Marvell 2025 Industry Analyst Day: An In-Depth Analysis
Introduction: This Was No Ordinary Analyst Day
On December 9, 2025, Marvell Technology held its Industry Analyst Day in Santa Clara, California. On the surface it was the usual "walk analysts through the company's outlook" event, but in reality it read more like a public "transformation manifesto."
First, some key context:
In Q3 of fiscal 2026, Marvell delivered
revenue of US$2.075 billion, up 37% year over year.
But the earnings are just the cover of the story. What really matters is:
where the company's future growth engines will be.
The message from the event was very direct:
Marvell no longer wants to be a networking IC company making a bunch of "pretty good" products — it wants to be the "architect" of AI data center infrastructure.
The core pivot happens along two dimensions:
Corporate strategy and capital allocation:
Shifting resources away from the steady but fairly linear automotive business toward high-risk, high-upside optical interconnect and custom compute silicon.
Technology and product roadmap:
Betting aggressively on optical interconnect, 1.6T Ethernet, custom XPUs, and CXL memory architectures.
What exactly is Marvell betting on
The real meaning behind the Celestial AI acquisition
Why optical interconnect is becoming the "hard currency" of the AI era
How it finds its place between Broadcom and NVIDIA
1. Marvell's Strategic U-Turn: From Diversified to a Pure-Play AI Infrastructure Company
Marvell's transformation didn't happen overnight.
Its acquisitions and product changes over the past few years trace a clear trajectory:
Acquiring Inphi: adding high-end optical interconnect and DSP capabilities
Acquiring Innovium: filling the gap in data center switch silicon
Before 2025: still a broad-based IC company spanning "networking + storage + some emerging areas"
Then in 2025, the whole story changed in kind.
1.1 The Key Capital Allocation Move: Sell Automotive, Buy Optics
The most concrete expression of this transformation was two moves in quick succession:
Selling the automotive Ethernet business to Infineon for US$2.5 billion in cash.
Generating a pre-tax gain of about US$1.8 billion.
Lightening the asset base and significantly boosting the cash position.
Then immediately announcing the acquisition of Celestial AI for about US$3.25 billion.
A startup still in its "technology growth phase," not yet commercialized at scale.
Its technology focuses on in-package optical interconnect and Photonic Fabric.
From a capital allocation standpoint, this is a very clear signal:
Marvell is willing to give up a stable automotive market with predictable growth and pile all its chips onto an optical interconnect technology that is "highly uncertain, but offers exponential returns if it succeeds."
A long-cycle automotive market with dispersed customers and lengthy qualification,
has been swapped for fast-moving, rapidly iterating AI cloud data centers with highly concentrated customers.
This isn't "diversifying risk" — it's a very deliberate "concentration of risk."
1.2 Defining Infrastructure in the AI Era: The Data Center Is the Computer
Marvell CEO Matt Murphy put forward a key idea at the event:
In the AI era, the single server is no longer the unit of analysis — the entire data center is one "supercomputer."
Under this architecture, the truly hard problem shifts from "how powerful is one GPU" to:
How to make tens of thousands of GPUs/XPUs operate as a single logical entity
How to link model training and inference across racks, data halls, and even geographic locations
Within this framework, Marvell breaks its market opportunity into three dimensions:
Scale-Up:
GPU ↔ GPU and XPU ↔ XPU interconnect within a rack or server
Scale-Out:
Rack-to-rack Ethernet interfaces, moving from 400G and 800G to 1.6T
Scale-Across:
Long-haul, coherent-optics-based DCI (Data Center Interconnect) between data centers
The core thesis behind this segmentation:
Whether the eventual winner is NVIDIA, custom ASICs, or a mix of players, every AI chip needs massive high-bandwidth, low-latency interconnect. Marvell wants to be the "interconnect arms dealer" to every camp.
2. The Celestial AI Acquisition: A Big Bet on the Optical Interconnect Endgame
Unsurprisingly, the technical focus of this Analyst Day was Celestial AI.
Because it represents:
Marvell not just "selling optical modules," but "redefining the physical layer of in-package, chip-to-chip interconnect."
2.1 How Far Can Copper Go?
Let's start with the physics.
