13 Nov 2026
How Does AI Networking Work, and Why Do Optics Matter?
AI networking is the equipment that connects thousands of accelerator chips so they can work together as one system: switches that route traffic, copper cables for very short links, and optical transceivers and fiber that carry data as light over longer distances. Because large models are trained across many chips at once, the speed of those connections can limit the performance of the whole cluster. As cluster sizes have grown, demand for high-speed optical components has risen sharply, and lasers and transceivers have become a supply constraint of their own.
Why do AI clusters need so much networking?
A single AI accelerator cannot hold or train a frontier model alone. Training splits the work across thousands of chips that constantly exchange results. Inference on large models also spreads across multiple chips. Every exchange crosses a network link, and if any link is slow, the chips on either end wait.
AI clusters therefore use two kinds of networks. A “scale-up” network connects chips within a server or rack at very high speed, often with copper. A “scale-out” network connects racks across a data hall or campus, usually with optics. Traditional cloud data centers moved most traffic between servers and users; AI data centers move far more traffic between servers, which multiplies the number of high-speed links per rack.
What are the main components?
Switches. Devices built around switching chips that route data between servers. Each generation doubles bandwidth per chip.
Network interface cards. Hardware in each server that connects it to the network.
Copper cables. Direct-attach and active electrical cables used for links of a few meters, where they are cheaper and use less power than optics.
Optical transceivers. Pluggable modules that convert electrical signals into light and back. Common speeds in AI clusters are 800 gigabits per second and 1.6 terabits per second.
Fiber and connectivity. Fiber cables, connectors and patch panels that carry light between transceivers.
Lasers and photonic chips. The light sources, modulators and detectors inside each transceiver.
Why do optics matter more as speeds rise?
Electrical signals degrade faster over copper as speeds increase, so the distance copper can cover shrinks with each generation. Beyond a few meters, optical links become the practical choice. As AI clusters grow and link speeds rise, the share of connections that must be optical increases.
The numbers reflect that shift. TrendForce estimated that 800-gigabit-and-faster modules would rise from 19.5% of global optical transceiver shipments in 2024 to more than 60% in 2026.1 LightCounting projected 120% growth in Ethernet optical transceiver sales in 2026.2 Yole Group forecast more than 10 million 1.6-terabit modules in 2026 and about 150 million 800-gigabit-and-faster transceivers across 2026 and 2027, while warning that laser supply had become a significant upstream bottleneck.3
How fast are link speeds increasing?
Network speeds in AI clusters have roughly doubled with each product generation. Links at 400 gigabits per second were common in cloud data centers a few years ago; 800 gigabits became the standard for AI clusters in 2024 and 2025, and 1.6-terabit modules began ramping in 2025 and 2026. Industry standards groups have defined 800-gigabit Ethernet and are completing work on 1.6 terabits, built on 200 gigabits per electrical lane.2 Each step requires new lasers, signal processors and switching chips, which refreshes demand across the supply chain and can reshuffle which suppliers win designs.
What is co-packaged optics?
Today most optics are pluggable modules that slot into the front of a switch. Co-packaged optics (CPO) moves the optical engine onto the same package as the switching chip, shortening the electrical path. The main benefit is power. A paper presented by Meta at a 2025 industry conference measured about 15 watts for one pluggable 800-gigabit transceiver, compared with about 5.4 watts per 800 gigabits for the optical engine in a co-packaged switch.4
CPO also carries trade-offs. If a laser fails in a co-packaged system, repair can be harder than replacing a pluggable module. Adoption has started in switches, and the pace of wider use remains uncertain. A shift toward CPO could change which suppliers capture value, moving some spending from module assemblers toward chip and packaging companies.
Where are the supply chain pressure points?
Lasers. High-speed lasers and the indium phosphide wafers used to make them are produced by a limited number of suppliers.
Digital signal processors. Chips inside transceivers clean up signals, although “linear” designs aim to remove them from some short links to save power.
