Evercore: Marvell's Networking IP Is Becoming the Backbone of Every Major AI Architecture — and Hyperscalers Are Noticing
Evercore Global TMT Conference, June 2, 2026 — CFO Willem Meintjes and SVP IR Ashish Saran lay out an expanding revenue roadmap and a quietly dominant competitive position
Jensen Huang's Endorsement and What It Actually Means
When NVIDIA's Jensen Huang singled out Marvell as the next trillion-dollar market cap company at a recent event in Taiwan, the comment made headlines. But the more substantive story, as Marvell's CFO Willem Meintjes and SVP of Investor Relations Ashish Saran explained at the Evercore Global TMT Conference on June 2, is what sits underneath that endorsement: a multi-year thesis about connectivity becoming the defining bottleneck in AI infrastructure is now visibly playing out, and Marvell has spent a decade assembling the IP stack to address it.
Evercore senior semiconductor analyst Mark Lipacis noted that checks conducted over the past six months across roughly two dozen sources at hyperscalers consistently returned the same finding — Marvell is increasingly viewed as a strategic partner, not merely a component supplier. That shift is showing up across both the networking and custom silicon sides of the business, and it is now being reinforced by a formal three-pillar agreement with NVIDIA that covers optics collaboration, network fungibility between merchant and custom AI infrastructure, and AI RAN baseband integration.
The NVIDIA Partnership Is Bigger Than the Headlines Suggested
Saran walked through the three pillars of the NVIDIA relationship in detail, and each one addresses a structurally distinct problem. The first extends a longstanding collaboration on pluggable optics into the emerging scale-up optics market, where both companies are now co-developing the path forward for near-packaging optics and co-packaged optics. The second, and arguably the most commercially significant, tackles the fragmentation problem inside hyperscaler data centers. As Saran put it, "Today, if you think about hyperscalers, they've really got two different infrastructures. They've got a completely merchant infrastructure, NVIDIA primarily, and then they obviously have custom. And they don't talk to each other. In the long run, that's not how you want to build a data center. You want complete fungibility. That's where Marvell comes in." The third pillar integrates NVIDIA compute capability into Marvell's OCTEON baseband processors, enabling software-defined RAN upgrades without rebuilding base station infrastructure from scratch. The practical read-through is that Marvell now sits at the junction of merchant and custom AI infrastructure in a way that expands the addressable market for both companies.
Revenue Targets Keep Moving Up, and the Math Is Getting Large
Meintjes recapped the progression of Marvell's long-range revenue targets with striking clarity. In September of last year, the company introduced guidance of $10 billion for the current fiscal year and $13 billion for the year following. One quarter later, those numbers moved to $11 billion and $15 billion respectively. After last week's earnings, they stand at $11.5 billion for this year and $16.5 billion for next. Each upward revision has been demand-driven, supported by ongoing tightness in supply that the company says it has managed effectively through a rigorous five-year forecasting process shared proactively with its supply chain.
The interconnect business is the primary engine of this acceleration. Saran described it simply: "Our interconnect business is absolutely on fire. That thing is growing at 70% plus for this year." Within interconnect, the scale-out business — anchored by PAM DSPs, TIA drivers, and broadband analog components — is growing even faster, while scale-across, the coherent-technology-based business connecting data centers over longer distances, is just beginning to contribute meaningful revenue. Marvell expects both to be billion-dollar-plus contributors in the outer years, alongside a TIA and driver business that Meintjes noted was underappreciated by most investors and already stands at $1 billion in annual revenue today.
Scale-Up Optics: Celestial AI Is in Productization, Not Development
One of the more concrete updates from the conference was the status of Celestial AI, the photonics acquisition that has drawn significant investor attention and some skepticism given the absence of production revenue to date. Saran was direct in addressing this: "The Celestial product is basically well past development. At this point, it's really entering a manufacturing cycle. We've lined up sources of supply. It's going to be in production essentially next year." The company continues to target $500 million in quarterly run-rate revenue exiting calendar 2027, doubling to $1 billion exiting calendar 2028, which implies roughly $1 billion in cumulative revenue over approximately 15 months from production start. Importantly, Marvell has broadened the framing of this opportunity — the $300 million revenue target for the nearer term now encompasses the full suite of scale-up optics products, including MRM with TSMC, MZM technology derived from its legacy DCI business, and EAM, not just the Celestial-specific solution.
