Marvell is acquiring XConn to strengthen its PCIe, CXL, and UALink switching capabilities, combining electrical and optical interconnect expertise to become a vendor-neutral platform provider for large-scale AI data centers.
In sum – what to know:
Strategic deal – Marvell buys XConn for $540m, adding high-speed PCIe and CXL switches and engineering talent to its UALink and AI interconnect portfolio.
Full stack – combining its CXL controllers with XConn switches enables large, coherent multi-rack AI systems, supporting high-fidelity accelerator comms.
Platform play – move promises long-term lock-in by deepening Marvell’s position in AI infra, helped also by its $3.25bn deal for optical-interconnect outfit Celestial AI.
Marvell Technology, making high-end chips to connect advanced processing units in AI infrastructure, is to buy XConn Technologies for $540 million (as widely reported). XConn supplies high-speed PCIe and CXL switching chips used to link processors, memory, and accelerators in large AI systems. The deal expands Marvell’s switching and interconnect portfolio, positioning it as a more complete provider for the latest – and largest – AI data centres.
Specifically, California-based XConn is a specialist in advanced PCIe and CXL switching silicon; the deal gives Marvell deeper expertise here, to go with its UALink portfolio, to connect and orchestrate large-scale AI systems – growing from single rack clusters with a few GPUs to multi-rack configurations with sometimes-hundreds of accelerators, all geared around latency, bandwidth, and coherence.
UALink is an open standard, built on “decades of PCIe ecosystem innovation”, designed to let many accelerators operate as a single coherent system, and deliver scalable low-latency connectivity in large AI systems. Marvell wants to be the dominant silicon supplier for UALink switches. The deal also strengthens its CXL capabilities. By combining its CXL controllers with XConn’s CXL switches, it can offer a full-stack CXL solution rather than just components.
Marvell’s controllers handle device-level memory access, while XConn’s switches provide high-speed routing to let multiple accelerators and memory modules communicate efficiently across servers or multi-rack clusters. XConn gives it a ready portfolio of PCIe 5 / CXL 2.0 switches, a near-term PCIe 6 / CXL 3.1 roadmap, and engineers who know how to build high port-count switches.
Marvell stated: “Together, Marvell and XConn bring together a significantly larger, integrated team to fully address the rapidly emerging opportunity in UALink switching as well as comprehensively support the growing list of customers and partners who want to work with Marvell in evolving their next generation AI platforms.”
The strategy for Marvell is to be an AI infrastructure company, combining its position in interconnect switches and relationships with hyperscaler companies, plus its pending $3.25 billion acquisition of Silicon Valley startup Celestial AI to add optical interconnect (photonic fabric) technology to its AI data‑centre portfolio. (Celestial AI has a ‘photonic fabric’ platform that uses light/optics instead of electrical signals to link AI computing chips, memory, and interconnects with high bandwidth, low latency, and improved energy efficiency.)
Together, its developing proposition covers electrical interconnect (PCIe, CXL, UALink), optical interconnect (rack-to-rack, board-to-board), and memory expansion and coherence; the backbone of future AI data centres, as such. The logic is that it is better to own layers in the AI stack than just to sell hardware – on the grounds you control all the standards-adjacent plumbing, which is critical to get AI right and also hard to replace.
Switch silicon has long design cycles and is deeply embedded; it creates strong customer lock-in. Marvell’s talk about a “scale-up fabric” looks like a sticky gamble with a huge payoff, potentially: if customers adopt UALink and CXL at scale, it might be the go-to supplier for this ‘plumbing’.
Hyperscalers are actively reducing dependency on single AI vendors, building heterogeneous accelerator fleets, and pushing open standards like UALink and CXL. Marvell claims to be vendor-neutral and standards-aligned, and interconnect silicon will always matter – no matter whose accelerators win. XConn has “more than 20 customers”; Marvell said its CXL and PCIe gear will contribute revenue by the second half of 2027, and hit $100 million by 2028.
Which is modest by hyperscale standards. But, again, strategic control of scale-up connectivity matters more; it is a longer-term platform bet. It is not about trendy GPUs or AI/LLM models, but about staking a claim on the future architecture of AI data centres, and the fabric for AI systems to scale beyond current limits. If AI becomes larger, more distributed, and more composable, then connectivity becomes the bottleneck – and the power position.
Matt Murphy, chairman and chief executive at Marvell, said: “This combination creates a compelling switching platform for accelerated infrastructure, advancing Marvell’s connectivity strategy for next-generation AI… We add proven PCIe and CXL switch products, IP, and engineering talent to expand our UALink switch team. Combined with our acquisition of Celestial AI, we will be well positioned to deliver the performance, flexibility, and architectural choice as AI systems grow in size and complexity.”
Gerry Fan, chief executive at XConn, said: “We have built the industry’s highest port-count advanced PCIe 5 and PCIe 6 switching portfolio to support the next generation of accelerated infrastructure. Marvell brings cutting-edge SerDes technology, a leading process roadmap, deep hyperscale customer relationships, and global scale. We share a common vision for high-speed connectivity as the foundation of modern data centers, and we look forward to working together to help customers drive new waves of AI innovation.”