Assembly and test. Much high-volume transceiver assembly takes place in China and Southeast Asia, which exposes the supply chain to trade policy and export controls.
Fiber and connectors. Large AI campuses require very high fiber counts, along with connectors and cabling hardware built to tight tolerances.
Our articles on the AI data center bill of materials and the semiconductor supply chain show how networking fits with compute, memory and power.
What does this mean for investors?
Networking and optics are a distinct part of AI infrastructure. Spending tends to grow with cluster size and link speed rather than with chip count alone. The segment includes switch makers, chip designers, optical component suppliers, cable makers and fiber producers. Risks include rapid technology transitions, customer concentration, pricing pressure as volumes rise and trade restrictions. Our guide on how to evaluate a thematic ETF explains how to review whether a fund includes this layer.
Frequently asked questions
What is an optical transceiver?
An optical transceiver is a small module that converts electrical data signals into pulses of light for transmission over fiber, and converts incoming light back into electrical signals. In AI data centers, transceivers plug into switches and servers and typically run at 800 gigabits or 1.6 terabits per second. Each contains lasers, photodetectors and, often, a signal-processing chip.
Why not use copper for everything?
Copper is cheaper and uses less power over very short distances, which is why it remains common inside racks. As data rates rise, however, signals over copper weaken more quickly, so the usable distance shrinks with each generation. For links between racks and across a data hall, optical fiber carries data farther with less signal loss.
What is the difference between scale-up and scale-out networks?
A scale-up network links accelerators within a server or rack so tightly that they can act like one large processor, often using proprietary high-speed connections over copper. A scale-out network links many racks across a data center, usually with Ethernet or similar protocols over optics. AI clusters need both, and each has its own suppliers.
What is co-packaged optics?
Co-packaged optics places the optical components on the same package as a switch chip rather than in separate pluggable modules. Shortening the electrical connection reduces power use per bit. The approach can make repairs harder and requires new manufacturing processes. Industry adoption began in switches in 2025 and 2026, and how widely it spreads is still being determined.
Why have lasers become a bottleneck?
High-speed transceivers need specialized lasers built on compound semiconductor wafers, and only a limited number of suppliers produce them at the required quality. Demand for 800-gigabit and 1.6-terabit modules rose faster than laser capacity. Yole Group identified laser light sources as a significant upstream constraint in its 2026 analysis.
How do thematic ETFs include networking?
AI and semiconductor ETFs may hold switch makers, networking chip designers, optical component suppliers and fiber producers. Some data center or digital infrastructure funds include cable and connectivity companies as well. Exposure levels vary widely by methodology. Reviewing a fund’s holdings shows whether networking is a meaningful part of its portfolio or a small allocation.
- TrendForce, “Google’s High-Speed Interconnect Architecture to Push 800G+ Optical Transceiver Share Past 60% by 2026,” February 10, 2026.
- LightCounting, quarterly market update, as reported by IEEE ComSoc Technology Blog, October 2, 2026.
- Yole Group, “Optical Transceivers at the Chokepoint of AI Growth and Supply Chain Constraints,” 2026.
- Meta, paper presented at the European Conference on Optical Communication (ECOC), 2025, as summarized by 100gmodules.com.
Cite this page: Adam Patti, "How Does AI Networking Work, and Why Do Optics Matter?," VistaShares, updated November 13, 2026, https://www.vistashares.com/ai-networking-optics-explained/. Plain-text version.
This material is provided for educational purposes only. It is not investment, legal or tax advice, and it is not a recommendation or an offer to buy or sell any security. Information is drawn from sources believed to be reliable as of the date of publication, is not guaranteed as to accuracy or completeness, and may change without notice. Figures attributed to third parties are their estimates. Projections and forecasts are not guarantees of future results. Investors should consult their own financial, legal and tax professionals before making investment decisions. Investments involve risk, including possible loss of principal. Companies in a single industry or theme may be affected by common economic, regulatory or technological developments and may be more volatile than the broad market. References to companies, governments or programs are for illustration only.