GlobalFoundries announced the agreement on September 17, and The Register first highlighted its relevance to the push toward near-packaged and co-packaged optics. Reuters independently reported that the arrangement increases SiGe capacity at the Vermont facility for Marvell’s data-center connectivity products. The key takeaway for infrastructure buyers is straightforward: a company whose growth thesis increasingly depends on optical interconnects has moved to lock down a less visible but necessary part of the supply chain before demand reaches the next speed step.
The companies have not disclosed the value of the contract, the wafer volumes allocated to Marvell, the share of Burlington’s existing capacity involved, or when the incremental production will become available. That omission makes this an important supply commitment, but not evidence yet of a factory expansion or of immediate relief for every optics customer.
Burlington capacity is being directed at a specialty process
GlobalFoundries’ Burlington operation is an established 200 mm specialty fab, not one of the giant leading-edge logic plants that manufacture the newest CPUs and AI accelerators. GF says the site has operated since 1957, was acquired from IBM Microelectronics in 2015, and has total capacity of 600,000 200 mm-equivalent wafers annually across several technologies. It is also a major Vermont employer, with about 1,800 employees and contractors.
That background matters because SiGe is a different manufacturing proposition from cutting-edge GPU logic. Silicon germanium processes are valued for fast analog and mixed-signal circuits: the electrical driver amplifiers, receiver electronics, clocking, and signal-conditioning functions that sit beside photonic components in a high-speed optical module. The material does not itself turn electricity into light; it helps the surrounding electronics operate cleanly at frequencies where ordinary interconnect design becomes increasingly difficult.
GlobalFoundries says its current SiGe technology supports 200 Gbps per lane optical connectivity. In practical terms, eight lanes at 200 Gbps can support a 1.6 Tbps optical module. The company says it has made advanced SiGe for more than a decade and is combining it with silicon photonics and advanced packaging work intended for tighter optical integration.
The new agreement builds on that existing production relationship. This is therefore capacity expansion within an existing manufacturing relationship, rather than a newly announced fabrication plant or a guarantee that a new SiGe process has reached mass production. GlobalFoundries described the added capacity only as “significant,” leaving customers and investors without a measurable unit of output or a production-ramp date.
Why 200G lanes are changing the networking bill of materials
For years, copper remained the economical answer for short links inside a server rack. Direct-attach copper and active electrical cables avoided the cost, optical packaging complexity, and power draw associated with pluggable optics. That balance gets harder to maintain as each electrical lane moves from 25 Gbps to 50, 100, and now 200 Gbps.
At 200 Gbps per lane, signal loss and heat become central design problems. NVIDIA’s current LinkX specifications show copper options for some 1.6 Tbps Ethernet deployments reaching roughly 2.5 to 3 meters, while single-mode optical transceivers can reach hundreds of meters or several kilometers depending on the module. The important distinction is not that copper disappears; it remains useful at short distance. It is that the distance over which copper is practical stops growing with the bandwidth demands of AI clusters.
This is where optics moves deeper into the data center. A conventional pluggable transceiver sits at the front of a switch and connects over electrical traces to the switch ASIC. Near-packaged optics, or NPO, moves the optical engines much closer to the switching silicon. Co-packaged optics, or CPO, puts optical components alongside or within the same package as the ASIC. Both approaches shorten electrical paths that become problematic at very high speeds.
The tradeoff is operational. Pluggable transceivers can be swapped from a switch faceplate, a familiar procedure for network teams. Co-packaged designs can reduce power consumption and eliminate some high-speed electrical signal-conditioning components, but place more valuable optical hardware closer to the switch package. NVIDIA promotes its CPO platforms as reducing latency and power consumption, while also claiming reliability gains. Those are vendor performance claims, and the practical service model will depend on how system vendors implement repairable optical subassemblies.
