The AI infrastructure buildout is creating a new bottleneck beyond GPUs: moving data between increasingly dense computing systems. Sivers Semiconductors is responding with a $30 million investment to expand its indium phosphide (InP) manufacturing facility in Glasgow, Scotland, targeting more than 100 million continuous-wave distributed-feedback (CW DFB) lasers in annual production capacity once the expansion becomes operational in late 2027.
AI data centers are becoming networking systems as much as computing systems. As accelerator clusters grow larger, the amount of data moving between processors, switches and storage systems is rising sharply, putting optical connectivity under increasing pressure.
That is the backdrop to Sivers Semiconductors’ decision to invest $30 million in expanding its InP manufacturing facility in Glasgow.
The Swedish semiconductor company announced the investment September 3, saying the expansion is designed to support expected customer production ramps tied to AI data centers and advanced optical networking. Construction is scheduled to begin in the second half of 2026, with the expanded facility expected to become operational in the fourth quarter of 2027.
Once complete, Sivers expects the site to produce more than 100 million CW DFB lasers annually. The project will also introduce additional manufacturing processes, automation and production flexibility.
The significance goes beyond Sivers’ own capacity.
Optical components have become an increasingly important part of the AI infrastructure supply chain because conventional electrical connections face growing challenges as bandwidth and distances increase. Fiber-optic links can move large volumes of data over longer distances with different power and signal-integrity characteristics than copper-based connections.
In AI clusters, those connections are becoming critical.
Training and inference workloads increasingly distribute computation across large numbers of accelerators. That creates substantial east-west traffic inside data centers, making the network fabric a potential constraint on overall system performance.
Sivers specializes in photonic semiconductor technologies, including InP lasers and semiconductor optical amplifiers. InP is particularly important in optical communications because it can be used to manufacture light sources and other photonic components operating at telecommunications wavelengths.
The company’s Glasgow investment therefore targets a component sitting relatively deep in the AI infrastructure supply chain: the laser technology that enables optical data transmission.
The timing reflects a broader surge in demand. LightCounting estimated that sales of optical transceivers and active optical cables reached roughly $10 billion in the first quarter of 2026, more than 90% above the same quarter a year earlier. Its research has also identified shortages of InP lasers as one factor that constrained optical-transceiver shipments during the earlier stages of the AI infrastructure expansion.
That supply-chain pressure is encouraging manufacturers to expand capacity before the next generation of AI clusters reaches full production.
Sivers is also changing how it approaches manufacturing.
The Glasgow investment forms part of the company’s transition from a Fab-Lite model toward what it calls a Hybrid Manufacturing strategy. Rather than moving all production in-house, Sivers plans to combine its expanded internal capabilities with external foundries, packaging providers and manufacturing partners, including partners in Asia.
That approach addresses a difficult problem in semiconductor manufacturing: capacity must increase without creating an inflexible cost structure.
For AI infrastructure suppliers, production flexibility can be particularly important because optical standards and product requirements are evolving rapidly. Data-center operators are moving from 400G and 800G connectivity toward 1.6T-class technologies, while co-packaged optics and other architectures could change where optical components sit within the networking stack.
NVIDIA, for example, is pushing silicon photonics and co-packaged optics as part of its strategy for scaling AI factories. Its Spectrum-X Ethernet Photonics platform integrates optics directly with switching silicon and is designed for extremely high-bandwidth AI networks.
NVIDIA has also deepened its relationship with optical suppliers. In March 2026, it announced a multiyear strategic agreement with Coherent that included a $2 billion NVIDIA investment and future access to optical component capacity for next-generation AI infrastructure.
Sivers’ strategy is different in scale and position, but the underlying market signal is similar: optical supply is becoming strategically important enough that capacity, manufacturing location and component availability are becoming part of AI infrastructure planning.
For cloud providers and hyperscalers, that creates a more complicated procurement equation. GPU availability remains critical, but the ability to connect those GPUs efficiently is increasingly inseparable from overall cluster economics.
Amazon, Google, Microsoft and Meta are all investing heavily in AI data-center infrastructure, while networking companies and optical component manufacturers are developing higher-speed architectures to keep pace.
LightCounting’s latest research estimates that Ethernet optical-transceiver sales could grow 73% in 2026, with the market potentially reaching $80 billion by 2031. The firm also notes that AI-cluster demand is one of the principal forces behind the acceleration.
The opportunity does come with risks.
AI infrastructure spending could slow, new optical architectures could reduce demand for particular laser designs, or accelerator supply constraints could limit data-center expansion. LightCounting has specifically noted that shortages in GPUs and other accelerators can ultimately affect demand for optical transceivers because networking deployments depend on the expansion of the underlying compute clusters.
That makes Sivers’ hybrid approach strategically relevant. Maintaining in-house expertise while accessing external manufacturing capacity gives the company a way to increase supply without betting entirely on one production footprint.
The Glasgow expansion ultimately illustrates a broader shift in AI infrastructure economics. The next phase of AI scaling will depend not only on more compute, but on the photonic technologies that allow that compute to communicate.
As AI clusters become larger and more distributed, the lasers hidden inside optical modules and networking systems are becoming an increasingly important part of the technology stack.
Market Landscape
AI data-center networking is moving toward higher bandwidth, lower power consumption and greater optical integration.
Traditional pluggable optical transceivers remain important, but the industry is also investing in linear-drive optics, co-packaged optics (CPO), silicon photonics and next-generation optical engines.
Sivers sits upstream in this ecosystem, supplying photonic semiconductor components rather than complete networking systems. Its InP lasers can ultimately become part of optical transceivers and other communications hardware deployed by data-center operators.
The competitive landscape includes major optical and photonics suppliers such as Coherent, Lumentum and Broadcom, alongside networking companies including NVIDIA, Cisco and Arista Networks.
The market is also becoming more strategically important for hyperscalers. NVIDIA’s investment in Coherent and its growing silicon-photonics ecosystem demonstrate how leading AI infrastructure companies are seeking greater control over optical supply.
LightCounting’s 2026 research points to exceptionally rapid optical-networking growth, with AI clusters emerging as one of the principal demand drivers.
Top Insights
- Sivers is investing $30 million in Glasgow to expand InP manufacturing capacity beyond 100 million CW DFB lasers annually.
- The expansion targets an increasingly important AI infrastructure bottleneck: optical connectivity between high-density accelerator and networking systems.
- Sivers’ hybrid manufacturing strategy combines internal production with external foundries and partners to improve scalability and supply-chain flexibility.
- NVIDIA and other infrastructure leaders are accelerating silicon photonics and co-packaged optics as AI clusters demand higher bandwidth and better power efficiency.
- Growing optical demand creates opportunities for photonics suppliers, although GPU availability and evolving networking architectures remain important market risks.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
