GlobalFoundries and Marvell expand SiGe deal for AI optical networking

GlobalFoundries (GF) and Marvell Technology have extended their long-running manufacturing partnership through an expanded multi-year agreement, directing additional silicon germanium (SiGe) production capacity at GF's Burlington, Vermont fab toward next-generation optical connectivity components. The deal is designed to support Marvell's growing demand for pluggable optical transceivers, Near-Packaged Optics (NPO) and Co-packaged Optics (CPO) used in high-density AI and cloud data centre deployments.

The two companies did not disclose the financial value of the agreement, but characterised it as a significant capacity addition building on more than a decade of existing production collaboration. Marvell is one of GF's strategic customers across photonics and radio-frequency applications, and the Burlington facility has been producing advanced SiGe technology for that relationship throughout that period.

The technology case

SiGe sits at the intersection of high-frequency radio-frequency engineering and photonics. GF's current process node supports 200G-per-lane optical connectivity, with a stated roadmap targeting higher speed generations as bandwidth requirements inside AI clusters continue to grow. The company also co-integrates SiGe with silicon photonics and advanced packaging to support emerging optical architectures such as NPO and CPO, both of which are seen as routes to reducing the power and latency penalty of moving data between GPU and custom-ASIC accelerators at close range.

Robb Johnson, Vice President of Foundry Technology at Marvell, said the industry is moving toward higher-bandwidth optical architectures and that the expanded GF collaboration "will help ensure we have the SiGe technology and manufacturing capacity to support the significant growth we see ahead." Shankaran Janardhanan, GF's Senior Vice President of Photonics and RF businesses, noted that GF's RF and analogue heritage is what translates into optical performance at these speeds, framing SiGe as core rather than peripheral to the company's roadmap.

Market context and competitive landscape

The race to connect AI accelerators at ever-higher bandwidth is reshaping the semiconductor supply chain. Optical interconnect demand is being driven by the scaling constraints of electrical copper traces inside server racks: beyond certain lane speeds and distances, photonic links become the only practical option for maintaining signal integrity and managing power budgets. This dynamic is giving SiGe-capable foundries a meaningful role in the AI infrastructure buildout that sits alongside, rather than in competition with, the better-publicised logic and memory supply chain.

GF's primary competitive differentiator here is its specialisation in compound-semiconductor processes at scale. Most leading-edge logic foundries, including TSMC and Samsung, focus overwhelmingly on advanced CMOS nodes; GF has instead concentrated on differentiated process technologies including SiGe, silicon photonics and RF-SOI. In the optical transceiver segment, GF competes with fabs offering indium phosphide and silicon photonics process lines, but the combination of SiGe with co-integration capabilities for CPO and NPO is a narrower field.

The US geography of the Burlington facility also carries policy relevance. The CHIPS and Science Act has directed substantial federal funding toward domestic semiconductor manufacturing, and compound-semiconductor and photonics capacity has been identified as a strategic priority alongside leading-edge logic. GF has been among the beneficiaries of that programme, and agreements that expand US-based SiGe output align with the broader industrial-policy goal of onshoring specialist process capacity. Marvell, which designs custom AI networking and accelerator silicon for hyperscaler clients, has strong commercial incentives to secure a reliable domestic supply base for components tied to its optical product lines.

Looking ahead, the trajectory of CPO adoption will be a critical test for the partnership. CPO integrates optical components directly into the switch or compute package, cutting transceiver module costs and power consumption substantially, but it requires tighter co-design between the optical chiplet and the host package than pluggable transceivers demand. Industry analysts expect CPO to move from niche deployments into broader data centre adoption across the latter half of this decade, making this capacity expansion timely if GF and Marvell can execute on the roadmap milestones they have signalled.