- All
- Product Name
- Product Keyword
- Product Model
- Product Summary
- Product Description
- Multi Field Search
Views: 0 Author: Site Editor Publish Time: 2026-05-12 Origin: Site
In the grand narrative of artificial intelligence, GPU chips are often the stars in the spotlight. With the constantly rising computing power indicators, manufacturers like NVIDIA have attracted the most capital attention and public gaze. However, no matter how powerful a GPU is, it cannot operate in isolation. When AI clusters with tens of thousands or even hundreds of thousands of cards are booming, how data flows at high speed between chips and servers is becoming the key lifeline restricting the development of artificial intelligence. Optical fiber, this seemingly insignificant filament, is precisely an indispensable "invisible pillar" in the intelligent computing system. It carries every transmission and interaction of massive data in the vast computing power network, quietly supporting every step of the large model's leap from training to inference. This article will deeply analyze how optical fibers have become the true cornerstone of the artificial intelligence era from multiple dimensions such as market fever, technical bottlenecks, cutting-edge breakthroughs, and future prospects.
To understand why optical fibers have become an indispensable support for AI, it is first necessary to understand what kind of tremendous changes the current optical communication market is undergoing. Over the past few years, the optical fiber industry has been in a slump - overcapacity, price wars, and price cuts by operators through centralized procurement, with the market remaining sluggish for a long time. However, in the second half of 2025, the situation took a dramatic turn. The global optical fiber market suddenly shifted from being "ignored" to "one fiber hard to come by" .
The requirements for bandwidth and latency in traditional data centers and AI supercomputing centers are completely not in the same league. For a cluster of tens of thousands of GPU cards, internal interconnection requires hundreds of thousands of fiber optic connections, which is 5 to 10 times the usage of traditional data centers. The astonishing growth trajectory of data can be intuitively felt from the above chart: from the "embryonic period" in 2018 when a single model only needed a few gigabytes of data transmission, to 2020 when BERT-Large opened the door to the "billion-level era" with hundreds of gigabytes of parameters, and then to 2023 when GPT-4 entered the "trillion-level threshold" relying on terabytes of data. By 2025-2026, the demand for single-GPU interconnection bandwidth from cutting-edge models has exceeded 800Gbps, and NVIDIA's GB200 NVL72 has an in-rack data transmission bandwidth as high as 1.8TB/s. Data shows that the proportion of AI-related optical fiber demand will surge from 5% in 2024 to 30% in 2027.
The sudden surge in demand is directly reflected in prices. Take the most common G.652.Dsingle-mode optical fiberas an example. Before New Year's Day 2026, its price was around 18 yuan per kilometer. By March this year, it had soared to 85 to 120 yuan per kilometer, with an increase of 372% to 650%. The G.654.E ultra-low loss optical fiber dedicated to AI computing power centers has also seen a growth rate of over 50%- 13%. North American fiber optic giant Corning has comprehensively raised long-term contract prices since February 2026. The price of its G.652.D fiber has risen by more than 400% since its low point in 2025, with some urgent orders even exceeding 250 yuan per core kilometer .
The optical fiber industry is entering a new cycle defined by "rigid supply" . Compared with the previous fiber optic boom cycle of 2017-2018, when the global fiber optic gap was only 40 million core kilometers, the gap has reached 180 million core kilometers by 2026. More importantly, the capacity utilization rates of overseas giants such as Corning and Fujikura have reached 100%, and there will be no effective release of new capacity before 2028 .
This imbalance between supply and demand reflects a deeper change: optical fibers are transforming from the "consumable" role in traditional telecommunications infrastructure to the "core blood vessels" in AI computing power networks, and the pricing power has completely shifted from a "buyer's market" to a "seller's market" .