Inside today's high-end AI servers, GPU ↔ GPU links still mostly run over copper:
Passive copper DACs
High-speed PCB traces
Active AECs as a transitional solution
But as rates move from 112G/lane to 224G/lane, copper hits several hard limits:
The skin effect severely attenuates high-frequency components
Rising dielectric loss sharply shortens the usable reach
Maintaining signal integrity requires thicker cables, more complex connectors, and tighter routing space
Marvell stated plainly at the event: beyond 224G/lane, the effective reach of passive copper may drop to 1 meter or even less.
This means:
It is very hard to support "large GPU clusters spanning racks" with copper interconnect alone.
Maintaining high bandwidth and low latency between servers requires bringing in optics at scale.
Traditional pluggable optics can provide optical transmission, but the problems are:
High power consumption
Large size
Limited area and power density, making them unsuitable for large-scale GPU ↔ GPU scale-up
This is exactly the gap Celestial AI targets.
2.2 Photonic Fabric and OMIB: Turning Optical Interconnect into "In-Package Infrastructure"
Celestial AI's Photonic Fabric is essentially an "in-package, chip-level" optical interconnect architecture.
OMIB (Optical Multi-Chip Interconnect Bridge) is a key concept within it:
Conventional CPO mostly places optical engines at the package edge, so it remains constrained by "shoreline"
OMIB's design allows optical interconnect to be routed out from various locations on the die
In other words, it isn't "running a fiber" outside the package to connect two ASICs
but letting optical paths serve as a high-speed die-to-die backbone inside the package
The practical implications include:
Bandwidth density: far more interconnect capacity than traditional electrical I/O within the same package area
Interconnect topology: no longer limited to "edge routing," enabling more varied network structures inside the package
Scalability: maintaining an efficient interconnect architecture as the number of XPUs and chiplets grows
Marvell said at the event that Photonic Fabric can deliver:
About 25x higher bandwidth and memory capacity
About 10x lower latency
Roughly half the power of copper interconnect at the same distance
The exact figures may not be the point; what matters is that the "order-of-magnitude relationship" has fundamentally changed.
2.3 Thermal Stability and 3D Stacking: Attaching Optical Functions Directly to the XPU
Modern AI accelerators routinely draw 700W or 1,000W, with highly uneven temperature distribution inside the package.
And optical components are especially sensitive to temperature.
In this environment, putting optical interconnect inside the package is an extremely difficult problem in itself.
Celestial AI's technology stack includes:
An external laser source (ELS), avoiding placing the light source directly on the high-power die
A modulator design that is relatively insensitive to temperature changes
A package structure that supports vertical 3D stacking
These designs allow Photonic Fabric to achieve:
3D integration with the XPU
Interconnect paths shortened to the micrometer scale
Stable operation even in high-power, high-heat-flux environments
This isn't a single-component breakthrough, but a rewrite of the entire package-level system architecture.
2.4 What the Deal Structure Reveals About Timing and Confidence
The structure of the US$3.25 billion acquisition is itself revealing:
Upfront consideration:
About US$1 billion in cash
About US$2.25 billion in Marvell stock (about 27.2 million shares)
Additional performance-based earnout:
Up to about US$2.25 billion in stock
Trigger condition:
Celestial AI products must exceed US$2 billion in cumulative revenue by the end of Marvell's fiscal 2029
These terms reveal two things:
Marvell expects Photonic Fabric to enter real volume adoption around 2027–2028.
If it succeeds, the scale won't be a few hundred million dollars, but billion-dollar-level revenue contribution.
In one sentence:
Marvell knows full well this is a high-risk, high-barrier long-term bet — but if it pays off, it will directly reshape the rules of the game for AI packaging and interconnect.
3. Scale-Up: Co-opetition with NVLink and Extending Copper's Life
In the most central battlefield of the AI server, scale-up is currently NVIDIA's home turf.
