GlobalFoundries, Marvell Expand US Optical Chip Output
Update, Oct. 20: the Vermont deal is aging into Marvell's broader optics push
TheStreet reported Marvell is showing 2nm optical chips at ECOC 2026 in Málaga — while noting Broadcom's 3nm signal processor reached 400 gigabits per lane first in March. On Marvell's August 27 earnings call, executives said the amplifier-and-driver business the GlobalFoundries deal supplies is on or ahead of pace toward a $1 billion annual run rate. GlobalFoundries' Shankaran Janardhanan, SVP of photonics & RF businesses, framed the partnership's technical core: “SiGe is where GF's decades of high-performance radio frequency and analog leadership translate directly into breakthrough optical performance.” (DataCenterDynamics) The same capacity race is visible elsewhere — Winbond's $11.2B Infineon deal revived a different corner of the chip-supply map.
The AI boom has a bottleneck that doesn't show up in any earnings headline: the tiny optical links that carry data between chips inside a data center. On Thursday, two American chip companies moved to break it.
GlobalFoundries and Marvell announced a multi-year agreement under which GlobalFoundries will expand its silicon-germanium chip manufacturing capacity at its Burlington, Vermont facility to supply Marvell's optical connectivity products. The deal targets the pluggable optical transceivers, near-packaged optics and co-packaged optics that move staggering volumes of data between AI accelerators — the connective tissue of every large AI cluster.
Investors noticed immediately. GlobalFoundries shares jumped nearly 6 percent in premarket trading, while Marvell rose about 4.5 percent, according to Reuters. In a market that has spent the year bidding up anything attached to AI infrastructure, a domestic supply deal for the industry's most constrained component class was always going to move.
Photo: Sangitiana Fararano, CC BY-SA 3.0, via Wikimedia Commons
What the deal actually covers
The agreement is straightforward in structure: Marvell gets secured, expanded access to GlobalFoundries' silicon-germanium (SiGe) manufacturing capacity in Vermont; GlobalFoundries gets a committed, multi-year customer for one of its most strategically valuable technologies. Robb Johnson, Marvell's vice president of foundry technology, said the expanded collaboration "secures our access to the silicon germanium technology and manufacturing capacity needed to meet the growing demand for optical connectivity solutions."
The technology at the center of the deal is worth understanding. Silicon-germanium is the material platform on which high-speed optical transceivers are built — the lasers, modulators and detectors that convert electrical signals into light and back again. GlobalFoundries says it has manufactured advanced SiGe technology for more than a decade, and that its platform co-integrates SiGe with silicon photonics on the same die, a combination that enables faster, more power-efficient optical links. The current generation supports 200 gigabits per second per lane, with faster generations on the roadmap.
That matters because of where AI networking is headed. Inside today's AI clusters, data moves between GPUs over copper cables and pluggable optical transceivers — modules that plug into the front of a network switch. But as clusters scale toward hundreds of thousands of accelerators, the industry is migrating toward near-packaged and co-packaged optics, which move the optical engines right next to the switch silicon, slashing power consumption and latency. The GF–Marvell deal explicitly covers all three generations: pluggable, near-packaged and co-packaged. It is a bet that covers the present and the transition.
The optical bottleneck nobody talks about
For all the attention paid to GPUs, the AI build-out has an equally binding constraint in interconnect. A modern AI training cluster is not one giant computer; it is tens of thousands of accelerators that must exchange data constantly, and the speed of that exchange determines how efficiently the whole cluster trains. As model sizes have grown, the networking inside the data center has become as important as the compute — and optical links are the only technology that can move that much data without melting the power budget.
The physics behind the shift is unforgiving. Copper interconnects lose signal integrity rapidly as data rates climb, forcing the use of power-hungry retimers and equalizers — and at 200 gigabits per lane and beyond, the power and cost math simply stops working. Optical links sidestep the problem by moving data as light, but they demand precisely manufactured lasers, modulators and detectors — components built on SiGe processes that only a handful of foundries worldwide can run at volume. That scarcity is what makes a multi-year capacity reservation in Vermont strategically valuable rather than merely convenient.
This is why optical connectivity has become one of the hottest corners of the semiconductor industry. Just last week, Ayar Labs raised $650 million to commercialize optical I/O for AI chips. Marvell itself agreed last October to acquire Celestial AI for $3.25 billion, a bet on optical interconnect for next-generation systems. Marvell has also been shipping 1.6-terabit optical transceiver technology and co-developing 2-nanometer DSPs — the digital signal processors at the heart of optical modules — positioning itself as one of the two or three companies that matter in AI optical networking, alongside Broadcom and a handful of challengers.
