Santa Clara, California, March 31, 2026, 06:17 PDT
- Nvidia has poured $2 billion into Marvell, broadening their AI infrastructure collaboration with the rollout of NVLink Fusion.
- Marvell is now projecting nearly $11 billion in revenue for fiscal 2027, coming off a record-setting fiscal 2026 and a 21% jump in fourth-quarter data-center revenue.
- Marvell’s latest deal strengthens its move into custom AI chips, networking, and optical links—key areas where Broadcom is competing for hyperscaler business, too.
Nvidia just put $2 billion into Marvell Technology and is tightening its AI infrastructure ties with the company, bringing Marvell even closer to Nvidia’s orbit. Marvell shares jumped sharply in premarket trading Tuesday after news of the deal broke.
Investors are shifting attention past the headline GPU battles and zeroing in on the underlying plumbing of AI clusters — networking, switching, and optical connections that actually move data around. Marvell projected on March 5 that fiscal 2027 revenue could come in near $11 billion, after a 21% jump in fourth-quarter data-center sales. Management expects that pace to pick up throughout the year.
Marvell is set to handle custom processors and the scale-up networking—those high-speed connections that tie together major AI systems—under the new deal. Nvidia, for its part, will bring in its Vera CPU, ConnectX network cards, BlueField DPUs, and Spectrum-X switches. Both firms plan to collaborate on silicon photonics, aiming to push data faster and more efficiently using light on chips.
Nvidia’s Jensen Huang declared, “The inference inflection has arrived.” Marvell’s Matt Murphy zeroed in on what matters to his company: “growing importance of high-speed connectivity” as AI models balloon in size. Marvell’s target isn’t Nvidia’s main GPU turf—it’s the custom silicon and interconnects that tie everything together. Marvell Technology, Inc.
Marvell posted fiscal 2026 revenue of $8.195 billion, a 42% jump over last year—a record for the company. Looking ahead, guidance for first-quarter fiscal 2027 landed at roughly $2.4 billion, give or take 5%. CEO Matt Murphy flagged an unprecedented pace of bookings, with fiscal 2026 design wins also at their highest ever.
The lineup at Marvell has grown quickly. Reuters noted earlier this month that demand is picking up for the company’s optical digital signal processors—the chips powering high-speed optical links—as well as for custom ASICs built for dedicated tasks. Then on March 17, Marvell rolled out what it described as the first 260-lane PCIe 6.0 switch designed specifically for scaling up AI infrastructure.
Marvell isn’t the only one making that move. Broadcom is stepping up its custom AI chip efforts for hyperscalers, too. Earlier this month, Reuters said Broadcom is forecasting more than $100 billion in AI-chip sales next year. Both Marvell and Broadcom are working with cloud providers on custom processor designs—an alternative, or supplement, to Nvidia’s off-the-shelf AI chips.
Marvell president and COO Chris Koopmans, speaking after the March earnings release, described hyperscalers’ spending as “still growing massively.” For Kinngai Chan at Summit Insights, the improved forecast seemed “more of a relief for investors” since Marvell had trailed some of its AI competitors. Reuters
Still, even with a tighter relationship with Nvidia, the central risk remains. Marvell needs to convert those design wins into real shipment volumes—custom silicon, optics, switching all included—and if hyperscalers slow construction, defending its 2027 and 2028 targets gets tougher.
The announcement on Tuesday expands on an alliance first mapped out in May 2025—back when Marvell said its custom cloud silicon would be paired with Nvidia’s NVLink Fusion, aiming to let customers assemble semi-custom AI systems. After a year dominated by talk of GPUs, this step shows the AI race is pushing deeper into the gear that links all those machines together.