3.1 NVLink's Walled Garden vs. UALink's Open Front
Today's GPU ↔ GPU high-speed interconnect market leaves customers with little choice:
Using NVLink means using NVIDIA GPUs, NVIDIA switches, and NVIDIA's packaging and software stack
For cloud providers, this is a high-performance, very sticky — but also heavily locked-in — architecture
Marvell's response is pragmatic and smart:
On one hand, actively participating in UALink
Working with AMD, Intel, Broadcom, and others
Aiming to build an open alternative to NVLink
Marvell plans to launch UALink-capable switches and related interconnect chips
On the other hand, partnering with NVIDIA on NVLink Fusion
Custom XPUs that Marvell designs for customers can integrate NVIDIA's NVLink IP
In other words, custom ASICs can still connect seamlessly into NVIDIA GPU clusters
For Marvell:
It can't ignore the closed ecosystem, nor can it give up on the open one. So it chooses to "bet on both sides," ensuring it won't be shut out of any possible future.
3.2 AEC and the "Golden Cable": Copper's Transitional Generation
Although optical interconnect is the long-term endgame, copper still has cost and maturity advantages in the near term.
Marvell uses AECs (Active Electrical Cables) to extend copper's life:
Embedding its own Alaska A DSPs at both ends of the cable
Allowing thinner cables while maintaining reach and speed
Enabling more flexible in-rack cabling and reducing weight and bulk
This approach means copper won't be phased out overnight, but instead becomes:
a very pragmatic intermediate technology in the transition from 400G and 800G to 1.6T optical interconnect.
3.3 PCIe Retimers: The "Signal Relay Stations" on the Server Motherboard
Inside the server, PCIe remains the basic backbone between CPUs, GPUs, NICs, and accelerator cards.
As PCIe Gen 6 and Gen 7 keep pushing data rates higher:
Trace lengths on a single motherboard become more sensitive
Bends, vias, and connectors all become signal integrity challenges
Marvell's Alaska P PCIe retimer plays this role:
Receiving the attenuated high-speed signal
Compensating and reshaping it, then passing it to the next link segment
Giving system designers more freedom in slot placement, card positions, and trace lengths
For Marvell, this is a very "invisible" product line, but one whose attach rate will keep rising.
4. Scale-Out and Scale-Across: 1.6T and Coherent DCI
Once you leave the rack, the network's lead actors become Ethernet and coherent optics.
4.1 1.6T Nova DSP: The Next Standard for AI Networking
Marvell reaffirmed its leadership in PAM4 DSPs at the event and introduced the 1.6T Nova platform:
Built on a 3nm process
200G per lane
Supports 200G/lane electrical and optical interfaces
On timing:
Already sampling to customers
Volume production expected between 2H 2025 and 2026
What this means for the industry:
As NVIDIA Blackwell, next-generation custom ASICs, and large-scale AI NICs adopt 800G/1.6T
optical modules, AOCs, LPO, and CPO products all have a high probability of carrying a Marvell DSP inside
4.2 COLORZ and 1.6T ZR/ZR+: Linking Data Centers Across Regions into "One AI Factory"
For longer-reach scenarios, such as interconnect between cities and between data centers,
Marvell's focus is its coherent optics product line:
COLORZ 800
1.6T ZR / ZR+
Key technical features include:
Coherent optics with higher-order modulation (such as 16QAM)
Transmission over hundreds to a thousand kilometers on standard fiber
While maintaining 800G- and 1.6T-class throughput
The practical value for hyperscalers:
Once a single data center maxes out its power, land, and cooling, operators can build multiple mid-sized facilities in different locations and use high-end DCI to "logically merge" them into a single AI factory.
In this area, Marvell has deep technical expertise and a first-mover advantage.
5. Custom Compute: A Platform, Not Just a Design Service
Marvell's custom ASIC business has been its fastest-growing segment in recent years and was a key emphasis at the event.
5.1 The Biggest Difference from Traditional ASIC Houses: A Platform-Based IP Portfolio
When people think of ASIC design, many picture "a design house turning your RTL into a manufacturable chip."
Marvell's model is closer to "providing an entire platform":
High-speed SerDes: 112G / 224G
Memory controllers: HBM3 / HBM4
Packaging technology: 2.5D / CoWoS-type, 3D stacking
Optical interconnect: Celestial AI's Photonic Fabric IP
For cloud customers, this means:
They can focus their effort on the most differentiated part — the compute core
while outsourcing highly complex but "non-core-differentiating" parts like I/O, interconnect, and packaging to Marvell
Development risk and time-to-market are both dramatically reduced
Put simply, what Marvell wants to do is:
let "cloud providers who want to build their own chips" stand on Marvell's platform and build upward, instead of building an entire system from scratch.