The GF deal addresses the supply side of that equation. Optical components require specialized manufacturing — SiGe processes that few foundries offer at scale — and securing capacity in advance is how a company like Marvell makes sure it can actually ship what its customers order. In an industry where lead times are measured in quarters and data center build-outs are measured in gigawatts, a multi-year capacity reservation is as much a competitive weapon as a purchase order.
Vermont, of all places
There is a pleasing irony in the geography. The chips that will carry the world's most advanced AI traffic will be made in Burlington, Vermont — not Taiwan, not Arizona, not upstate New York. GlobalFoundries' Vermont facility is one of its longest-running US sites, and expanding SiGe production there is a concrete data point for the onshoring narrative that has dominated US chip policy since the CHIPS Act.
The domestic angle is not incidental. Optical interconnect is increasingly viewed in Washington as strategic infrastructure — the kind of capability the United States wants onshore. A supply chain that runs from a Vermont fab to Marvell's US-designed products to American data centers is, from a policy perspective, the ideal shape of the AI supply chain. It is also a commercial hedge: customers building hundred-megawatt AI campuses want to know their critical components are not hostage to a single overseas fab.
GlobalFoundries has been leaning into this positioning for years, marketing its US and European manufacturing footprint as an alternative to the Asia-concentrated supply chain. The Marvell deal is validation of that strategy from one of the industry's most demanding customers — a company that could, in principle, source SiGe capacity elsewhere, and chose Vermont.
What it means for the AI infrastructure race
Step back and the deal fits a pattern we have been tracking all year: the AI infrastructure race is moving from the obvious bottlenecks to the subtle ones. The first phase was about GPUs. The second was about power and memory — the dynamic we examined in the SK Hynix–Intel Ohio talks. The emerging third phase is about the plumbing: the optical links, the DSPs, the packaging technologies that determine whether a data center full of accelerators actually performs like one machine.
For Marvell, the deal de-risks its optical roadmap at exactly the moment its customers — the hyperscalers building the largest AI clusters on Earth — are placing their biggest-ever orders. For GlobalFoundries, it fills advanced capacity with a marquee customer and deepens its moat in a process technology where it has a genuine lead. For the industry, it is another signal that the optical transition — from pluggable modules toward co-packaged optics — is not a research project anymore. It is a capacity-planning exercise.
The market's verdict was swift: both stocks up sharply before the opening bell. In a year when investors have learned to look past the GPUs to everything around them, a Vermont fab making the chips that carry AI's data turned out to be exactly the kind of story Wall Street wanted to hear.
What is silicon-germanium (SiGe), and why does optical networking need it?
Silicon-germanium blends silicon with germanium to make high-frequency analog front ends — the components that drive and read lasers, modulators and detectors in optical links. Only a handful of foundries run SiGe at volume, which is why GlobalFoundries' Burlington capacity matters: it is the bottleneck behind the bottleneck. The current generation supports 200G per lane, and GlobalFoundries' roadmap points toward co-integrated SiGe and silicon photonics with advanced packaging aimed at higher-speed generations, including near-packaged and co-packaged optics. The constraint is not just physics but equipment — the equipment bottleneck behind the capacity race is what decides how fast any of this capacity actually arrives.
GlobalFoundries–Marvell optical deal FAQ
What did GlobalFoundries and Marvell agree to? A multi-year agreement under which GlobalFoundries expands SiGe capacity at Burlington, Vermont, for Marvell's optical products.
Which technologies does the deal cover? Pluggable optical transceivers, near-packaged optics and co-packaged optics; the current generation runs at 200G per lane.
Why Vermont? GlobalFoundries' long-running US fab, the CHIPS-Act onshoring narrative, and Washington treating optical interconnect as strategic infrastructure. The onshoring push extends beyond Vermont — a US–Japan push around GlobalFoundries' chip footprint and TSMC's planned second US hub are the same race from different angles.
What is the deal worth? No financial terms were disclosed. Marvell's amplifier-and-driver business was on or ahead of pace toward a $1 billion annual run rate per its August 27 earnings call (TheStreet).
Sources: GlobalFoundries press release (Sept. 17, 2026); Reuters (Sept. 17, 2026) for market moves; Ayar Labs funding and Marvell–Celestial AI deal per prior TechPulse USA coverage. Internal link: SK Hynix in talks with Intel on Ohio memory production.
Comments
Post a Comment