The Great Shift in Optical Packaging (Part 4): After Marvell Bought Celestial AI, SiPh Interposers Are the Endgame for Scale-Up

3 hours ago

6 min read

In December 2025, Marvell announced its acquisition of Celestial AI. The market read the deal as Marvell bolstering its AI ASIC business, but that's only the surface. The deeper meaning is that Marvell has acquired the SiPh interposer, a key asset seen as the endgame architecture for scale-up CPO, putting it alongside independent player Lightmatter as one of two companies able to turn the PIC into an active interposer.

What is a SiPh interposer? Put simply: interposers used to be passive, silicon carrying RDL and TSVs purely for electrical routing. A SiPh interposer turns the silicon into an active component: light goes straight from the GPU die into the interposer's waveguides and is routed to another GPU or HBM without converting back to electrical. It pushes CPO logic one step further, from light reaching the side of the ASIC to light going straight into the middle of it.

This article breaks down the technical approaches of Marvell-Celestial AI and Lightmatter, why this path is the scale-up endgame, and what it means for the advanced packaging industry.

1. Why It's the Scale-Up Endgame: Copper's Physical Limits

Scale-up networks address ultra-high-bandwidth, low-latency interconnect between GPUs and between GPU and HBM. Today the mainstream relies on copper interfaces such as NVLink, Infinity Fabric and ICI, which can deliver 800 GB/s+ per link over very short distances, but at a cost:

High energy per bit: copper SerDes power can't be pushed down at these speeds, typically 1+ pJ/bit
Shoreline-limited: the GPU die perimeter is finite, which physically caps how many NVLinks you can attach
Reach-limited: signal integrity degrades quickly as copper reach grows, making it hard to scale to rack level

Putting optics next to the GPU (standard CPO) solves part of this, but it's still limited by where the EIC-PIC interface sits: light can only enter and exit at the ASIC edge, not cross the die interior. The SiPh interposer's answer is to use the entire die surface as optical I/O.

Lightmatter calls this 40x interconnect density, because waveguides are an order of magnitude denser than fiber.

In other words, the future architecture looks like this:

GPU dies and HBM dies sit directly on a SiPh interposer
Die-to-die communication runs through waveguides in the interposer, not copper
One interposer can connect dozens of chiplets across the entire package
Light exits the interposer edge through fiber arrays to the next package

From an advanced packaging perspective this is still 2.5D (chiplets side by side on an interposer), but from a photonics perspective it's 3D integration (the PIC becomes the interposer layer and performs active optical routing).

2. Celestial AI Photonic Fabric: What Marvell Bought

Celestial AI's core product is Photonic Fabric, a full-stack solution: photonics + mixed-signal ASIC + advanced packaging + software. It splits the SiPh interposer into three product forms:

PFLink: a chiplet or licensable IP providing die-to-die optical links
PFSwitch: a low-latency, high-bandwidth scale-up switch
OMIB (Optical Multi-Chip Interconnect Bridge): a package-level optical bridge

Technically, Photonic Fabric promises:

Within chip: data can reach any point on the die, with no beachfront limit (mm-scale distances)
Within package: die-to-die or die-to-chiplet through a photonic interposer (mm scale)
Package-to-package: connected through an FAU (fiber array unit), meter scale
Package-to-fabric system: also through FAUs, meter scale

Integrating all four layers on one platform is the key. Conventional CPO only handles the third layer (package-to-package); Photonic Fabric uses light starting on-die, eliminating NVLink and PCIe links end to end.

That's the point of the acquisition: Marvell already has an ASIC business (custom XPU design), SerDes IP and cloud customer relationships. With Photonic Fabric in-house, it can offer hyperscalers that build their own ASICs, such as Google, Amazon and Meta, a complete "ASIC + SiPh interposer" package that bypasses the NVIDIA-led standard CPO path.

Marvell-Celestial AI's real rival isn't TSMC or Broadcom; it's NVIDIA's command over cloud-grade GPUs.

3. Lightmatter Passage: The Independent Track

Lightmatter's counterpart product is Passage, in two versions:

Passage M1000: a 3D photonic interposer that connects chiplet processors with light

Passage L200: 3D CPO, a more aggressive integration

Technical details:

Diced from 300mm SOI wafers using the GlobalFoundries Fotonix 45CLO process
Uses an external multiband laser chip as the light source, avoiding the reliability issues of integrated lasers
Provides standard SerDes I/O interfaces to chiplets
Modulates electrical signals onto multiple laser wavelengths using microring resonators
Combines multiple wavelengths into one optical circuit with a multiplexer
Routes signals with interferometers and ring resonators
One Passage can connect 48 customer chips, with <2ns switching latency between any two

Passage and Photonic Fabric are fundamentally the same idea: the PIC as an active interposer. The differences:

Dimension	Lightmatter Passage	Marvell Photonic Fabric
Company type	Independent fabless	Part of Marvell
Wafer fab	GlobalFoundries	(Details undisclosed; possibly multiple fabs)
Packaging partner	Amkor	(Marvell's existing supply chain)
Business model	Sells products directly	IP licensing + integration into Marvell ASICs
Customer reach	Requires customer co-design	Through Marvell's ASIC customers

Lightmatter is working with Amkor on the industry's largest multi-reticle 3D photonic package, aiming to push the compute density a single package can hold to the limit. That competes head-on with TSMC's CoWoS path; Lightmatter simply replaces the silicon interposer with a SiPh interposer.

