The problem between one chip and another
An artificial-intelligence data center is not a single electronic brain. It is a network of processors, memories and communications equipment that must constantly exchange enormous amounts of data. As available computing power grows, the route between components becomes more important. An accelerator may wait for instructions or results even when it is ready to work: the speed of a chip alone does not guarantee the speed of the system.
On September 17, GlobalFoundries and Marvell announced an expanded multiyear agreement to increase production capacity for silicon-germanium semiconductors at the Burlington, Vermont, facility. Their stated aim is to serve demand for optical connections in data centers. Reuters reported the agreement on the same day, identifying the link between the capacity expansion and high-speed connections. This is an announced industrial decision, not evidence that any particular data center has already become faster.
The decision matters because it makes a little-discussed part of the AI supply chain visible. Accelerators and models dominate the headlines; behind the scenes are the components that convert, amplify and move signals between machines. If those components are missing, a growing number of processors is not enough to build an efficient infrastructure. That is where a specialized manufacturing facility becomes a technology story as well as an industrial one.
What is made in Burlington
The material at the center of the agreement is silicon-germanium, often abbreviated as SiGe. GlobalFoundries says it has manufactured this technology for more than ten years and uses it in the high-frequency circuits needed for optical networking. The companies specify that the expansion covers components for pluggable optical transceivers, optical solutions placed near the chips, and architectures in which optics and electronics are brought into the same package. These are three different ways to shorten or strengthen the path of information, not three products that are already equivalent in the market.
A transceiver performs the crucial conversion between an electrical signal and a signal carried by light. Fiber covers distances that, within and between buildings, become increasingly difficult to span with copper alone at the required speeds. But fiber does not remove the need for electronics: circuits must drive and read the signals. SiGe technology sits precisely at this interface. Its importance does not come from a vague promise about photonics, but from the need to manufacture repeatable components in adequate volumes.
GlobalFoundries says its current technology supports optical connections at 200 gigabits per second per lane and that its development roadmap targets faster generations. The figure describes a technical capability stated by the company; it does not tell us how many systems have adopted it, at what cost, or what energy savings it delivers in a real facility. Nor does the announcement quantify the additional investment or the number of new units that will leave the factory. It would therefore be wrong to speak already of doubled production.
The 400-gigabit demonstration
In the same week, Marvell presented a series of demonstrations based on two-nanometer processes at the European Conference on Optical Communication in Málaga. They included a PAM4 optical connection at 400 gigabits per second per lane. Converge Digest explained the architectural significance of the test: moving from 200 to 400 gigabits per lane could pave the way for 3.2-terabit-per-second modules without increasing the lane count compared with a 1.6-terabit module made up of eight 200-gigabit lanes.
The conditional is essential. A demonstration at a trade-show booth is not a component that can be ordered, qualified by customers and installed in thousands of racks. Converge Digest also notes that Marvell has given no sampling or production timeline for that technology. Moreover, “two nanometers” is the name of a manufacturing generation, not a simple measurement of the width of an individual line visible in a chip photograph. The number indicates an engineering path, but on its own it is not enough to compare performance and power consumption.
Marvell also showed coherent links for different distances and an optical platform integrated close to switching equipment. The common thread is reducing friction as data moves. Links within a group of machines, between buildings and between data centers require different solutions. A faster cable alone cannot solve every section of the network, just as a more powerful engine cannot automatically fix every road in a city.
From a demonstration to real demand
The GlobalFoundries–Marvell agreement brings together two technology timelines. The first belongs to the factory: adding capacity in a plant that already produces components and trying to secure supplies over several years. The second belongs to design: testing circuits and architectures that could enter the next generation of networks. Confusing the two creates a false sense of imminence. The announced expansion may support current and future products; the more advanced demonstrations still have technical and commercial tests to pass.
What should we watch in the coming months to see whether this frontier becomes a business? First, figures on the capacity actually added in Burlington, which the announcement does not specify. Next, availability dates, delivered samples, customer qualifications and contracts that distinguish a promising project from a stable supply. Finally, comparable measurements of energy used per bit transferred and of link reliability as a whole, including conversion, cooling and maintenance. These are editorial assessment criteria, not results that the companies have already established in the cited documents.
Price matters, too. An optical connection can be technically superior yet remain limited to specific routes if module costs, integration or equipment replacement make the change uneconomic. Copper may plausibly continue to coexist with fiber where distance and density allow it. That is a hypothesis about how an operator might choose an architecture, not a market-share forecast. The value of the announcement is that it puts manufacturing capacity behind a technological demand that designers see as growing.
Finally, there is a lesson about the geography of AI. Knowing where accelerators are assembled is not enough: the system also depends on factories able to produce the analog circuits that communicate with light. Burlington is one link in that chain, not its only center. For a European reader, the interesting question is not who wins the race to the biggest numbers, but what skills and industrial capacity are needed to connect the next generation of machines credibly. The future of AI may depend less on one spectacular chip than on the quality of the ordinary, verifiable connections among many chips.
