Marvell and Lumentum: Lasers Join the Semiconductor Stack

On Monday, Marvell announced an agreement to acquire Lumentum for $115B in stock. This is one of the largest technology acquisitions of all time. For those unfamiliar with Lumentum, Marvell, and the state of photonics in the AI data center, here's why.

• Compute interconnects in AI data centers are shifting from copper to optical networking. The physics of electrical and optical signaling means cluster size is now limited by network reach, and copper can no longer scale. Building larger clusters requires optics.
• Lasers are in massive long-term shortage. Without lasers generating photons, optical fiber is just inert glass.  Indium phosphide lasers are the most sought-after parts in AI infrastructure, and in October Lumentum told Bloomberg it was sold out through early 2029. Lasers have always been the hardest part of the optical signal chain and arguably the most difficult to manufacture at scale in the AI datacenter. This specialty manufacturing knowledge takes decades to develop, and only a few companies possess it.
• Optics is now part of the semiconductor value chain. Marvell already makes the chips at each end of an optical link (DSPs, drivers, and amplifiers), plus modules for links between data centers. Lumentum adds the lasers, the modulators, and the module transceivers that connect servers inside the data center. That’s vertical integration of the whole link, down to the laser, and nearly every optical link starts with a laser. With complete supply chain control, Marvell can now execute extreme co-design from the photon up, a capability almost no other semiconductor company (Broadcom also has in-house laser capability) can match. 

What comes next?

At Scintil, we believe the solution to the laser and III-V device manufacturing bottleneck isn't more of the same telco-era manufacturing techniques, but a fundamental paradigm shift from discrete photonics to integrated photonics. Electronics went through this transition in the 1970s, which led to the rise of integrated circuits. Photonics hasn't made this transition until now.

Three nested disruptions

1. Traditional compute to AI-based compute - the rise of LLMs and AI itself
2. Compute interconnect transition from copper to optical 
3. Optical transition from discrete components to integrated photonics 

• DWDM and Optical scale-up is where discrete lasers hit a wall. DWDM scale-up links reduce power per bit by running more channels at lower speeds as separate wavelengths on one fiber (dense wavelength-division multiplexing, or DWDM). A discrete laser's wavelength is set by its own grating, so it's hard to make one land on an exact wavelength. Suppliers make many and sort them by wavelength, a step called binning. Four is workable. At eight and 16, the wavelength spacing keeps halving, and the target keeps shrinking. An eight-wavelength bidirectional (BiDi) link needs 16 discrete lasers, each assembled and multiplexed separately. This scaling crisis will hit the industry next. With heterogeneous (InP + SiPHo) integration, the wavelengths are set collectively with precision semiconductor manufacturing equipment and can easily meet the most stringent requirements at scale.
• Optics needs massive economies of scale and a semiconductor mindset. Lasers already ship in relatively high volume, but they’re still made like a craft: grown, cleaved, and coated, then aligned and epoxied in place. As AI networks scale up, the number of lasers will climb much faster, and so will the demands on their performance and packaging. I have seen this movie before when MEMS devices replaced cleaved quartz crystals with silicon timing chips and mylar and magnet-based ECM microphones with silicon-based MEMS. Before Scintil, I ran Vesper, a MEMS microphone company which shipped close to 100M microphones and accelerometers. MEMS now ships more than 30 billion units a year, by Yole's count, and it got there by moving onto silicon wafers, where high volume is much easier to manufacture. Lasers need to become a design block on a chip with 16 or more lasers and other components. Each wafer costing a few thousand dollars at 300mm must contain many hundreds of these chips. This will completely redefine what is possible with optical networking and compute and allow the industry to scale to tens of billions of integrated lasers and other II-V devices (SOAs, high-speed modulators, etc.) per year. 
• Integration is the path. That's the bet behind SHIP™, our wafer-level photonic integration process. SHIP™ adds lasers to a standard silicon photonics flow: the indium phosphide gain material is die-to-wafer bonded only where a laser will be, and each laser's wavelength is set by a grating patterned in the silicon with deep-UV lithography. LEAF Light™, our single-chip DWDM laser source and the first product built on SHIP™, has demonstrated up to 16 lasers on a single chip, spaced as close as 100 gigahertz (GHz) apart with ±10 GHz accuracy. SHIP™ is validated on Tower Semiconductor's silicon photonics production lines.

Marvell and Lumentum are betting on vertical integration, from the chips down to the laser, and I think it's a smart bet, but the next-generation laser technology is coming fast. The industry’s next step is making the lasers as manufacturable as the chips beside them. After OCP, I wrote that the winning architecture is the one you can manufacture. This deal makes that case louder.

If you're planning a scale-up fabric or a channel plan, I'm glad to compare notes.

Matt Crowley is the CEO of SCINTIL Photonics.