COMPUTEX 2026 The sun is slowly but surely setting on copper interconnects, Marvell CEO Matt Murphy claimed in his Computex keynote this week. Within the next decade the IP house expects photons to take the place of electrons and change the way datacenters are built and run in the process.<\/p>\n
And, if Nvidia CEO Jensen Huang is right, the widespread transition to silicon photonics technologies could make Marvell the next trillion dollar company.<\/p>\n
With a market cap of $191 billion, Marvell still has a long way to go, not that stopped Wall Street investors from sending the company\u2019s share price on a 30 percent rally on the proclamation.\u00a0<\/p>\n
However, Huang\u2019s prediction, made during Marvell\u2019s Computex keynote this week, may be more than flattery. The large-scale deployment of AI infrastructure for training, inference, and agentic systems is already reshaping datacenter networks and pushing copper interconnects to the limit.<\/p>\n
\u201cThe distance a signal can travel over a copper cable is inversely proportional to the bandwidth, so every time you double the bandwidth, you have to cut the distance in half,\u201d\u00a0 Murphy explained.<\/p>\n
Today the fastest network interconnects operate at 200 Gbps per lane, but at these speeds copper cables can only carry a signal about 2.5 meters, effectively limiting interconnects.<\/p>\n
With the launch of its next-gen NVSwitch silicon in its Vera Rubin platform, Nvidia will double this again to 400 Gbps, halving copper’s reach once again. There\u2019s a reason the NVL72\u2019s switches are located in the middle of the rack.<\/p>\n
\u201cGoing forward, even the connections within the rack will become optical,\u201d Murphy said. \u201cThe whole industry knows this is coming. So, we’ve been preparing for this moment, not just Marvell, but the industry.\u201d<\/p>\n
Optics offer much greater reach, but the tech isn\u2019t without compromise. Pluggable optics are not only power hungry but they also fail.<\/p>\n
Power consumption is one of the reasons why Nvidia first revealed<\/a> its NVL72 rack systems, Huang explained that using optics would have added another 20 kilowatts to the system\u2019s then monstrous 120 kilowatt load.<\/p>\n \u201cYou use optics wherever you must, you use copper wherever you can,\u201d Huang said during Marvell\u2019s keynote.<\/p>\n While Huang expects copper interconnects to remain relevant for a while longer, Marvell is preparing for a future in which even PCB traces will be replaced by fiber optic cables.<\/p>\n In 2020 Marvell acquired<\/a> Inphi, which specialized in building\u00a0 optoelectrical interconnects,\u00a0 and more recently the company dropped billions to acquire<\/a> Celestial AI\u2019s silicon photonics interconnect tech. Then in March, Nvidia invested<\/a> $2 billion in Marvell to, among other things, advance its silicon photonics interconnect tech.<\/p>\n \u201cWe build optical modules that contain all the electronics needed to drive and modulate the laser and transmit data over long distances,\u201d Murphy said.<\/p>\n At copper’s end<\/p>\n \u201cThink about 10 years in the future and it’s a world where a lot of the copper connections are gone,\u201d Murphy said. \u201cThis is a world where then distance doesn’t matter… that’s a profound change.\u201d<\/p>\n All modern datacenter infrastructure and software has been designed around the constraints of distances.<\/p>\n \u201cWith optics, distance doesn’t matter. So now we can change the size of the scale up domain from 72 or 144 XPUs or GPUs to 1,000 or more, all optically interconnected,\u201d he said. \u201cThe implications for workloads are enormous.\u201d<\/p>\n But it\u2019s not just GPUs. Murphy explains that when everything from CPUs and GPUs to memory and storage are optically interconnected, they will no longer need to be in the same box.<\/p>\n \u201cModern AI servers are composed of a certain number of CPUs, XPUs, memory, and network interfaces, and the reason they’re all on the same system is because of distance,\u201c he explained. \u201cImagine a completely disaggregated architecture, XPUs in one system, memory in another, agentic CPUs in another.\u201d<\/p>\n This means these resources can be reconfigured on the fly to achieve the ideal ratio of CPU to GPUs to system memory for a specific workload.<\/p>\n Google is already doing this to a lesser extent with its TPU clouds. While the ratio of CPUs and memory to GPUs can\u2019t be reconfigured on the fly, the use of optical circuit switches means the number and shape (topology) of Google\u2019s TPUs can be adjusted to maximize inference or training performance.<\/p>\n This also has implications beyond AI. Even if the bubble collapses and AI infrastructure demand evaporates, one can imagine AWS and other major cloud providers using silicon photonics or co-packaged optics to disaggregate compute resources and then reassemble them a la carte.\u00a0<\/p>\n Battling Broadzilla<\/p>\n Marvell is a long way from a trillion dollar market cap and getting there assumes a certain other IP house doesn\u2019t eat their lunch.<\/p>\n Broadcom, whose market cap already surpasses $2 trillion, and whose customers include some of the largest hyperscalers in the world, including Google and Meta, has also been amassing a broad portfolio of silicon photonics and optics tech over the past several years.<\/p>\n These technologies include co-packaged optics for switches and XPUs, as well as DSPs for high bandwidth pluggables.<\/p>\n