If data and the market provide the reasons why the demand for optical fibers is so vigorous, then the evolution of technical bottlenecks reveals the fundamental logic of why optical fibers are irreplaceable. Looking back at the trajectory of AI hardware investment, every shift in bottlenecks has given rise to new core beneficiary directions . The year 2023 will be a "computing power shortage" - the most scarce are GPU chips. 2024 marks the "memory wall and power wall" - HBM and liquid-cooled temperature control solutions are experiencing rapid growth. By 2025-2026, AI computing power clusters have expanded from the ten-thousand-card level to the hundred-thousand-card level, and the data exchange volume between Gpus has grown exponentially. The "Internet wall" has become the core bottleneck at the current stage. Once the network connection is congested, the huge investment in computing power cannot be effectively released.
In traditional short-distance transmission, copper cables are still competent. However, at the current ultra-high transmission rates of 800G or even 1.6T, the effective transmission distance of copper cables is strictly limited. As the scale of GPU clusters continues to expand, the transmission distance limitations of copper cables under high-speed conditions are becoming increasingly prominent, and the importance of optical communication solutions is rapidly rising . The current cutting-edge models already require an interconnection bandwidth of over 800Gbps for a single GPU. Traditional copper interconnection solutions are no longer able to meet such extreme transmission demands.
In 2025, the majority of AI cluster switch port shipments will operate at 800Gbps, while by 2027, the mainstream will transition to 1.6Tbps, and by 2030, most ports will run at 3.2Tbps-20. This iteration pace far exceeds that of the traditional communication era - in the past, it often took several years to upgrade from 100G to 400G. Now, cloud vendors directly drive deliveries based on orders, and the pace of optical module rate iteration has advanced from "annually" to "monthly" -21.
In the traditional architecture, data signals need to leave the ASIC, cross the circuit board and connectors, and then be converted into optical signals. This method causes severe electrical signal loss, reaching approximately 22 decibels on a 200Gbps channel, and the power consumption of each port increases to 30W. Nvidia's proposed CPO (Co-Encapsulated Optics) solution integrates the optical engine with the switching chip, almost immediately coupling the signal onto the optical fiber, reducing electrical loss to 4 decibels, lowering power consumption per port to 9W, increasing power efficiency by 3.5 times, and enhancing elasticity by 10 times .
Nvidia has planned to officially launch a CPo-based optical interconnect platform in 2026, including Quantum-X800 InfiniBand switches and Spectrum X800 Ethernet switches, marking that the optical interconnect technology layout has entered a critical stage of commercialization .
When the scale of a single AI cluster approaches its physical limit, collaborative training across multiple data centers is becoming a new exploration direction. The role of optical fibers in this process is further highlighted - they are not only the "blood vessels" within the cluster, but also the "aorta" connecting the entire computing power network.
The distributed training of large models has become the central technical path in the development of artificial intelligence. However, with the rise of the Hybrid Expert Model (MoE), the traditional static GPU interconnection architecture is facing new challenges - the dynamic and non-uniform communication requirements of the MoE model have plunged the traditional static architecture into an efficiency predicament . The MixNet architecture proposed by the Hong Kong University of Science and Technology and other institutions, through regional reconfigurable optical switching technology, has for the first time achieved real-time topology reconfiguration in distributed MoE training, increasing the network cost efficiency by 1.2 to 2.3 times.
Facing the supply and demand contradiction between "idle scattered computing power resources" and "difficulty in obtaining high-quality large computing power", institutions such as China Telecom are exploring the realization of cross-cluster and cross-regional multi-data center collaboration through optical networks, and adopting a super-large-scale distributed training method under multiple intelligent computing nodes . The GeoPipe project, a collaboration between Beijing University of Posts and Telecommunications and Huawei, has made breakthrough progress in cross-domain training of large models. It has completed cross-domain distributed training of large models over a distance of hundreds of kilometers and under a 400G lossless optical layer networking, revealing the "bandwidth sweet spot" existing in optical networks between intelligent computing centers. It provides theoretical and experimental guidance for the networking design and bandwidth optimization of optical networks in intelligent computing centers .
What is more worth noting is that the three major chip giants, AMD, Broadcom, and NVIDIA, have joined forces with hyperscale data center operators such as Microsoft, Meta, and OpenAI to jointly establish the OCI Multi-Source Protocol Alliance. The goal is to define a universal optical fiber interconnection infrastructure for future AI clusters and promote the migration of AI clusters from copper cable architecture to optical fiber hardware architecture Ultimately, achieve a transmission capacity of 3.2Tbps or more per optical fiber . This is a rare industry collaboration, which comprehensively reflects the industrial wave of optical interconnection technology evolving from "individual standards" to "universal norms", and also indicates that the status of optical fibersin AI infrastructure is upgrading from an "optional solution" to a "mandatory foundation".
