Optical communications becomes a central AI infrastructure business

Corning’s second-quarter 2026 results put a less conspicuous layer of the artificial-intelligence buildout into sharper focus: the physical network connecting accelerators, switches and data halls. The company reported that Optical Communications sales rose 32% year on year to $2.07 billion in the quarter ended June 30, while Enterprise Networks sales increased 65%. Corning said sales of products aimed at generative-AI deployments expanded faster still.

Those figures matter because they point to a shift in the economics of data-centre construction. Much public attention is directed at GPUs, custom accelerators, servers and the electrical infrastructure that feeds them. Yet a modern AI cluster also requires a dense system of fibre, cable assemblies, connectors, transceivers and increasingly specialised photonic components to move data between computing resources. If that connectivity cannot be manufactured, installed and operated at the required scale, buying more processors does not automatically translate into a larger usable cluster.

Corning’s results do not establish an industry-wide supply shortage by themselves. They do, however, show that hyperscale customers are placing greater strategic value on optical connectivity and are willing to secure capacity over multiple years. This makes networking hardware a more material constraint, and a larger opportunity, in the race to build AI data centres.

Why AI clusters require more optical hardware

Traditional data-centre networks have long relied on a mixture of copper and optical links. AI changes the balance because training and inference workloads distribute computation across many accelerators. The performance of the overall system depends heavily on fast, reliable exchanges of data among those accelerators and the networking equipment that ties them together.

As clusters expand, short copper links become less practical in parts of the network due to reach, bandwidth density and power considerations. Optical links can carry high data rates over longer distances, but their deployment requires considerably more than fibre itself. Operators need precisely engineered cable routes, high-density connectors, fibre-management equipment, clean and repeatable installation practices, and components capable of surviving frequent moves, additions and changes in an operating data hall.

Corning has explicitly framed the trend as both “scale out” and “optical scale up.” Scale out refers to connecting growing numbers of compute nodes across a cluster. Optical scale up describes the use of optical technologies deeper within the system, including links that previously would have remained electrical and copper-based. The implication is significant: connectivity demand rises not merely because data centres are getting bigger, but because optics may occupy a greater share of the network architecture.

This is a hardware constraint with different characteristics from semiconductor fabrication. Fibre and connectivity products are manufactured across materials, drawing, cabling, connectorisation, assembly and testing processes. They must also be deployed in real buildings, where rack layouts, pathways, cooling equipment and maintenance access matter. Expanding capacity therefore means adding plants, machinery, trained labour and qualified supply chains, rather than simply allocating more chip foundry output.

Long-term agreements turn demand into capacity plans

Corning’s recent customer announcements make the financial results more consequential. In May, the company and NVIDIA announced a multiyear commercial and technology partnership intended to expand US manufacturing of advanced optical connectivity for AI infrastructure. Corning said it planned to increase US-based optical-connectivity manufacturing capacity tenfold and raise US fibre-production capacity by more than 50%, including three new facilities in North Carolina and Texas.

In its July 28 earnings release, Corning also said Amazon had agreed to a multiyear, multibillion-dollar supply arrangement for optical fibre, cable and connectivity solutions for its expanding US data-centre infrastructure. These arrangements suggest that major buyers are seeking more certainty around components that are easy to overlook when evaluating an AI project at the level of chips and servers.

For Corning, the agreements potentially reduce the risk associated with large manufacturing investments. Building capacity ahead of demand is capital intensive, especially when products require high yields and strict performance specifications. Customer commitments can provide a clearer basis for plant construction and equipment purchases. For customers, the benefit is greater confidence that the network layer will not delay the commissioning of expensive computing systems.

The structure of these deals also changes the competitive discussion. Suppliers are no longer selling only commodity cable into a broad telecommunications market. They are being asked to deliver integrated connectivity systems tailored to fast-changing AI architectures. That favours companies able to pair materials expertise with systems design, manufacturing scale and close engineering relationships with data-centre operators and equipment makers.

Product density is as important as fibre volume

The response to AI networking demand will not be limited to making more of the same cable. Corning has been adding products designed to increase density and simplify deployment, including technology licensed around PRIZM TMT ferrules and its wider GlassWorks AI portfolio. Its product roadmap also includes multicore fibre, which places multiple cores within a fibre while retaining a standard cladding footprint.

The strategic objective is to carry more capacity through constrained physical spaces. Data halls have finite tray capacity, limited room around switch and server racks, and practical limits on the weight and complexity of cable bundles. Higher-density fibre and connector systems can reduce the amount of cabling and the number of connection points required for a given network design. That can help deployment speed and serviceability, not only raw bandwidth.

Co-packaged optics is another relevant direction. In this approach, optical components are brought closer to switching or computing silicon, reducing the length of electrical paths. The technical promise is higher bandwidth density and potentially lower power use, but it also increases the importance of fibre management and precision optical assembly. A network becomes harder to regard as an interchangeable collection of cables when its optical interfaces are integral to the hardware architecture.

Strong growth, but execution remains the test

Corning’s quarterly numbers show an enterprise-networking business benefiting from an exceptional investment cycle. Optical Communications represented roughly 44% of reported core sales in the quarter, making it the company’s largest growth engine. Company-wide core sales rose 17% to $4.74 billion, while core operating margin expanded to 20.9%.

Still, the growth outlook depends on execution. The planned capacity additions must be built, staffed and qualified on schedule. Demand forecasts from a concentrated group of large technology customers can change with AI capital-spending plans, product transitions or network-design choices. New connectivity technologies must prove reliable at production scale, not simply demonstrate attractive specifications. Corning itself notes risks around customer demand, manufacturing expansion, competition, pricing and supply-chain conditions.

The more durable conclusion from the quarter is not that optical hardware has displaced semiconductors as the defining AI investment. It is that advanced computing increasingly requires both at scale. Accelerators create computational capability; optical networks make that capability usable across a cluster. Corning’s results offer concrete evidence that the network layer is moving from a supporting role to a core capacity-planning issue for AI data centres.

Sources