Marvell fell more than 8% after hours on reports that it is in advanced talks to buy Lumentum, while Lumentum jumped about 10%. That is the market’s reflex for any large deal likely paid partly in stock. It prices the dilution on day one and leaves the strategy for later. The strategy here deserves a price now.
Every AI cluster that grows past a single rack runs on light. Marvell already owns the electrical half of that link. Its optical DSPs, the franchise it bought with Inphi for about $10 billion in 2021, sit inside most of the 800G and 1.6T transceivers shipping to hyperscalers today. What Marvell doesn’t own is the laser. Lumentum makes the indium phosphide lasers, EMLs and pump lasers the rest of the optical chain waits on, and in this cycle laser capacity has been the slow thing to add. A new DSP can be taped out in a year. A new InP fab takes much longer, which is why Lumentum just bought a fifth one in North Carolina.
Put the two together and Marvell sells the DSP, the laser, the switch and the photonic fabric, with Celestial AI and XConn already folded in. That is close to the full optical bill of materials from one vendor, arriving just as the industry moves GPU-to-GPU links from copper to optics. Co-packaged optics makes the fit tighter still. Once the optics move onto the package, the light source and the silicon have to be designed together, and a vendor that controls both gets designed in earlier and stays in longer.
The moat argument for Marvell has rested on attach for a while. A hyperscaler can pull a compute die in-house. It is far harder to in-source SerDes, optical DSPs and lasers, because each takes years of process and yield know-how that no customer wants to rebuild. Buying Lumentum moves more of Marvell’s revenue into the layer that survives in-sourcing. Management already said scale-up optics and switching drive most of the raise to roughly $18 billion of FY28 revenue, with custom silicon behind them. Lumentum makes that driver bigger and much harder to copy.
This isn’t a mature asset bought for cost cuts. Lumentum guided to roughly $1 billion of quarterly revenue at a 35% to 36% non-GAAP operating margin, after a year in which it guided growth above 85%. It carries a backlog of more than $400 million in optical circuit switches alone, a product that barely existed inside data centers two years ago. Added to Marvell’s $2.74 billion fiscal Q2, the combined company would be running near $3.8 billion a quarter, with the faster-growing piece being the one that is hardest to replace.
Nvidia sits on both sides already. It holds about 2.5% of Marvell through its $2 billion convertible preferred, partners with Marvell on NVLink Fusion and silicon photonics, and committed $2 billion to Lumentum in March. A combination puts two of Nvidia’s optical bets under one roof that Nvidia already works with, and gives the largest AI buyer one counterparty for both DSP and laser supply.
On the stock, Marvell was up 184% for the year at the end of August and peaked at $329.88 before the latest pullback, leaving it near $221 billion in market value going into the report. The 8% drop is the dilution being marked to market before a single term is known. Inphi drew the same kind of reaction in 2020, and it turned into the business behind Marvell’s rerating. If Lumentum’s growth rate carries into the combined optics line, a couple of years of earnings growth would absorb the share count, and Marvell would own the scarcest part of the AI interconnect outright.
The number to watch is in the announcement itself: the cash and stock mix, and whether management gives a combined optics revenue target alongside it.