Marvell Bets $115 Billion on Lumentum to Own AI Optics

Marvell is acquiring photonics supplier Lumentum for $80 billion in stock, and with $35 billion in assumed debts and obligations the deal carries an enterprise value of $115 billion. The combination, announced with Marvell valued around $150 billion, is nearly twice the $67 billion Dell paid for VMware and EMC in October 2015, and well beyond the $61 billion Broadcom spent on VMware in May 2022.

The commercial logic is blunt: the order book for photonics lasers is sold out through early 2029, and laser supply threatens to gate Marvell's fast-growing custom silicon business. Marvell designs custom CPUs and XPUs for Amazon Web Services, Microsoft and Google, and vertical integration into indium phosphide lasers removes a dependency it cannot otherwise control.

What does the deal change in the laser supply chain?

Lumentum is one of only two key suppliers of the indium phosphide lasers that enable optical interconnects for scale-up networks across racks and potentially rows, plus scale-in links inside compute engines. The other is Coherent. Both received $2 billion earlier this year, in March, to expand manufacturing. A few weeks later, Marvell itself took a $2 billion infusion from Nvidia to help custom-chip customers adopt NVLink ports on their XPUs and possibly CPUs.

Nvidia, with a market capitalization near $5 trillion, could buy any of these companies with stock. It cannot buy both Lumentum and Coherent without inviting antitrust scrutiny, and buying either alone would secure only roughly half the laser supply, plus perhaps optical circuit switches, that future AI systems will need. Nor can it stand up its own laser operations quickly. Multiple suppliers pushing against each other should, in theory, hold down costs — though the HBM market shows how poorly that works in practice, with no supplier, Nvidia included, able to buy the HBM capacity it wants at any price.

How does Marvell justify the price?

Lumentum's recent numbers explain some of the confidence. In its most recent quarter it earned $360.3 million in profit on $1.26 billion in sales, making it more profitable than Marvell, which posted $308 million in net income. Two quarters earlier, Lumentum booked a $7.16 billion loss tied to converting $7.76 billion of debt into stock; operating income that quarter was $279.3 million on $1.01 billion in sales.

Chief executive Matt Murphy framed the deal as capturing a market inflection, and drew a direct comparison to Marvell's acquisition of PHY maker Inphi, which became the dominant driver of its AI-related revenue rather than custom chip work. "We are at the inception of the transition as the copper wall moves from scale-out to the next wave, which is going to be scale-up and then scale-in beyond that," Murphy told Wall Street analysts. "It's hard to see sometimes. If you look back at Inphi, when we acquired it, the company looked like we were paying a very high multiple for it. But you can see when you capture a market inflection at the right time with the best asset and the best team and you execute well, you can really outperform."

Murphy added that the scale of the Lumentum business over the next few years is about an order of magnitude larger than what Marvell expected from Inphi five years ago — and that Lumentum's orders are already on the books, not projected. Marvell also spent $3.25 billion in December 2025 to acquire Celestial AI for its photonics interposer technology, positioning it for optical on-package compute engines its custom-chip customers want de-risked.

What is the market opportunity?

Marvell puts its total addressable market at $400 billion in calendar 2030, driven largely by scale-out and scale-across chips for AI clusters. Lumentum brought a $90 billion TAM, and the combined company sees an additional $60 billion to chase, mostly through near-package optics and co-packaged optics chips and packaging. Supply chains, not demand, limit those TAMs today.

The physics argument is simple. Every layer down the network hierarchy — from distributed datacenters, to rackscale systems within a datacenter, to elements within a rack, to components on a package or tray — demands another 10X increase in bandwidth. Copper signaling cannot deliver it. The row is becoming the new rack, and the system board the new compute engine.

What are the risks?

Marvell claims the deal makes it the largest pure-play provider of datacenter connectivity chips, from compute engine to cross-region. That claim is generous. Nvidia and Broadcom will run datacenter connectivity businesses several times larger than the combined company. The deal also concentrates eggs in one basket, though combining expertise in indium phosphide, silicon germanium and gallium arsenide under one roof may reassure customers seeking volume leverage.

Optical circuit switches add an option. Google has used its Palomar OCS, based on reconfigurable micromirror arrays, since 2020 to lash together nearly 10,000 TPUv4 accelerators in reconfigurable topologies. If other hyperscalers follow, Marvell is positioned; if not, it still sells Innovium TeraLynx Ethernet switch ASICs for backend networks.

Closure is expected in the first half of 2028 — a very long time in AI terms. Chinese antitrust regulators will likely object, Europe may involve itself, counteroffers remain possible, and the deal puts Coherent immediately into play, much as Intel's Altera purchase a decade ago put Xilinx on the market.