Direct:
Marvell Technology (MRVL) just framed its growth around two reinforcing moves. It wants to win custom AI silicon designs at the hyperscalers, then sell the connectivity that ships with every one of those sockets. The fiscal Q1 2027 call, reported May 27, put hard numbers behind it. Data center revenue hit a record $1.83 billion, up 27% year over year and 11% sequentially, against $2.42 billion in total revenue. At roughly three-quarters of the business, data center now carries the entire company.
The part worth focusing on is “attach.” When Marvell lands a custom XPU at Amazon (AMZN), Google (GOOGL), Microsoft (MSFT), or Meta (META), it tries to attach the adjacent custom silicon around that accelerator too: the NICs, retimers and co-processors, CXL and PCIe switching, and the optical interconnect feeding the chip. Winning the accelerator is what gets Marvell onto the platform, and the attached content is what makes the socket lucrative over time. There are 18 active custom programs now, a dozen of those devices already at the big four. Management puts custom silicon, roughly $1.5 billion in fiscal 2026, on a line of sight to more than $10 billion a year by fiscal 2029, with custom revenue set to more than double in fiscal 2028. Those figures come off an earnings call, so treat them as targets backed by signed design wins rather than booked revenue.
Marvell has been buying its way into that interconnect layer. It closed Celestial AI in February for $3.25 billion up front, with earnouts that take it to as much as $5.5 billion (the Photonic Fabric optical interconnect), picked up silicon photonics startup Polariton, and bought PCIe and CXL switch maker XConn for around $540 million. The logic is that as AI clusters scale out, moving data between accelerators becomes the bottleneck, and Marvell wants to own that layer across copper, optics, and silicon photonics.
Nvidia (NVDA) put $2 billion into Marvell in March, with an agreement to make Marvell's custom XPUs and scale-up networking compatible with NVLink Fusion. It gives Nvidia a stake in systems built around accelerators meant to compete with its own GPUs.
The caveat is Broadcom (AVGO), still holding somewhere around 60 to 70% of the custom ASIC market and adding named customers. Marvell is the clear number two. Google reportedly sources custom chips from both, so at the hyperscaler level this looks like a two-supplier market.
Drafted with AI assistance against parallel reporting.
Sources:
- DigiTimes, “Marvell eyes XPU and attach wins as pillars of custom silicon growth” (May 28, 2026, paywalled stub: headline and lead only)
- DigiTimes, “Marvell expands silicon photonics, switching, and interconnect roadmap” (May 28, 2026, paywalled stub: headline and lead only)
- Marvell Technology, Q1 FY2027 financial results and earnings call (May 27, 2026): data center $1.83B, total revenue $2.42B, over $10B custom silicon line of sight for FY2029
- The Motley Fool, “Marvell (MRVL) Q1 2027 Earnings Transcript” (May 27, 2026)
- Marvell Technology, “Marvell Completes Acquisition of Celestial AI” (Feb 2, 2026): $3.25B up front, up to $5.5B with milestone earnouts
- DatacenterDynamics, “Marvell acquires PCIe and CXL switch provider XConn Technologies for $540m” (2026)
- DatacenterDynamics, “Marvell acquires silicon photonics device startup Polariton Technologies” (2026)
- Tom's Hardware, “Why Nvidia just poured $2 billion into AI ASIC competitor Marvell” (Mar 31, 2026)
- TradingKey, “Marvell's Data Centre Revenue Grew 42% and Its Custom AI Chips Are Landing at Hyperscalers” (2026)
Reddit: https://ift.tt/8O3W9ZC