Nvidia Backs Marvell With $2B Bet on Silicon Photonics and NVLink Fusion

Nvidia Corp. is buying a $2 billion strategic stake in Marvell Technology Inc. and cementing a partnership to co-develop silicon photonics chips and custom XPUs that plug straight into Nvidia’s AI-factory blueprint, according to announcements Monday from both companies and coverage by Bloomberg, CNBC, and GlobeNewswire. The deal gives Marvell immediate credibility—its shares jumped 11–13 % on the news—and positions Nvidia to tighten control over the supply chain for next-gen data-center networking.

The collaboration centers on four concrete deliverables. First, Marvell will design custom XPUs (CPUs, DPUs, and accelerators) that speak Nvidia’s NVLink Fusion protocol, letting them drop into existing Nvidia racks with no software rewrites. Second, the duo will jointly craft silicon photonics components—tiny lasers and modulators etched onto chips—to push data between boxes at terabit speeds while burning less power than copper. Third, Marvell gains early access to Nvidia’s Vera CPU and ConnectX NIC roadmaps, ensuring pin-compatible parts. Fourth, both firms will publish unified firmware so hyperscalers can mix-and-match silicon without re-architecting clusters.

This is the fourth $2 billion check Nvidia has written in six months, following similar moves with CoreWeave, Nebius, and Coherent, CNBC notes. The pattern is deliberate: instead of buying whole companies, Nvidia is buying influence—securing preferential supply, locking in architectural compatibility, and preventing rivals from courting key partners. Marvell brings high-speed SerDes IP and a foundry relationship with TSMC that Nvidia itself lacks, giving it leverage over Broadcom and Intel in the custom-silicon race.

Broadcom, Marvell’s chief rival in custom ASICs, now faces a customer pipeline that increasingly has Nvidia whispering in its ear. Intel, pushing its own photonics and Gaudi accelerators, must convince hyperscalers to adopt a second architecture while Nvidia-Marvell offers turnkey integration. For cloud giants, the deal lowers switching costs: they can buy Marvell parts that behave like native Nvidia silicon, preserving their software investments. Smaller AI labs win too—start-ups that couldn’t afford full Nvidia stacks can now license Marvell XPUs that drop into the same ecosystem.

The photonics angle is bigger than a faster cable. Silicon photonics replaces heavy copper backplanes with light, cutting power per bit by 5-10× and enabling kilometer-scale AI clusters without signal loss. Nvidia tried home-grown photonics in 2022 and shelved it; partnering lets it leapfrog to market using Marvell’s proven IP while sharing R&D costs. If the duo hits volume in 2027, hyperscalers could see rack-to-rack latency drop below 200 nanoseconds—fast enough to treat a 10,000-GPU pod as a single logical device.

Marvell gets a near-term revenue bump—analysts at 247WallSt estimate $300-400 million in additional 2027 sales from NVLink-compatible chips—and a halo that lifts its valuation multiple. Nvidia gains a hedge against Broadcom price hikes and secures another photonics supplier ahead of 1.6 Tb/s networking rollouts. For developers, the first NVLink-Fusion-ready Marvell parts are slated to tape out late 2026, with reference boards arriving in Q2 2027. Expect SDK updates this summer that expose new low-latency collectives tuned for photonic links.