4. Why GlobalFoundries Is Key to This Path

Neither SiPh interposer company uses TSMC: Lightmatter uses GlobalFoundries, and Celestial AI also has strong GF ties. Why?

Reason one: GF Fotonix is a platform that integrates photonics and electronics on the same CMOS wafer. Electronics and optics run on the same wafer, a natural advantage for SiPh interposers, which need exactly that: logic and optics on the same piece of silicon.

Reason two: TSMC's CoWoS and COUPE focus on transceiver-style integration, not active interposers. TSMC follows EIC-PIC stacking logic, which doesn't fully overlap with the chiplet-on-PIC logic of SiPh interposers.

Reason three: capacity bottlenecks. CoWoS capacity is already tight. If SiPh interposer players had to queue at TSMC against NVIDIA, they'd never get a slot. GF offering SiPh plus active integration with capacity that can expand quickly is a rational choice for these fabless companies.

If SiPh interposers really do become mainstream for scale-up, GlobalFoundries will be a seriously underrated winner on this path.

Of course, TSMC won't ignore this market and will likely extend COUPE to active interposer applications in the coming years. But in the near-term 2026–2028 window, GF does have a first-mover advantage.

5. Impact on the Advanced Packaging Industry

If SiPh interposers do capture scale-up CPO, they'll rewrite the value structure of advanced packaging. Three concrete effects:

Impact one: silicon interposer value gets split. The value of a traditional silicon interposer (the one used in CoWoS) lies in RDL and TSVs: a high technical bar, but fundamentally a passive component. Adding optical routing makes each wafer worth a tier more, but suppliers expand from TSMC alone to TSMC + GF + possibly Samsung, so the value gets visibly split.
Impact two: fiber arrays and FAUs become far more important. SiPh interposers route internally with waveguides, but packages still connect to each other with fiber. A Passage-class SiPh interposer may need dozens of fibers to exit, a direct tailwind for fiber array makers such as Senko, Sumitomo Electric, Corning, Browave and TFC Communication.
Impact three: NVIDIA's leverage gets diluted. NVIDIA's standard CPO + CoWoS path is a vertically integrated architecture. The SiPh interposer path gives hyperscalers an option to build high-bandwidth GPU clusters without NVIDIA; Google TPU, Amazon Trainium, Meta MTIA and Microsoft Maia could all adopt it. If that ecosystem truly grows, NVIDIA's bargaining power in cloud-grade AI will decline.

6. Risks and Counterarguments

The narrative is attractive, but two practical constraints need to be spelled out:

Constraint one: technology maturity and ecosystem validation are still early. Lightmatter's Passage M1000 has been demoed, but no deployment in a hyperscaler production environment has emerged. Integrating Marvell and Celestial AI takes time, and first commercial shipments are expected around 2027–2028. Until then, all scale-up interconnect will keep running on NVLink and Infinity Fabric.

Constraint two: it competes with standard CPO rather than complementing it. NVIDIA and Broadcom are betting on standard CPO with 3D EIC-PIC stacking, which is already in volume production. SiPh interposers are a more radical long-term path, and in the near term hyperscalers won't bet on an unproven architecture, especially when standard CPO already delivers 30–50% power savings. Why take the risk?

That means the two paths will coexist through 2026–2027, with a winner emerging only in 2028+.

For readers: this axis is worth tracking, but don't bet on it short term. It's a 2030+ story, not a 2026 story.

7. Conclusion: The Real Endgame for Scale-Up CPO Has Emerged

Condensed into one sentence: standard CPO is the answer for scale-out, the SiPh interposer is the answer for scale-up, and these are two different axes.

Marvell buying Celestial AI and Lightmatter deepening its work with Amkor are two independent signals pointing the same way: the SiPh interposer path really will reach industrialization rather than stalling at the demo stage.

For Taiwan's supply chain, the biggest implication is that GlobalFoundries' SiPh wafer capacity will become a strategic resource. Taiwanese firms can still take back-end chiplet integration in advanced packaging (ASE and SPIL can play Amkor's role), but Taiwan currently has no player in the SiPh wafer segment. That's a natural point of supply-chain diversification, and a gap Taiwan's industrial policy should pay attention to.

For investors, this axis remains a narrative trade until 2027, and real revenue will wait until 2028+. But if you follow long-term themes, Marvell, Lightmatter (private), GlobalFoundries and Amkor are the four direct beneficiaries.