Marvell Bets Big on Optical (Again)

Marvell’s role in AI has been behind-the-scenes but substantial. It’s helping hyperscalers design ASICs, such as the Trainium3 launched by AWS this week. And its high-end networking and optical chips for datacenter connectivity are applicable to AI scenarios.

The acquisition of Celestial AI could boost Marvell on both fronts. Marvell announced yesterday that it’s buying the optical startup for roughly $3.25 billion in cash and stock, a deal expected to close in Q1 2026.

Founded in 2020, Celestial AI has raised $520 million, including a $255 million Series C that closed in August 2025.

Optical Ambitions

Marvell is not a large company in relative terms. It just reported third-quarter 2026 revenues of $2 billion (the fiscal year ends in January) compared with Broadcom’s Q3 2025 revenues of $16 billion. But under CEO Matt Murphy, who joined in 2016 and became chairman in 2023, Marvell has been determined to punch above its weight, especially in datacenter and AI circles.

That includes designing ASICs for the likes of AWS and Microsoft and a datacenter networking portfolio that includes high-end optical networking technologies, which originated with the $10 billion acquisition of Inphi in 2021.

Now, Marvell is ready for another step in optical, one that can boost the ASIC business as well. Celestial AI is one of a few companies investigating optical interconnection, alongside Ayar Labs and Lightmatter.

It’s about scaling up AI, as you might expect. Workloads can require connectivity between servers in different racks, so that the datacenter becomes less about individual servers and more about accelerator “domains” that span multiple racks.

This is partly related to the trend of OEMs like Dell and Supermicro producing rack-scale products. That has more to do with expediency—it’s a faster way to fill datacenters—but it’s another indication of the rack-level thinking that has overtaken AI datacenters. (This theme comes up in our recent report, AI, GPU Clouds, and Neoclouds in the Age of Inference.)

Celestial AI’s Photonic Fabric addresses rack-to-rack communications. It also provides optical connectivity for faster chip-to-chip communications within a processor package.

A central element of the Photonic Fabric is PFLink, a connectivity technology available as a chiplet or as licensed intellectual property to go into a chip design. PFLink is based on the UCIe protocol, which support speeds up to 14.4 Tb/s and distances up to 50 meters. It was designed for connecting heterogeneous chiplets within one package. Celestial AI is also developing a switch chip called PFSwitch for allowing multiple XPUs to collaborate as a single virtual “super XPU,” as the company puts it.

Marvell’s press release also notes that getting the network to operate at AI’s scale requires micro-refinements at the chip level. Here, it’s relevant that Celestial AI has developed a packaging technology called Optical Multi-Chip Interconnect Bridge (OMIB), which provides connectivity between disparate chips within one package.

All this technology will go toward putting optical interconnections on-chip and enabling more optical networking around the datacenter. These aren’t new ideas, but they’re continually battling the economics of copper. The long-term assumption is that AI will push beyond the limits of copper. Marvell and Celestial AI certainly buy into that theory, but it’s not yet a guarantee.

Visions of 2029

Celestial AI will take time to make its mark, as the terms of the deal imply. Celestial AI shareholders are eligible for a payout of an additional $2.25 billion in Marvell stock if certain milestones are met. They get one-third of that amount if Celestial AI earns cumulative revenue of $500 million by the end of fiscal 2029. They’ll get the full payout if the revenue figure exceeds $2 billion.

The 2029 range makes sense given the time required to design and validate complex semiconductors. AWS’s Trainium4 chip is likely to be announced next year, which means the design is likely completed (steps such as emulation would eat up the remaining time before launch)—meaning Celestial AI isn’t arriving soon enough to aid in that design. Celestial AI might not even arrive in time to help with Trainium5. It’s feasible that Marvell might have licensed Celestial AI intellectual property already, but aside from that possibility, any impact on Marvell’s bottom line is going to take years.

The acquisition accomplishes one more thing: It takes Celestial AI away from Broadcom. The two companies were reportedly working on prototypes back in 2023.