Marvell Technology just doubled down on a semiconductor technology most people have never heard of, and the logic behind the move explains something real about how the AI boom is reshaping American manufacturing.

On September 17, Marvell shares jumped more than 4% and GlobalFoundries climbed more than 6% after Reuters reported the two companies had expanded an existing agreement to increase production of silicon germanium, or SiGe, chips at GlobalFoundries' facility in Burlington, Vermont. The deal is multi-year, the capacity is being expanded now, and the timing is not an accident.

What SiGe actually does

Silicon germanium is a material that lets chips run faster while consuming less power. That combination makes it valuable in one specific, critical place: the optical connections that move data at high speed between servers inside AI data centers. When you ask a large language model a question, the answer travels through layers of networking hardware before it reaches you. SiGe chips are embedded in that hardware, in components like pluggable optical transceivers and co-packaged optics. As AI workloads grow, those connections have to carry more data, faster. SiGe is one of the few technologies that can keep up.

Robb Johnson, Marvell's Vice President of Foundry Technology, said the expanded deal will help ensure the company has the "SiGe technology and manufacturing capacity to support the significant growth" it expects ahead. That's corporate language for: we think demand is coming faster than we can currently supply it.

The factory in Vermont is already overwhelmed

GlobalFoundries' numbers tell the same story from the supply side. The company reported $1.786 billion in revenue for the second quarter of 2026, up 6% year over year. Its communications infrastructure and data center segment, which includes SiGe and silicon photonics, grew 62% compared to the same period last year. It was the seventh consecutive quarter of double-digit year-over-year growth in that segment, and the fastest quarterly expansion since 2022.

GlobalFoundries CEO Tim Breen said demand for SiGe "remains strong" and that the company is "oversubscribed throughout 2027." That phrase is worth pausing on. It means customers want more than the factory can currently produce, not just today but through the end of next year. The Burlington facility is essentially sold out before the expanded capacity even comes online.

Marvell, for its part, reported record revenue of $2.739 billion in its most recent quarter, up 37% year over year, driven largely by AI data center infrastructure.

Why this lands in Vermont, not overseas

The Burlington facility matters beyond the financial story. SiGe manufacturing is technically demanding and relatively rare. GlobalFoundries is one of the few foundries in the world producing it at scale, and it is doing so inside the United States. At a moment when Washington is actively trying to bring advanced semiconductor manufacturing back onshore, this expansion represents exactly the kind of domestic capacity policymakers say they want. The CHIPS Act directed tens of billions toward that goal. A multi-year deal between two publicly traded companies to expand a Vermont chip factory is how that policy intent actually shows up in the physical world.

The deeper pattern here is about bottlenecks. AI spending has been heavily covered as a story about Nvidia and data center construction. But the real constraint may be in the connective tissue: the optical networking hardware that links all those servers together. If data centers can't move data fast enough between chips, adding more chips doesn't help. Marvell and GlobalFoundries are betting, with real money and real manufacturing commitments, that the optical layer is the next chokepoint in the AI build-out. The Burlington facility is their hedge against being unable to deliver when the demand wave hits.