After reading this, the reader knows GlobalFoundries and Marvell are adding US chip capacity for faster optical links inside AI data centers.

GlobalFoundries Expands AI Optics Capacity

A multi-year agreement targets silicon-germanium components used in pluggable, near-packaged and co-packaged optical networking.

GlobalFoundries and Marvell have expanded a multi-year manufacturing agreement for silicon-germanium technology at GlobalFoundries’ Burlington, Vermont, facility. The companies say the added capacity will support optical connectivity for AI and cloud data centers.

The announcement does not specify wafer volumes, capital spending or financial terms. It says the expansion builds on existing production and is expected to add significant capacity for Marvell’s requirements.

Why it matters

AI systems are distributed machines. Thousands of accelerators exchange parameters, activations and stored data while training or serving a model. Faster processors provide little benefit if networking cannot feed them or coordinate work quickly enough.

Electrical links become harder to scale as distance and data rates rise. Optical systems can carry more information across a data center while managing signal loss and power differently. That is why the components behind transceivers and optical packaging are becoming strategic parts of the AI supply chain.

GlobalFoundries says its current silicon-germanium technology supports 200 gigabits per second per lane and has a roadmap for faster generations. Silicon germanium, usually abbreviated SiGe, combines silicon manufacturing with material properties useful for high-frequency analog and optical-interface circuits.

Marvell uses such technology in products for pluggable optical transceivers, near-packaged optics and co-packaged optics. Pluggable modules sit at equipment ports and can be replaced separately. Near-packaged and co-packaged designs move optical components closer to the switching silicon, reducing the distance high-speed electrical signals must travel.

Capacity is the immediate claim

The agreement is about manufacturing supply, not a new model or accelerator. It gives Marvell more planned access to a process used in the links around compute. GlobalFoundries gains a multi-year customer commitment at a US facility, while Marvell aims to reduce the risk that optical demand outruns component availability.

Neither company quantified the capacity increase. The phrase “significant capacity” is therefore directional rather than measurable from the release. There is also no disclosed timetable for each increment, customer allocation or revenue effect.

The announcement names current 200G-per-lane support, but an end-to-end optical link depends on more than the foundry process. Lasers, drivers, receivers, packaging, fiber, switching equipment and system software all affect speed, reliability and power. A process roadmap does not guarantee that every associated product is ready for deployment.

Even with those limits, the deal illustrates a change in AI infrastructure priorities. The first wave of attention centered on accelerator availability. As clusters grow, memory movement and interconnect efficiency increasingly determine how much useful work the installed chips can complete.

Domestic manufacturing is another part of the announcement. Burlington gives the companies an established US production site for a specialized technology. That can matter to customers seeking supply diversity, but the release does not claim that every material, tool or packaging step is US-based.

The commercial test will be whether added capacity arrives when customers need it and whether the resulting components meet cost, yield and power targets. Optical architectures are evolving quickly, and co-packaged designs require coordination across chipmakers, network vendors and data-center operators.

This is therefore a supply-chain commitment with technical significance. It does not remove networking bottlenecks by itself, but it adds manufacturing resources to a layer that determines how efficiently large AI systems use their processors.

Verification

  • Tier 1 — VERIFIED: GlobalFoundries and Marvell announced an expanded multi-year agreement on 17 September 2026 to increase SiGe capacity in Burlington, Vermont. Source: https://gf.gcs-web.com/news-releases/news-release-details/globalfoundries-and-marvell-expand-collaboration-next-generation
  • Tier 1 — VERIFIED: The release identifies pluggable transceivers, near-packaged optics, co-packaged optics and current 200G-per-lane technology. Same source.
  • Tier 1 — PARTIALLY VERIFIED: The companies call the capacity addition significant but disclose no volume, spending or financial terms. Same source.
  • Tier 2 — ANALYSIS: Explanations of bottlenecks, supply diversity and deployment risks are editorial analysis.

Glossary candidates

  • Silicon germanium: A semiconductor material system used for high-frequency and optical-interface circuits.
  • Co-packaged optics: Optical components placed close to switching silicon in the same package or assembly.

Cold-reader sentence: GlobalFoundries and Marvell are reserving more US manufacturing capacity for the optical links that connect large AI systems.