For enterprise administrators, this is not a reason to replace installed Ethernet or InfiniBand equipment. It is a signal that the cost, availability, and support model for optics will carry more weight in the procurement of high-density AI systems, especially where clusters span multiple racks and use 800 Gbps or 1.6 Tbps-class ports.
Marvell is protecting a growing optical business
The manufacturing agreement is more meaningful because Marvell’s dependence on data-center connectivity has become unusually large. In its 2026 proxy filing, Marvell said data-center products accounted for roughly three-quarters of its fiscal 2026 revenue. It also said optical interconnect had grown at an approximately 50 percent compound annual rate for five straight years and represented about half of its data-center revenue.
That makes the Burlington agreement a supply-side response to a business Marvell has already identified as central to its financial performance. Marvell’s first-quarter fiscal 2027 filing also attributed data-center growth to AI-related demand across electro-optics, custom silicon, storage, and switching. The company completed its $3.5 billion acquisition of Celestial AI in February 2026, adding technology for optical scale-up interconnects, and separately expanded its switching position through the acquisition of XConn.
The timing also follows NVIDIA CEO Jensen Huang’s widely reported June 2, 2026, comment at Computex that Marvell could be “the next trillion-dollar company.” The remark drove a sharp share-price reaction, but it was an endorsement of Marvell’s place in future AI infrastructure, not proof of a valuation outcome. The more concrete development is Marvell’s effort to secure the components required to deliver that connectivity at volume.
NVIDIA has its own reasons to care about optics supply. Its Quantum-X InfiniBand Photonics and Spectrum-X Ethernet Photonics platforms use co-packaged optics and 200G SerDes technology. NVIDIA says its Spectrum-X Ethernet Photonics switches, due in the second half of 2026, integrate CPO directly with the ASIC and provide up to 409.6 Tbps of switching bandwidth. NVIDIA’s 2025 launch materials also named a broad set of optics and manufacturing partners, underscoring that no single chip company or fab controls the whole production chain.
Marvell and GlobalFoundries are not claiming that this agreement supplies NVIDIA hardware directly. Neither company identified end customers, module makers, laser suppliers, packaging partners, or deployment schedules. Treating the Burlington deal as a dedicated NVIDIA supply contract would go beyond the public record.
The supply-chain bottleneck is broader than SiGe
The announcement points to a real constraint, but it does not solve the entire optical supply problem. A high-speed optical link relies on photonic integrated circuits, lasers, modulators, photodetectors, fiber attach, packaging, test, digital signal processors in many designs, and the analog electronics that SiGe can provide. A shortage or yield problem in any one of those areas can cap shipments.
GlobalFoundries’ own technical material makes this clear. It presents SiGe as one piece of a co-integration strategy that includes silicon photonics, advanced packaging, and eventually materials beyond silicon for speeds at or above 400 Gbps per wavelength. Its 200G-per-wavelength position is relevant to 1.6 Tbps modules today, but 3.2 Tbps-class optics will require both higher per-lane performance and a mature manufacturing chain across several component types.
The company’s roadmap language should also be separated from shipping product commitments. GlobalFoundries says it is exploring hybrid integration with materials such as thin-film lithium niobate, barium titanate, and electro-optic polymers to move beyond 200G-per-wavelength limits. Those are technology directions, not announced Marvell products, customer deployments, or guaranteed delivery dates.
For GlobalFoundries, the arrangement is a useful validation of Burlington’s role as a U.S.-based specialty manufacturing site. For Marvell, it reduces the risk that its optical ambitions are limited by its ability to obtain the high-performance analog silicon that its modules and optical systems require. For buyers of AI infrastructure, however, it changes no near-term compatibility list or product configuration.
The measurable milestones still missing are the ones that will show whether this agreement changes supply in practice: a disclosed production ramp, a wafer-capacity figure, qualified 200G-per-lane Marvell products, and systems shipping in volume with NPO or CPO designs. Until then, the September 17 agreement is best read as an early reservation of industrial capacity for a network transition that is already underway, rather than a completed expansion that has reached the data center floor.