5.2 Growth Projections: The Power of Attach
Marvell laid out its expected CY23–CY28 CAGR by product segment, and the numbers are striking:
Product segment | Projected CAGR (CY23–CY28) | Notes |
Custom XPU Attach | 90% | Fastest growth, driven by rising attach rates for companion components such as DSPs, retimers, and optical chips |
Custom XPU | 47% | Reflects hyperscalers accelerating in-house silicon; Marvell benefits directly as design and platform supplier |
Interconnect | 35% | The 800G → 1.6T upgrade drives growing demand for DSPs and optical modules |
Switching | 17% | Despite a formidable rival in Broadcom, there are still new opportunities in AI-specific switches |
Storage | 9% | A mature market, sustained mainly by iterations of enterprise SSD controllers and related front-end ICs |
The key takeaway from this table:
The real explosion isn't the "single XPU" — it's "the whole string of interconnect and memory chips that ship alongside each XPU."
In other words, Marvell doesn't just earn from "one ASIC" — it earns from "the ring of components around every AI node."
5.3 Customers: Expanding Outward from the Top 4 Hyperscalers
Marvell said at the event that it has won custom orders from "the top four U.S. cloud service providers,"
and is working with other emerging hyperscalers.
Although the company didn't name them one by one, the industry context suggests:
AWS: the Trainium / Inferentia ecosystem
Google: TPU / Axion-type products
Microsoft: Maia / Cobalt
Meta: the MTIA series
All of these platforms are moving toward in-house ASICs and custom compute, and Marvell's role is:
amid the wave of compute self-sufficiency, to become the common supplier providing the "interconnect backbone and packaging know-how."
6. Memory and CXL: A Piece That Hasn't Taken Off Yet, but Sits in a Critical Position
In most AI discussions, memory gets reduced to "it's just HBM,"
but in reality, system-level memory architecture is undergoing structural change.
6.1 What Problem Does CXL Solve?
A pain point of traditional servers:
Memory is mostly tied to a single CPU / socket
If that CPU's workload isn't full, the memory attached to it is wasted too
In AI and cloud environments, this "stranded memory" drives up TCO significantly
CXL brings two important capabilities:
Creating a shared memory pool so multiple CPUs / XPUs can share one pool of memory resources
Creating multi-tier memory hierarchies, placing hot and cold data in different tiers to optimize cost and performance
6.2 Structera: Where Marvell's CXL Product Line Fits
Marvell is entering the CXL market under the Structera brand, focusing on:
Memory expansion controllers
Repackaging older DDR memory into a CXL-attached capacity tier
For cloud providers with large DDR inventories, this is a tool for "turning sunk costs into assets"
Near-memory compute
Adding lightweight processing such as compression/decompression to the memory controller
Reducing the bandwidth pressure and power of shuttling data back and forth between the host processor and memory
In terms of revenue share, CXL isn't a lead actor yet,
but in mid- to long-term AI infrastructure architecture, it could well become a key piece of the puzzle.
7. The Competitive Landscape: Broadcom, NVIDIA, and Marvell's Three Kingdoms
7.1 Broadcom: The Most Direct Rival
Competition between Broadcom and Marvell spans almost everything:
ASIC / Custom Compute
Switch ASIC
SerDes / DSP
Some storage-related ICs
The biggest difference is strategy:
Broadcom leans toward "heavy vertical integration," with much of its IP and design tightly bound to specific large customers
Marvell emphasizes "platform-based IP + custom flexibility," aiming to be a long-term partner to more cloud providers
On optics, their paths also differ:
Broadcom mostly partners with optical module suppliers while controlling the ASIC and some electro-optical front-end capabilities itself
Marvell has brought Celestial AI in-house, pushing deep into in-package optical interconnect
In short:
Broadcom has the advantage in "today's economies of scale," while Marvell hopes to stake out a position in "the next generation of packaging and interconnect architecture."
