Marvell Paid $115 Billion for the Wire, Because Copper Stops at the Rack

A GPU rack is a cheap neighborhood. Inside it, accelerators talk over copper traces and short cables, and that works because the distance is short. Move the same conversation a few meters and the signal has to turn into light. Electrical lanes lose too much of themselves over distance at the rates these clusters run.

Marvell just paid for the company that makes that conversion. It paid as if the conversion is now the thing that decides how big a cluster can get.

The cheap zone ends at the rack boundary

Every AI cluster crosses three zones and only the first one is cheap. Inside a rack, copper carries the scale-up traffic that lets a group of GPUs behave like one accelerator. Between racks, the signal converts to optics at the port. That conversion costs power at both ends, adds a component that can fail, and lands on a networking team that already owns too much of the incident pager.

Which is why AI optical interconnect is now the line that decides cluster size. With copper, size is a cabling decision. With optics, size is a power and parts decision, and you buy it by the port. Lumentum supplies the lasers, the transceivers and the compound semiconductor manufacturing behind those ports. Marvell brings the electrical side, the switching and the custom accelerator work.

A fabric you can repatch turns cluster shape into a scheduling problem

The part of Lumentum worth the most attention is not the transceiver business. It is optical circuit switching, which moves light paths around the fabric instead of routing every packet through a switch. Connections between groups of accelerators can be rearranged without touching a cable.

If that holds at cluster scale, topology stops being a permanent artifact. It becomes a decision that has an owner at runtime. Placement, gang scheduling and bin-packing all assume a fixed shape today. They would need to be told the shape.

We covered the supply half of this in October, when NVIDIA started prepaying for optics and substrates that ship in 2028. The working assumption then was that optics were a component shortage to wait out. This deal says they are the architecture.

The next allocation fight is lasers and ports

Marvell is not buying a components supplier out of weakness. Revenue went from $3 billion in fiscal 2021 to $8.2 billion in fiscal 2026. It closed a $3.5 billion purchase of Celestial AI in February, and it signed a custom silicon deal with Google worth up to $120 billion in sales through fiscal 2033, with Google holding the right to buy a $12.2 billion stake.

It is buying because the optical layer has its own supply chain, running through laser fabrication and compound semiconductors, and that chain is narrower than demand. Lumentum competes with Coherent and with China’s Zhongji Innolight. A proposed US ban on new Chinese optical transceivers would push demand toward exactly the domestic capacity Marvell is now inside.

Broadcom sells the same stack under one roof, fiber optics included, which is why Marvell frames this as a race rather than a purchase.

The consequence for anyone planning a 2027 or 2028 cluster is a procurement one. Ask about ports per rack, not accelerators per rack. Ask which optics sit under allocation, whether the fabric can change shape without a maintenance window, and who is allowed to change it.

Three questions worth applying to your own cluster. Can the fabric change shape while jobs are running, and would the scheduler notice if it did? What share of 2028 interconnect spend is optics, and is any of it contracted? When a transceiver fails at 3am, does the workload route around it, or does the job restart from the last checkpoint?

Related reading. NVIDIA Shipped Vera Rubin With 13 Percent Less Bandwidth Than It Advertised covers what actually moved inside the rack this year. NVIDIA Is Prepaying for Capacity That Ships in 2028 covers the optics and substrate queue this deal now sits inside.