For decades, the semiconductor industry has relied on one simple material to move data between chips: copper.
Copper traces, cables and connectors have become extraordinarily sophisticated. NVIDIA CEO Jensen Huang has even joked that his company has made “copper sexy again.” But the extraordinary demands of artificial intelligence are exposing the physical limits of electrical connectivity—and pushing the computer industry toward a fundamentally different way of moving information: photons instead of electrons.
The shift is not happening because copper has suddenly become obsolete. Rather, AI is creating workloads in which the amount of data moving between processors is growing faster than copper-based connections can efficiently handle it. At some distances, copper remains the most practical option. But as AI clusters grow from racks to entire data-center campuses, optical interconnects are increasingly essential.
One of the clearest descriptions of this transition came on September 30 from Yannick De Koninck, a former NVIDIA silicon-photonics engineer who spent five years at the company. In an interview, De Koninck explained that AI models have become too large to run on a single GPU, requiring multiple GPUs to work together.
“Traditionally, that's been copper interconnects,” he said. “But as these GPUs get faster, they need more data, and they need more data at a faster rate—and copper is running out of steam.” The result, he said, is a transition to optical interconnects capable of much higher data-transfer bandwidth.
That observation goes directly to the heart of the issue.
AI is Creating an Opportunity to Problem Solve
Modern AI systems are not simply collections of increasingly powerful GPUs. They are enormous distributed computing systems in which thousands—and eventually hundreds of thousands or millions—of processors must exchange information continuously.
The faster the processors become, the more information has to move between them. That creates what engineers increasingly refer to as a bandwidth wall.
Copper has several disadvantages as electrical signaling speeds increase. Signal loss and interference become more difficult to manage, while the electrical connections themselves consume more power. Engineers can compensate with increasingly sophisticated SerDes technology, retimers, equalization and other techniques—but those solutions add complexity and energy consumption.
Optical connections approach the problem differently. Instead of pushing increasingly high-speed electrical signals through copper, data is converted into light and transmitted through optical fiber or integrated optical waveguides.
The physics become particularly compelling as distances increase.
Huang explained the tradeoff during a 2026 NVIDIA analyst Q&A: “You should use copper as long as you can, as much as you can. You should use optics whenever you must.”
Inside very short connections (about a meter), copper remains inexpensive, reliable and efficient. But when AI systems require enormous numbers of high-bandwidth connections extending across racks and facilities, optics increasingly becomes the technology of choice.
In March 2026, AMD, Broadcom, Meta, Microsoft, NVIDIA and OpenAI became founding members of the Optical Compute Interconnect Multi-Source Agreement (OCI-MSA), an effort to establish an open specification for optical scale-up interconnects. The group's announcement explicitly said traditional copper connectivity is reaching physical-reach limitations as AI clusters scale.
The industry is therefore not merely researching photonics. It is building products around it. Optical interconnects carry data as pulses of light rather than electrical current, enabling far higher bandwidth and greater energy efficiency than copper at the speeds modern AI clusters require. As AI model sizes and GPU counts have grown, copper interconnects have reached practical limits in bandwidth, reach and power consumption, making optical the leading replacement technology.
So, optical interconnects represent one of the clearest cases where a genuine engineering constraint — copper's physical limits — is forcing a technology transition. POET Technologies is among the companies built for the significant opportunity that transition presents.
The Challenge: Bringing Optics Closer to the Chip
The landscape of optical transceivers is rapidly evolving. Here are three architectures that are currently in the market:
- 1.Traditional pluggable optics: These are optics at the system edge with the optical transceiver physically separated from the switch ASIC or processor. Electrical signals travel from the chip across the board through copper traces to the transceiver, where they are converted into light.
- 2. CPO: Co-packaged optics are integrated into the package. The optical components are packaged alongside the switch ASIC, making the electrical connection extremely short.
- 3.NPO: Near-packaged optics are designed closer to the chip. In this case, the optical engine is placed very close to the ASIC/package. The goal is to dramatically shorten the high-speed electrical connection between the ASIC and the optical engine.
