Celestial AI Explained: The $3.25B Bet Marvell Made to Put Light Inside the Chip

• Marvell closed its acquisition of Celestial AI in February 2026 (about US$3.25B, all stock). What it bought is not "yet another CPO" but in-die optical I/O: the ability to deliver light to any point inside a GPU/ASIC die, not just its edge.

• The core product, Photonic Fabric, relies on OMIB (an optical take on EMIB) plus GeSi EAM modulators. The company claims 25x the bandwidth and 10x lower latency than conventional CPO, with a single PFLink chiplet reaching 14.4–16 Tbps.

• The real moat is the GeSi EAM process (IP inherited from the Rockley Photonics acquisition) and CoWoS-L integration, not the pretty architecture slides. But this is a 2028+ story for scale-up, and for now it is still a narrative.

  
  

1. The Pain Point: AI Chips Don't Lack Compute, They Can't Be Fed Fast Enough

Think of a modern AI accelerator as a giant factory. The problem stopped being machine speed long ago (FLOPs are cheap). The problem is that raw materials can't get in and finished goods can't get out. Moving data from HBM into the compute cores, and shipping results to the neighboring GPU, is the real bottleneck.

That bottleneck has two physical limits. The first is the beachfront limit: compute and memory capacity scale with die area, but copper I/O can only sit along the die's edge. As dies grow, the pins along the edge don't grow proportionally, so the bigger the chip, the hungrier it gets. The second is distance latency: copper trace delay is roughly length × 1 ns/mm, which is why HBM has to sit right next to the GPU.

Light shouldn't be forced to queue at the edge of the chip. It should walk straight into the middle of it.

2. Clearing Up a Misread: Marvell Didn't Buy "Another CPO"

The market's first reaction was "Marvell is shoring up AI optics," lumping the deal in with ordinary CPO (co-packaged optics). That's a misread. To see where Celestial AI sits, line up the three most-watched silicon photonics (SiPh) startups. They represent three entirely different integration dimensions: Ayar Labs paves a side road (optical engine beside the GPU, side-car), Celestial AI builds upward (optical engine buried beneath the compute die, in-die), and Lightmatter replaces the foundation (all chips planted on a full silicon photonic wafer, wafer-scale interposer).

One-line takeaway: Ayar gives the chip broadband, Celestial adds floors, Lightmatter swaps the foundation. Ayar is the most mature but bound by the chip edge; Lightmatter is the most radical; Celestial takes the middle path, the most pragmatic and also the hardest to execute: it breaks the edge limit while reusing existing packaging flows.

3. The Technical Core: Moving Light from Beside the Chip to Underneath It

Celestial's key move is to shift the data ports from the die's edge to its entire area. Conventional CPO can only route signals out from the die perimeter; Celestial's Vertical Area I/O lets light enter and exit vertically across the whole die footprint, so bandwidth scales quadratically alongside compute.

So how do you bury the light? The answer is OMIB (Optical Multi-Chip Interconnect Bridge). The most intuitive way to think about it: it's an optical version of Intel's EMIB. EMIB is a small silicon bridge embedded in the substrate that connects GPU and HBM inside the package; OMIB is likewise embedded in the substrate, but it links the compute die to an entire optical network, and it uses TSMC's CoWoS-L advanced packaging.

4. Breaking Down Photonic Fabric: Four Key Pieces

Celestial's product platform is called Photonic Fabric, a full-stack solution (photonics + mixed-signal ASIC + advanced packaging + software).

4.1 Three product forms: PFLink (a licensable optical-link chiplet/IP, 14.4–16 Tbps per chiplet, about ten times today's 1.6T port), PFSwitch (a low-latency, high-bandwidth scale-up optical switch), and OMIB (a package-level optical bridge).

4.2 Four layers of integration: Photonic Fabric uses light starting on-die: on-die (mm scale) → in-package (die-to-die) → package-to-package (via fiber array units, FAU, meter scale) → package-to-fabric. Conventional CPO only handles the outermost layer; Photonic Fabric replaces copper NVLink and PCIe links with optics end to end.

4.3 What OMIB actually looks like: it is essentially a large active silicon photonic IC (active PIC), built on SOI and embedded in the CoWoS-L substrate. The XPU sends signals vertically into the OMIB through micro-bumps on its underside, with the EIC (TSMC 4/5nm driver die containing SerDes/TIA/driver) stacked on top. Non-optical regions carry TSVs for power delivery, the laser sits in an external ELS for thermal isolation, and fiber exits through an FAU.

4.4 The freed-up edge goes back to memory: moving optical I/O to the center frees the die edge that optical interfaces used to occupy, so it can be reassigned to 4 DDR plus 2 HBM controllers. This is the physical basis of memory disaggregation: optical latency is only about 0.02 ns/mm (copper is 1 ns/mm), so HBM can move to a rack meters away and form a memory pool while the GPU accesses it as if it were next door.

