As cloud providers, AI startups, and sovereign entities scale out their data center architectures, Marvell sees increasing demand not only for primary compute engines—such as custom CPUs, GPUs, and accelerators—but also for a broad range of supporting silicon elements including network interface controllers, power management ICs, and memory expanders.
At Marvell’s 2025 AI Investor Day, CEO Matt Murphy outlined the company’s growing role in powering AI infrastructure, emphasizing two key developments shaping the market: the rise of new AI infrastructure builders beyond the traditional top four hyperscalers, and the rapid emergence of XPU-attach components as a significant new category of custom silicon. These trends, Murphy noted, are contributing to a much larger and more diverse total addressable market than previously projected.
As cloud providers, AI startups, and sovereign entities scale out their data center architectures, Marvell sees increasing demand not only for primary compute engines—such as custom CPUs, GPUs, and accelerators—but also for a broad range of supporting silicon elements including network interface controllers, power management ICs, and memory expanders.
Surging AI Infrastructure Investment
Murphy detailed how global data center capital expenditures have exploded, driven by hyperscaler and sovereign AI buildouts. Top-tier U.S. hyperscalers—AWS, Microsoft, Google, and Meta—grew their combined capex from $150 billion in 2023 to over $300 billion in 2025, with forecasts pushing past $1 trillion globally by 2028. Marvell believes a large share of this spend will be directed at highly customized silicon platforms.
Expanded Market Opportunity
Marvell has revised its total addressable market (TAM) upward to $94 billion by 2028, a 25% increase over last year’s estimate. The breakdown:
- $40B for custom XPU (accelerators, CPUs, GPUs, DPUs) — growing at 47% CAGR
- $15B for custom XPU Attach (NICs, power ICs, security co-processors, memory expanders) — growing at 90% CAGR
- The remainder spans switching, interconnect, and storage—sectors in which Marvell is also active
Murphy emphasized that XPU Attach is a breakout category, noting that “AI compute platforms are no longer defined by a single chip. They’re complex systems with an explosion of custom sockets—each one a new opportunity.”
Murphy concluded with a bullish outlook: “Last year, we had three custom compute sockets and a $75 billion TAM. This year, we’re at 18 sockets, a $94 billion TAM, and a growing pipeline of 50+ opportunities. The AI infrastructure market is evolving fast—and Marvell is right in the middle of it.”
Custom Socket Wins and Momentum
Marvell has secured 18 custom sockets to date:
- 5 XPU sockets with leading U.S. hyperscalers
- 13 XPU Attach sockets (including production and in-design engagements)
- 6 sockets (2 XPU + 4 Attach) with emerging AI infrastructure players such as xAI and Tesla
Each socket win, Murphy explained, represents a multi-generational commitment. “Once you’re in, you’re in pole position for the next design,” he noted. With architectural refresh cycles accelerating, new sockets are emerging faster than ever, dramatically expanding the opportunity landscape.
Marvell is tracking more than 50 active custom silicon engagements—a mix of XPU and Attach programs—across over 10 customers. These span cloud hyperscalers, emerging AI startups, and national AI infrastructure initiatives. The company estimates that these pipeline opportunities represent $75 billion in lifetime revenue potential, over and above the 18 sockets already secured.
Murphy stressed that this market is no longer dominated by a handful of mega-sockets. “Back in 2023, one socket accounted for 75% of the TAM. By 2028, no single socket will be more than 10–15%. That’s a massive diversification—and it plays right into our strengths.”
https://www.marvell.com/company/events/custom-ai-investor-event.html