The industry is also exploring increasingly deep photonic-electronic integration alongside and potentially within computing packages. The reason is simple: The shorter the electrical path between the chip and the optical interface, the less electrical signaling has to travel through copper.
The POET Optical Interposer™ is designed to integrate electronic and photonic components at wafer level, using semiconductor manufacturing techniques. The company’s architecture can reduce conventional optical assembly, alignment and testing requirements while providing a scalable platform for optical engines.
POET’s “semiconductorization of photonics" approach is particularly relevant as the industry moves toward 1.6T, 3.2T and eventually higher-speed optical connectivity.
The optical interposer provides a highly integrated platform for building the optical engines inside high-speed transceivers and optical modules. It combines optical, electronic and passive components in a compact, scalable architecture, enabling efficient coupling between lasers, modulators, photodetectors and fibers. For transceiver manufacturers, POET’s optical engine product set offers a path to higher bandwidth, smaller footprints, lower power consumption and simplified assembly as data rates move to 1.6T and beyond. By integrating critical optical functions on a common interposer, POET can reduce component count and manufacturing complexity while supporting scalable, repeatable production—creating potential advantages in performance, cost, reliability and time-to-market for next-generation AI and data-center connectivity.
POET has also developed external light-source technologies based on its optical interposer architecture. Its POET Blazar™, for example, is a high-power, multi-channel external light source designed for high-bandwidth chip-to-chip optical links. POET’s wafer-level architecture is intended to reduce cost, improve manufacturing scale and use less of the scarce materials that generate light in lasers, such as Indium Phosphide.
Blazar is a clear example of a new product that is designed for what the industry needs — a low-cost, high-performance optical technology that is closer to the processor.
Industry Economics Reflect the Transition to Optics
The growth forecasts for optical interconnects are robust. LightCounting estimated that the ethernet optical-transceiver market reached $16.5 billion in 2025 and could reach $26 billion in 2026, representing approximately 60% growth in each of those years.
Its subsequent April forecast raised its 2026 growth expectation to 65%, while its July 2026 forecast became even more aggressive. LightCounting’s new expectations call for ethernet optical-transceiver sales to increase 73% in 2026, with the market reaching approximately $80 billion by 2031.
The research firm has even suggested that annual sales of optical interconnects used in AI clusters could potentially approach $100 billion by 2030, although it emphasizes that such an outcome depends on several factors lining up—including continued AI infrastructure investment and sufficient supply of GPUs, switch ASICs, lasers and other components.
TrendForce likewise expects CPO to gain steadily in AI data centers, projecting CPO penetration could reach 35% by 2030.
A Fundamental Change in Computing Is Upon the Industry
The significance of optical interconnects ultimately goes beyond networking.
For years, the industry focused primarily on making individual processors faster. AI is changing that equation. The performance of an AI system increasingly depends not just on the computational capability of each processor, but on how efficiently thousands of processors can operate together.
That makes the interconnect part of the computer.
As De Koninck put it in his recent discussion of AI infrastructure, “A GPU is no longer one chip.” At the largest scale, the system itself becomes the computer.
And when the computer consists of tens or hundreds of thousands of processors, the technology connecting those processors can become just as important as the processors themselves.
The Transition to Light Is Accelerating
For the chip-making industry that makes photonics much more than an enabling technology. It is becoming the foundation on which the next generation of AI computing will be built.
The demand for optical interconnects inside AI infrastructure is not a short-cycle trend. Every new generation of AI processor demands more bandwidth between chips and between servers. The transition from 800G to 1.6T and eventually 3.2T is a roadmap that plays out over years, not quarters — and each step requires more sophisticated photonic integration than the last.
POET aligns with that roadmap. The POET Optical Interposer platform is intended to serve successive speed generations without a fundamental redesign. It is one of the ways POET intends to help guide the evolution of photonics, which in 2026 has proven it will be essential to the age of AI.
Read More from the POET Blog: Why POET's Optical Interposer Is Built for the AI Cluster Era
