Marvell announced its acquisition of XConn Technologies for $540 million on May 11, per industry coverage. XConn builds high-performance interconnect silicon — primarily PCIe switches and CXL fabric components used to wire GPUs and CPUs together at rack scale. Marvell described the deal as part of its broader push into AI and cloud connectivity.
What's actually new here is which problem the AI hardware industry is now solving. The first phase of AI infrastructure expansion was about GPU supply — get more chips, install more chips, train more models. That phase is mostly capacity-limited rather than performance-limited. The second phase, which Marvell-XConn is targeting, is about getting GPUs to talk to each other and to memory at speeds that don't bottleneck the chips themselves. Rack-scale inference needs CXL coherency; multi-rack training needs photonics-grade interconnect. The connectivity layer is where the next round of margins lives, which is why both Marvell and Nvidia (with its $40B in photonics-related strategic equity bets) are positioning here.
If you're a hardware buyer building inference infrastructure, expect the architectural conversation to shift from "how many GPUs" to "what's the interconnect topology" inside the next two product cycles. Plan procurement accordingly.