The artificial intelligence revolution has fundamentally altered the economics and technical trajectory of data center networking. Where once network upgrades follo
wed a predictable cadence driven by enterprise IT refresh cycles, today’s AI training clusters demand bandwidth increases that outpace Moore’s Law. The optical transceiver industry is now racing to keep up—and the finish line keeps moving.
The AI Bottleneck That Changed Everything
For years, data center networks were designed around a simple premise: compute was expensive, and networking was cheap. That equation has inverted. Modern large language models and generative AI workloads generate staggering volumes of east-west traffic during training, with GPU clusters requiring near-lossless, ultra-low-latency communication to maintain utilization rates above acceptable thresholds. A single NVIDIA GB200 server, for example, requires up to 72 1.6T optical modules to interconnect its GPUs—a level of density that would have been unthinkable just three years ago.
The math is brutal and revealing. Global AI computing demand spiked 400% year-over-year in early 2026, while effective supply grew only 128%, creating a 46% capacity gap. Every 8-GPU AI server requires high-speed interconnection, and a single NVL72 pod demands hundreds of 400G switches to maintain optimal performance. Networking is no longer a supporting actor in the AI infrastructure story—it has become the bottleneck that determines whether trillion-parameter models train in weeks or months.
1.6T Transceivers: The New Baseline for AI Clusters
The industry’s response has been swift and decisive. 1.6T optical transceivers have emerged as the critical interconnect infrastructure for next-generation AI clusters, bridging the gap between GPU throughput and network capacity. The global datacom optical module market exceeded US$6.25 billion in 2023 and is projected to reach US$25.8 billion by 2029, a compound annual growth rate of 27%. More specifically, global demand for 1.6T optical modules is projected to reach 3–5 million units in 2025, with a market value exceeding US$1 billion, and some analysts expect that number to surpass 10 million units by 2026.
From a technical perspective, achieving 1.6T transmission requires 200G/lane modulation—eight lanes running at 200Gbps each. The OSFP-XD form factor has become the absolute mainstream for 1.6T transceivers due to its larger size and superior heat dissipation capability, which can accommodate the complex Digital Signal Processors (DSP) and optical engines required for high-power 1.6T operation. The QSFP-DD form factor, by contrast, faces significant challenges at the 1.6T rate, with limited heat dissipation and power headroom making it viable only for specific short-reach scenarios.
Silicon photonics is emerging as a key battleground among major manufacturers. While EML (Electro-absorption Modulated Laser) has traditional advantages in single-mode solutions, silicon photonics offers immense potential in terms of integration, cost, and power consumption. TrendForce estimates that the global shipment share of 800G-and-above optical transceiver modules will climb from 19.5% in 2024 to over 60% by 2026, positioning these modules as standard components in AI-focused data centers.
InfiniBand’s Role in the High-Performance Networking Stack
While Ethernet remains the workhorse of general-purpose data center networking, InfiniBand has carved out a dominant position in AI and HPC environments where latency and lossless fabric are non-negotiable. NVIDIA’s Quantum-2 InfiniBand platform, built around 400G NDR (Next Data Rate) technology, delivers 64×400Gb/s ports in a compact 1U chassis, achieving 51.2Tb/s bidirectional throughput with ultra-low latency around 200 nanoseconds.
The 400G OSFP/QSFP112 InfiniBand NDR transceivers are at the heart of these deployments. These modules use 4×100G PAM4 modulation per lane and are available in both OSFP and QSFP112 form factors. The QSFP112-SR4-400G, for example, supports transmission distances up to 50 meters over OM4 multimode fiber and is primarily used with NVIDIA ConnectX-7 adapters. On the switch side, these connect to OSFP-SR8-800G optical modules on QM9700/9790 switches. The QM9790 switch features 32×800G OSFP physical interfaces that support two InfiniBand links per connector, enabling 64×400G NDR ports.
The shift from active optical cables (AOCs) to pluggable transceivers in InfiniBand NDR (400G) systems marks a significant architectural evolution. NDR connections now use pluggable 400/800G SR4/SR8 and DR4/DR8 transceivers instead of AOCs, providing greater flexibility and density. The latest XDR systems are already moving to 1.6T DR8 and 2xFR4 pluggables, including some LPO (Linear-drive Pluggable Optics) variants. Global demand for NDR 400G InfiniBand switches—including the QM9790 and QM9700—reached 4,500–5,000 units in May 2026, representing a 35% month-over-month increase.
Beyond the Hyperscale: The Long Tail of High-Speed Connectivity
It would be a mistake, however, to view the optical transceiver market solely through the lens of AI hyperscalers. The broader ecosystem of enterprise networking, telecom infrastructure, and specialized applications continues to drive meaningful demand across the entire speed spectrum.
Consider the 10G BiDi SFP+ transceiver—a technology that might seem almost quaint in an era of 1.6T optics, yet remains indispensable for a vast installed base of enterprise and service provider networks. These BiDi (bidirectional) modules transmit and receive on different wavelengths over a single fiber strand, effectively doubling fiber utilization without requiring new cable infrastructure. With transmission distances ranging from 10 kilometers to 100 kilometers over single-mode fiber, 10G BiDi SFP+ modules are widely deployed in data center interconnects, metro Ethernet networks, high-availability architectures, and FTTX applications.
The coexistence of 10G SFP+ and 1.6T OSFP-XD transceivers within the same data center ecosystem illustrates a fundamental truth about networking: bandwidth requirements are not monolithic. Front-end networks handling management traffic, storage area networks, and legacy enterprise workloads continue to operate efficiently at 10G and 25G speeds, while back-end AI fabrics push the boundaries of what is technically possible at 1.6T and beyond. This stratification creates a diverse and resilient market where innovation at the high end eventually trickles down to improve cost and performance across all speed grades.
The Road Ahead: LPO, CPO, and the Next Frontier
Looking beyond 1.6T, the industry is already preparing for the next wave of innovation. LPO (Linear-drive Pluggable Optics) and CPO (Co-packaged Optics) represent two competing paradigms for reducing power consumption and latency in high-speed interconnects. LPO eliminates the DSP from the optical module, shifting signal processing to the switch ASIC, while CPO integrates optics directly onto the switch package. NVIDIA is actively developing both approaches, with deployments expected to begin in scale-up networks starting in 2026-2027 and reaching high volumes by 2028.
Google’s Apollo optical circuit switch (OCS) architecture offers another glimpse of the future. By using MEMS micro-mirrors to enable direct fiber-to-fiber connections and avoid repeated optical-electrical-optical conversions, Apollo achieves roughly a 95% reduction in power consumption compared with traditional switches—around 100 watts versus 3,000 watts. Critically, upgrading bandwidth from 800G to 1.6T in this architecture requires only swapping in higher-speed optical modules, rather than rebuilding the entire system.
The optical transceiver market is no longer a slow-moving commodity business. It is now a strategic battleground where the winners will shape the infrastructure of the AI era. With silicon photonics penetration expected to reach 70% and the “optical supercycle” already underway, the next few years will determine which technologies and which companies define the future of high-performance networking. One thing is certain: the bandwidth race shows no signs of slowing down.