Marvell (MRVL) and 6G: A Shrinking RAN Franchise Bets on the Nvidia Alliance
Marvell enters the 6G cycle from a position of strength that is quietly eroding. The company supplies the custom baseband silicon at the heart of the non-Chinese RAN — OCTEON Fusion processors and OCTEON DPUs sit inside Nokia’s ReefShark chipsets and Samsung’s massive-MIMO base stations. That franchise defined Marvell’s role in 5G. It does not obviously survive into 6G intact, and the company’s maneuvering over the past year reads as an attempt to convert a declining incumbency into something more durable.
The Legacy Franchise Is Contracting
The math is unforgiving. Worldwide RAN spending fell to roughly $35 billion in 2025, down from $45 billion in 2022, with no recovery expected. Silicon is a small fraction of that. Marvell earned about $542 million from all telco customers in the twelve months to November — a rounding error against a company now valued on its data-center trajectory.
Worse, the two customers that matter are both moving away. Nokia, after accepting a $1 billion investment from Nvidia, disclosed plans to build 5G-Advanced and 6G equipment on Nvidia GPUs, framing the shift from proprietary to general-purpose hardware as a way to redirect spend toward software differentiation. Samsung has publicly forecast the end of custom baseband silicon in the RAN and is developing its own alternatives. Both statements point at the exact seat Marvell occupies. Custom silicon still outperforms general-purpose compute on baseband tasks, but developing bespoke ASICs for a single vendor in a shrinking market is becoming impossible to justify on returns.
Two Bets Against Obsolescence
Marvell’s response has two prongs, and they sit in tension with each other.
The first is a pitch to make itself the industry’s shared baseband platform — one common silicon design for every non-Chinese RAN vendor, excluding Huawei and ZTE. The logic is a compromise between two bad extremes: general-purpose GPUs and CPUs carry better economics but suboptimal baseband performance, while per-vendor custom silicon excels at the job but no longer pays for itself. A single chip amortized across the whole industry threads that needle. Whether competitors agree to standardize on a rival’s platform is the entire question, and history offers little encouragement.
The second is alignment with Nvidia rather than resistance to it. On March 31, 2026, Nvidia invested $2 billion in Marvell and the two announced a partnership linking Marvell into Nvidia’s AI-RAN ecosystem through NVLink Fusion. Marvell contributes custom XPUs, high-performance analog, optical DSP, and silicon photonics; Nvidia supplies Vera CPUs, ConnectX NICs, and BlueField DPUs. Management has indicated OCTEON processors will be enhanced to interoperate directly with Nvidia GPUs, letting operators run 5G and 6G radio workloads alongside AI applications on shared hardware.
What the $2 Billion Actually Buys
The partnership contains a contradiction worth stating plainly. Nvidia’s AI-RAN thesis — GPUs displacing fixed-function silicon across the radio access network — is a direct threat to the baseband business Marvell is trying to defend. The two companies are collaborators on interconnect and photonics and competitors on the RAN itself.
Read against Nvidia’s other recent moves, the $2 billion looks less like conviction in Marvell’s RAN chips than like access to its optical and photonics expertise. It fits the same pattern as Nvidia’s stakes in Coherent and Lumentum: securing the high-speed connectivity layer for AI factories, not the baseband layer for cell sites. For Marvell, the capital and the ecosystem tie-in are real; the strategic implication for its legacy franchise is more ambiguous.
The Franchise That Actually Matters
Framing Marvell primarily as a 6G story misreads the company. RAN is a small, declining slice of a business now driven by custom AI ASICs for hyperscalers — Amazon Trainium and Inferentia, Google TPU, Microsoft, Meta — and by optical interconnect for AI data centers, reinforced through the Celestial AI and Polariton acquisitions. That is where the growth, the margin leverage, and the valuation premium live.
Seen that way, Marvell’s genuine 6G role is dual. In RAN baseband, it holds a defensive and uncertain position under pressure from both its own customers and its new partner. In the optical and custom-silicon plumbing that AI-native 6G networks will run on — and that the AI infrastructure behind them depends on — it holds a far stronger one. The connectivity fabric of 6G is a better business for Marvell than the radios themselves.
The Timing Caveat
None of this resolves soon. 6G radio standards under IMT-2030 are not expected to be finalized until roughly the end of 2030, which means production silicon does not land until 2031 at the earliest. Everything currently announced — the shared-platform pitch, the Nvidia alliance, the GPU-interoperable OCTEON roadmap — is positioning ahead of a standard that does not yet exist. The competitive structure of 6G RAN silicon is being set now, years before there is a market to sell into. For Marvell, the risk is that the seat it is fighting to keep will be a smaller one by the time the cycle actually arrives.