5. The Moat: Why Only Celestial Dared to Bet on GeSi EAM

Skeptics will ask: the architecture is elegant, but can others copy it? The answer lies in a part most people overlook, the modulator, the switch that turns electrical signals into optical ones. The industry splits into three camps:

Approach	Key players	Size	Thermal stability	Positioning
MRM (micro-ring)	NVIDIA / Lightmatter / Ayar / Intel	~15µm	Poor (needs heaters)	Maximum density
MZM (interferometric)	Broadcom	>1000µm	High	Switch CPO leader
GeSi EAM	Marvell / Celestial AI	~50µm	High (wide temp, >85°C)	Short-reach, high-density interconnect

Celestial chose the third path, and that's precisely its hardest moat, for three reasons:

• IP that's hard to replicate: the key GeSi EAM IP came from the acquisition of Rockley Photonics, which spent years solving germanium-silicon lattice mismatch and dark current.

• A delicate process: making a Ge PD that only absorbs light is easy, but an EAM must operate right at the bandgap cliff between absorbing and transmitting, which demands extremely precise strain engineering and composition control.

• A double wall of patents and process: any competitor wanting to switch to GeSi EAM must clear both Rockley's patent portfolio and a from-scratch process development effort.

The EAM camp is betting that customers would rather have a slightly larger but rock-solid solution than babysit a heat-sensitive, fragile component next to the GPU.

6. Industry Links: Where Marvell Plugs It In, Who Benefits, and How Far Along It Is

6.1 Marvell's calculus: Marvell already had Inphi's DSPs, MZMs and EMLs (well suited to longer-reach switching). What it lacked was short-reach, high-density interconnect. With Celestial's in-die EAM, it can offer hyperscalers building custom ASICs a complete Optical CXL Appliance: the Structera CXL controller as the brain, UCIe as the interface, PFLink as the wings, with total cross-rack latency around 135 ns. The core strategy is to bypass NVIDIA-dominated CoWoS plus standard CPO and go straight at hyperscalers with in-house ASICs.

6.2 Upside scenario: Google's MPU: reports say Google is in talks with Marvell to co-develop an MPU (Memory Processing Unit) aimed at the data-movement bottleneck, which is exactly Photonic Fabric's home turf. Technical fit is high, but integrating it into a custom ASIC requires the XPU to place SerDes/UCIe at the die center aligned with the OMIB, and to manage the heat of an OMIB sandwiched between the heat source and the cooling path. The hard part is co-design and thermal management, not the optics themselves. (Information strength: the MPU collaboration is a media report and not finalized; integration feasibility is inferred and timing is speculative. Confidence: medium.)

6.3 Who in the supply chain rides along:

Segment	Key players	Why they benefit
SiPh foundry	GlobalFoundries / TSMC / UMC	Active interposer / PIC wafers; GF is underrated
CW / external lasers	Lumentum / Coherent / LandMark Optoelectronics / Luxnet	ELS thermal-isolation architecture needs high-power CW lasers
FAU / fiber arrays	Senko / Sumitomo / Browave / FOCI	Package-to-package links exit via fiber, driving FAU demand
Bumping / OSAT	Amkor / ASE / SPIL	Back-end chiplet integration and optoelectronic module assembly
Substrates / long reach	Unimicron; Lumentum / Coherent / Ciena	CoWoS-L organic interposer; cross-DC DCI and OCS

The bump angle deserves a note: OMIB's full-area vertical I/O moves micro-bumps from the edge to the entire die underside, so bump count per XPU rises significantly. Pitch keeps shrinking (35µm mainstream, down to 5µm with hybrid bonding), and non-optical regions need TSVs for power, which lifts content value for both OSATs and substrate makers.

6.4 How far along it is: standard CPO is the scale-out answer and already in volume production (Meta measured 800G CPO at 65% lower power than pluggables with 3x the MTBF). In-die / SiPh interposer approaches like Celestial's are the scale-up endgame, but no cloud production deployment has appeared yet. The two tracks will coexist through 2028, and revenue will come after 2028. (Information strength: production status is established fact; timing is reasonable inference. Confidence: medium.)

Conclusion

In three sentences: First, Marvell bought Celestial for the ability to put light inside the chip (in-die optical I/O), not for another CPO. Second, the real moat is not a pretty architecture diagram but the GeSi EAM process and Rockley's patents; anyone wanting to copy it has to start R&D from scratch. Third, this is a 2028+ scale-up story; chasing it now means chasing the narrative and supply-chain positioning, not this year's revenue.

So what should you watch? The first hyperscaler order that actually puts Photonic Fabric into a production ASIC, especially whether the Google MPU lands with Marvell. That moment is the signal that this story has moved from slides into the data center.

This article is a technology and industry trend analysis focused on Celestial AI's products and technology. It does not constitute investment advice.