Marvell's $191B bet: silicon photonics could remake datacenters

Marvell CEO Matt Murphy just put a very large idea on the table: copper interconnects, the wiring that has quietly held datacenters together for years, are heading toward the exit. At Computex 2026, he said the sun is “slowly but surely setting on copper interconnects” and argued that, within the next decade, photons will replace electrons in datacenter networking. That is not a small tuning change. It is a rewrite of how AI infrastructure gets built, how far signals can travel, and which companies get paid every time the industry pushes bandwidth higher.

The market noticed fast. Marvell has a market cap of $191 billion, which is a long way from the trillion dollar club, but Wall Street still sent the stock up 30 percent after the keynote proclamation. The real attention-getter, though, was Nvidia CEO Jensen Huang's call that a broad transition to silicon photonics could make Marvell the next trillion dollar company. That is the kind of endorsement that turns a technical roadmap into a boardroom conversation. It also makes sense in context: AI training, inference, and agentic systems are already stressing datacenter networks, and copper is running into a hard physical limit.

Murphy laid out the math in plain English. The distance a signal can travel over a copper cable is inversely proportional to bandwidth. Translation: every time bandwidth doubles, distance gets cut in half. Today, the fastest network interconnects operate at 200 Gbps per lane, and at those speeds copper can only carry a signal about 2.5 meters. That effectively constrains how far you can place components and still keep them talking at full speed. Nvidia's next-gen NVSwitch silicon in its Vera Rubin platform will double that again to 400 Gbps, which cuts copper's reach in half once more. That's why the NVL72's switches sit in the middle of the rack. The layout is not aesthetic. It is physics with a corporate budget.

Optics, by contrast, can reach much farther, which is why Murphy said, “Going forward, even the connections within the rack will become optical.” But the transition is not a free lunch. Pluggable optics are power hungry and they fail, which is a nasty combination when your datacenter already looks like a small power plant. Huang said optics would have added another 20 kilowatts to the NVL72 system's then-monstrous 120 kilowatt load, which explains the old rule of thumb he repeated during Marvell's keynote: “You use optics wherever you must, you use copper wherever you can.” In other words, the industry is not swapping one cable for another because it is fashionable. It is doing it because physics, heat, and power are closing in.

This is where Marvell's strategy gets interesting. The company has been preparing for this moment for years. In 2020 it bought Inphi, which specialized in optoelectrical interconnects. More recently, it spent billions to acquire Celestial AI's silicon photonics interconnect technology. And in March, Nvidia invested $2 billion in Marvell to help advance that same silicon photonics interconnect work. Murphy said Marvell builds optical modules that contain the electronics needed to drive and modulate the laser and transmit data over long distances. That is not just component supply. It is infrastructure leverage. If the data center of the future is stitched together by light, the companies that own the optical stack get positioned closer to the center of gravity.

Murphy thinks the end state is bigger than faster links. “Think about 10 years in the future and it's a world where a lot of the copper connections are gone,” he said. “This is a world where then distance doesn't matter... that's a profound change.” Why does that matter? Because the entire modern datacenter stack has been designed around physical limits, especially the need to keep CPUs, XPUs, memory, storage, and network interfaces close together. If optics removes that constraint, then the system architecture itself can change. Murphy described a future where the scale-up domain grows from 72 or 144 XPUs or GPUs to 1,000 or more, all optically interconnected. He also pointed to a fully disaggregated architecture, with XPUs in one system, memory in another, and agentic CPUs in yet another. That would let operators reconfigure resources on the fly to match a specific workload, rather than forcing every job into a fixed hardware box.

That idea is already showing up in smaller form at Google, where optical circuit switches allow TPU topology to be adjusted to maximize inference or training performance, even though the CPU and memory ratios cannot yet be reconfigured on the fly. And the strategic reach goes beyond the AI boom. Murphy said that even if the bubble collapses and AI infrastructure demand evaporates, cloud providers like AWS could still use silicon photonics or co-packaged optics to disaggregate compute and reassemble it a la carte. That matters for anyone building capital plans around cloud infrastructure, because the value is not just in training the next model. It is in reshaping the economics of how compute gets packaged, cooled, and sold.

Marvell is not alone in chasing that future, though, and that is the catch. Broadcom, with a market cap already above $2 trillion, has also been assembling a broad portfolio of silicon photonics and optics technologies over the past several years. Its lineup includes co-packaged optics for switches and XPUs, plus DSPs for high-bandwidth pluggables. CEO Hock Tan shares the same long view, saying late last year that he can see a point in time when photonics becomes the only way to do it, while also stressing that “we are not quite there yet.” He added, “The final, final, straw is when you can’t do it well in pluggable optics. Then you go to silicon photonics.” So the race is on, but the finish line is not tomorrow. For executives watching from adjacent sectors, the lesson is simple: the datacenter's next big margin pool may not come from the compute itself, but from the invisible network layer that makes compute possible in the first place.