7.2 NVIDIA: A World-Class Customer, and a Potential Competitor
NVIDIA's strategy is to own compute + network end to end:
Its own GPUs and DPUs
Its own InfiniBand and NVLink
Its own fully defined software and communications stack
This naturally erodes Marvell's room in switches and some networking chips.
On the other hand:
Marvell remains one of NVIDIA's key DSP suppliers
Through its dual NVLink Fusion and UALink strategy, Marvell works with both NVIDIA and its competitors
The relationship is delicate, but it also highlights that:
at the AI infrastructure level, pure "friend or foe" relationships are increasingly rare — it's more about competing and cooperating at the same time across different technology layers.
8. Wall Street Views and Risk Review
8.1 Mainstream Analyst Views
Overall, analysts view Marvell's strategic direction fairly positively:
Rosenblatt: Maintained a Buy rating, arguing the Celestial AI acquisition cements Marvell's long-term lead in optical interconnect, especially in CPO and in-package interconnect.
Piper Sandler: Raised its price target to US$135, emphasizing Marvell's high revenue visibility into 2027–2028.
J.P. Morgan: Believes the market underestimates Marvell's penetration of the custom XPU market, particularly its work with the top U.S. cloud providers.
8.2 Risks That Can't Be Ignored
Of course, this path carries clear risks:
Execution risk
Celestial AI is still at an early stage of technology adoption, with significant engineering challenges before volume production.
Optical component yield, 3D packaging process maturity, and supply chain coordination are all variables that could delay products.
Technology direction risk
If process nodes and SerDes designs over the next few generations really extend copper's life by another one or two generations,
optical interconnect penetration may be slower than expected.
This would affect the pace and scale of returns on the Celestial AI investment.
Customer concentration risk
Marvell's rapid growth depends heavily on a handful of very large customers.
If major customers like AWS, Google, or Microsoft adjust CapEx or change their architecture strategy, Marvell's revenue will be hit in tandem.
9. Conclusion: Marvell Is Tying Itself to the Rising Curve of AI Infrastructure
Looking back at Marvell through this 2025 Industry Analyst Day:
It is no longer a networking IC company that "does a little of everything."
It is explicitly tying the company's future success to the central axis of AI data center infrastructure.
The core thesis can be distilled into one sentence:
In the AI era, compute is no longer the only bottleneck — what truly sets the system's ceiling is "connectivity" and "memory architecture."
Within this broad direction, Marvell has made three key moves:
Scale-Up:
Betting on the next generation of in-package optical interconnect through Celestial AI's Photonic Fabric.
Balancing NVLink partnership with the open UALink standard to avoid picking the wrong side.
Scale-Out / Scale-Across:
Using 1.6T DSPs and 800G / 1.6T coherent DCI to position for AI network upgrades and cross-data-center training demand.
Custom Compute:
Positioning itself as the "system-level platform supplier" to hyperscalers,
using high-attach-rate interconnect and memory products to amplify the revenue contribution of every XPU.
10. Key Metrics to Watch Over the Next Two to Three Years
Whether you're an investor, an industry practitioner, or a supply chain partner, the following metrics are worth tracking:
Celestial AI product validation and adoption pace
Whether concrete silicon proof points and demos are announced
Who the first customers are, at what scale, and on which XPU or AI platform
Ramp speed of the 1.6T product line
When clear adoption appears in mainstream 51.2T and 102.4T switches and AI NICs
Share shifts versus competitors (especially Broadcom's in-house DSPs)
Changes in the revenue mix of Custom XPU and Custom Attach
Whether the major hyperscalers keep increasing reliance on Marvell's platform for in-house silicon
Whether attach products (DSPs, retimers, optical chips) really grow at close to the 90% CAGR the company expects
Real-world adoption of CXL and Structera
Whether a major cloud customer publicly adopts CXL-based memory pooling with Structera as a key component
Overall, Marvell stands at a very critical crossroads:
If Celestial AI's technology lands on schedule and 1.6T and Custom Compute grow as expected, it will go from "an important supplier" to "a key player defining the next generation of AI infrastructure."
Comments