Marvell shares the latest AI progress: Custom computing is fully exploding, targeting a market of $94 billion

Marvell's recently concluded Custom AI Investor Day revealed a clear signal: AI chips are transitioning from "general GPU assembly" to "highly customized system-level collaboration." Marvell is no longer just a chip design company but a key infrastructure partner for AI companies building their computing foundation.

Marvell has revised its expectations for the potential market size (TAM) of data centers in 2028 from last year's $75 billion to $94 billion, with the market size for custom computing (XPU) and its supporting components reaching $55 billion.

Marvell pointed out that custom computing (XPU) and XPU Attach (supporting components) are the two fastest-growing areas. Over the past year, the custom computing market has grown by nearly 30%, while interconnect has grown by about 37%. Marvell's cloud revenue will fully transition to AI revenue in the future, reflecting the deep integration of AI in all applications and cloud infrastructure, with the cloud becoming the "factory" for AI.

AI Chips Have Entered the "System Customization" Era

Marvell emphasized that customized chips are key to meeting the demands of new workloads and are the future of AI infrastructure, occurring "in every cloud."

From core XPU chips to supporting high-bandwidth memory (HBM), interconnect control, power supply, and other "XPU Attach" modules, customers are building dedicated systems for their AI application scenarios.

Company CEO Matt Murphy pointed out at the conference: "If you are just starting to prepare for custom chips now, it is already too late." Marvell has been betting on customization since 2018 and has now formed a complete system.

TAM Significantly Revised Upward to $94 Billion

Based on last year's disclosed market expectation of $75 billion, Marvell has revised its target market size (TAM) for custom chips in 2028 to $94 billion, with a compound annual growth rate (CAGR) of 35%. Among them:

• Custom XPU: $40 billion, CAGR of 47%

• XPU Supporting Components: $15 billion, with a CAGR as high as 90%, nearly doubling every year

Marvell expects that cloud revenue will fully transition to AI revenue in the future, reflecting the deep integration of AI in all applications and cloud infrastructure.

The team also observed that workloads are becoming more diverse, and using dedicated chips has significant advantages in cost and performance. Marvell noted that compared to GPUs, XPUs can provide better performance in workloads. Currently, XPUs account for about 25% of the AI computing market, and Marvell believes that this share is expected to continue to increase with technological breakthroughs from ASIC suppliers

Changes in Customer Structure: Emerging Large AI Compute Builders and Sovereign AI Surface

The traditional cloud computing giants (Amazon AWS, Microsoft Azure, Google, Meta) remain the main players, but Marvell points out that Emerging Hyperscalers (new large AI compute builders, such as xAI, Tesla, etc.) are rising, beginning to build AI clusters independently and promote custom chips.

At the same time, Sovereign AI (locally driven AI infrastructure by governments) has also become a new growth direction, with multiple countries already investing, which will drive more demand for custom AI chips globally.

Marvell Has Secured 18 Customer Sockets, with Over 50 Projects in Negotiation for the Future

Currently, Marvell has secured 18 custom Socket (customer custom chip project units) projects, most of which are long-term multi-generation projects:

• Among them, 5 are core XPU projects

• 13 are related to XPU supporting components

In addition, Marvell is actively following up on over 50 new project opportunities, covering more than 10 customers, with a potential lifecycle revenue totaling up to $75 billion.

Marvell expects multiple Sockets to enter mass production in 2026-2027, particularly a "heavyweight XPU," which will be a key turning point for accelerating revenue.

Marvell has set a goal to achieve a 20% market share in the custom computing market by 2028. The company revealed that a few years ago, its market share was less than 5%, and it has now reached 13%. A 20% share is the "best estimate" and a reasonable next target.

When asked whether they have "deeply participated in the development of the next-generation project after the current 3-nanometer design" (with Amazon as the main customer), Marvell clearly answered "of course," and explained that the product iteration cycle requires them to develop chips in parallel (confirming the T4 and MAIA 3 projects).

Technical Breakthroughs: SRAM, HBM, and Die-to-Die Interconnect

Marvell showcased several key module technological breakthroughs at this event:

• 2nm Custom SRAM: Bandwidth density is 17 times that of mainstream IP on the market, with standby power consumption reduced by 66%

• Self-developed HBM Solution: Through bottom-layer Die-to-Die interconnect, it frees up 1.7 times the usable core area, with power consumption reduced by 75%

• Die-to-Die communication: Bandwidth density exceeds 10 Tbps/mm, with power consumption as low as sub-pJ/bit level.

• Co-Packaged Optics: Integrates optical communication directly into the package, addressing the low energy efficiency issues of traditional copper wires over long distances.

These modules are the foundational capabilities supporting the expansion of high-performance AI training and inference architectures