlock additional compute and memory capacity while reducing costs and improving performance. Marvell’s data-center business is currently above $2 billion per quarter in AI-related activity, Murphy said.

“This era of connectivity… is going to unleash a whole new wave of innovation,” Murphy said. “It’s going to enable new use cases. It’s going to enable cost to come down, performance to go up.”

Murphy said the size of the addressable market and the range of available technology opportunities are unlike anything he has seen during his tenure leading the company, while emphasizing that AI silicon and networking development will unfold over multiple years rather than all at once.

About Marvell Technology (NASDAQ:MRVL)

Marvell Technology Group is a global semiconductor company that designs and develops integrated circuits and related software for data infrastructure, networking, storage and connectivity markets. The company's product portfolio includes system-on-chip (SoC) solutions, Ethernet physical-layer transceivers (PHYs), switch and switch silicon, optical interconnect components, storage controllers, and security processors. Marvell's technology is used to enable high-performance data centers, carrier networks, enterprise and cloud storage, as well as connectivity in automotive and industrial applications.

Founded in 1995 and headquartered in Santa Clara, California, Marvell has grown through both organic development and strategic acquisitions to broaden its capabilities across networking and data interconnect.

This instant news alert was generated by narrative science technology and financial data from MarketBeat in order to provide readers with the fastest reporting and unbiased coverage. Please send any questions or comments about this story to contact@marketbeat.com.

Continue following MarketBeat

Add MarketBeat as your preferred source on Google to see our latest stories in your feed.

Add As Preferred Source