The significance of optical fibersgoes far beyond connection. As photonic large-scale interconnection technology becomes increasingly mature, more and more research is beginning to explore the direct establishment of computing cores on photonic logic and connections. The programmable three-dimensional photonic neural network chip developed by the team from Huazhong University of Science and Technology builds a three-dimensional waveguide interconnection network inside glass through femtosecond laser direct writing technology, with a theoretical computational throughput of 654 TOPS. The team from Xidian University has made significant progress in nonlinear photonic neuromorphic chips and photonic pulse reinforcement learning. Their software and hardware collaborative framework can be extended to various photonic pulse neural network hardware architectures and has broad application prospects in resource-constrained edge computing scenarios such as real-time decision-making of robots, autonomous driving, and somatic intelligence. The multi-functional programmable optoelectronic fusion gate array system LightIN, developed by the National Information Optoelectronics Innovation Center and other units, has achieved 4×4 bidirectional unitary matrix and 3×3 non-unitary matrix multiplication, with a computing speed exceeding 1.92TOPS and a photon core energy efficiency of 1.875pJ/MAC. This indicates that the integration of light and artificial intelligence is extending from the "physical layer" of data transmission to the "logical layer" of computing execution. The ultimate significance of optical fibers lies in enabling the entire AI infrastructure to operate at the speed of light.
According to LightCounting data, the global optical module market is expected to maintain a high growth rate of approximately 60% in 2026. By 2031, the global market size will approach 60 billion US dollars, with a compound annual growth rate of over 20% from 2025 to 2031. The latest research by TrendForce indicates that the global market for AI-specific optical transceiver modules is expected to expand from 16.5 billion US dollars in 2025 to 26 billion US dollars in 2026, with an annual growth rate of over 57% to 62%. The booming of the entire industrial chain of optical modules and optical fibers fully demonstrates that optical fiber and optical interconnection technology have become one of the most solid foundations in the intelligent era.
Looking ahead, the evolution of optical fiber and optical interconnection technologies will unfold along three core directions: First, the rate will continue to rise. The rate of AI cluster switching ports will move from 400G to 800G and steadily transition to 1.6T and even 3.2T, with the pace of optical module rate iteration continuously accelerating. Second, power consumption is continuously optimized. Silicon photonics technology and CPO/LPO solutions are accelerating the replacement of traditional high-power DSP architectures, effectively alleviating the pressure of power consumption and heat dissipation; Third, the application scenarios continue to expand, extending from internal interconnection within data centers to cross-data center collaboration, and from large model training to real-time edge inference coverage. Fiber optic networks are becoming intelligent computing infrastructure that covers the entire country and connects everything. China has taken solid steps in this field - the new architecture of intelligent computing DCN network based on all-optical switches by the Research Institute of China Telecom is effectively enhancing the performance, scalability and flexibility of ultra-large-scale clusters ; Nvidia is also expanding the high-speed connection that was originally limited within the cabinet to the scale of the entire data center and even across regions through NVLink Fusion technology. The framework of optical fiber networks is expanding at an astonishing speed, quietly outlining the grand outline of the future AI computing power landscape.
Optical fiber, this thin thread that connects Gpus, data centers, and optical computing chips, is the most silent yet indispensable invisible pillar in the world of artificial intelligence. From micron-level chip optical interconnection to optical switching matrices spanning intelligent computing center clusters, from the rapid evolution of optical modules to the breakthroughs in optical interconnection technology, fiber and photonics technologies are profoundly defining the physical boundaries of AI infrastructure. Understanding why optical fibers are the invisible pillar of intelligent computing is not only to unravel the fundamental logic behind why AI cannot do without light, but also to provide a solid and powerful perspective for grasping the technological lifeline of the entire intelligent era.