Looking further out, Marvell has acquired a company called Polariton, which Saran described as pre-revenue but possessing modulator technology capable of operating at speeds up to ten times higher than current solutions. The acquisition signals that Marvell is not treating photonics as a near-term product cycle but as a decade-long platform investment.
The XPU Business: IP Differentiation Is the Real Moat, Not Design Services
The custom ASIC business continues to generate confusion in the market, partly due to conflation of Marvell with contract design service firms operating in Taiwan. Saran drew the distinction plainly. Companies without a merchant semiconductor business have no proprietary IP — they perform physical design and layout work at the tail end of the chip development process. Marvell's engagement with XPU customers is entirely different: hyperscalers are accessing Marvell's networking IP — SerDes, die-to-die interfaces, custom HBM interfaces, advanced packaging know-how — which is directly relevant to the construction of large, disaggregated, multi-die AI accelerators. "The reason we're in the custom business is actually because of the expertise we've established from our merchant business," Saran said.
On size and trajectory, the XPU business is expected to grow from roughly $1.5 billion to approximately $2 billion this fiscal year, then more than double to $4 billion-plus the following year. Marvell has broken this growth into thirds: one-third from existing programs, one-third from ten-plus new XPU attach programs entering or approaching production, and one-third from a new XPU program. The attach programs — primarily custom NICs and CXL controllers — were each tracking toward $1 billion in the outer year when first communicated, and Meintjes indicated they are on track to reach or exceed that figure.
Agentic AI Is an Unmodeled Tailwind for CXL and NICs
One of the more underappreciated points from the conference is that Marvell's $3 billion-plus XPU attach revenue target for fiscal 2028 was built before the emergence of agentic AI as a meaningful workload driver. Saran was explicit: "The numbers we've outlined so far don't really include the agentic impact, just to be clear." The logic is straightforward — more CPUs in AI server configurations, driven by inference scaling and agentic workloads, means more NIC sockets and more CXL memory expansion demand. The company also flagged two additional XPU attach categories beyond NICs and CXL: storage accelerators, which offload SSD and HDD management from the central processor, and security accelerators, a Cavium-derived capability that Saran described as increasingly relevant as enterprises consume AI infrastructure as a service and require hardware-level encryption offload. Neither of these categories appears to be fully embedded in published forecasts.
PAM DSP Share Position Remains Intact Despite Market Scale
On the DSP competitive question — a recurring investor concern since the Inphi acquisition — Meintjes was straightforward: the anticipated market share erosion simply has not materialized. "When we acquired Inphi, there was always this expectation that we're going to lose market share. And frankly, it just really hasn't happened." At 1.6T, which entered production last year and is becoming the volume technology this year, Marvell was again first to market. The 3.2T generation is next on the roadmap. Saran noted the interconnect business is expected to substantially outgrow cloud CapEx again in fiscal 2027, which given current CapEx trajectory implies a very large absolute number. While the company acknowledged that at its current scale some incremental competition is inevitable, it described its share position across each generation as a very strong majority and showed no sign of concern about the near-term competitive picture.
Scale-Across Is Just Beginning, and the TAM Is Moving Higher
The scale-across opportunity — coherent interconnect connecting data centers over distances too long for PAM technology but now increasingly relevant for AI cluster networking — was described as being at the very beginning of its revenue ramp. Meaningful scale-across revenue likely starts in earnest next year, gated by customer migration to 1.6T bandwidth, which is the required foundation for economically viable scale-across deployments. Marvell has pulled forward its 1.6T coherent DCI module to sample in the second half of this year, built on a two-nanometer coherent DSP that the company claims is first to market. Saran characterized the original TAM projections for this category as almost certainly too low, not just because CapEx has surprised to the upside, but because the architectural complexity of modern AI workloads — mixture of experts, long-context inferencing, disaggregated memory — is generating far more inter-cluster traffic than was anticipated even eighteen months ago.