The Computing Bottleneck Shift of a 100,000-Card AI Cluster: How Optical Interconnection Becomes the Core of New Infrastructure?

Over the past two years, discussions about AI computing power have almost entirely revolved around GPUs: the H100 supply gap, the B200 performance specifications, and the architecture roadmap for next-generation GPUs, which together form the main storyline of the industry. However, as AI training clusters scale from the kilocard level to ten-thousand-card and even hundred-thousand-card levels, a deeper structural constraint is emerging—efficiency of data flow between GPUs is becoming the ultimate ceiling that determines the cluster’s overall computational throughput.

In early 2026, Tencent optical network architect Fu Sidong pointed out that from the Pascal architecture in 2016 to the Blackwell architecture in 2024, AI computing power increased by about 1000 times within eight years; over the past four years, inference computing power grew 32 times, and training computing power grew 16 times. Meanwhile, during the same period, network bandwidth increased only from 200G to 800G, which is just a fourfold increase. This imbalance—“computing power climbing like a rocket while the network moves forward on foot”—makes the data transmission speed between nodes a critical bottleneck for clusters at the ten-thousand- to hundred-thousand-card scale, seriously affecting overall cluster efficiency and resource utilization.

This reality is reshaping the investment logic and technical roadmap choices for AI infrastructure. As optical interconnect technology evolves from local performance compensation to a key capability supporting the large-scale, scaled operation of AI clusters, understanding its technical logic, market landscape, and industrial value becomes an unavoidable foundational issue for evaluating the AI computing power race track. At the same time, the investment side is undergoing a similar structural shift—moving from single-asset allocation to coordinated multi-market strategies—while the value chain connecting computing infrastructure and financial infrastructure is taking shape.

Communication Dilemmas of Hundred-Thousand-Card Clusters: A “Scissors Gap” Between Computing Power and Network

The efficiency of GPU clusters is not determined by the peak computing power of a single GPU, but by the time required for all GPUs to complete collaborative computation. In distributed training of large language models, frequent parameter synchronization and gradient exchange mean that the communication capability between nodes directly determines overall training efficiency. In its CPO technology white paper, H3C points out that in recent years, the rate at which single-card computing power improves has far outpaced the evolution of network interconnect bandwidth. As a result, many clusters keep adding GPUs on the compute side, but communication bandwidth expansion lags behind; consequently, the share of communication time in total training time keeps increasing. GPUs spend long periods waiting for data to arrive, and overall effective computing power becomes difficult to scale proportionally with the number of GPUs.

This phenomenon has clear quantitative support. Data from Tencent’s presentations shows that over the past four years, training computing power increased 16 times and inference computing power increased 32 times, while network bandwidth increased only from 200G to 800G—an increase of 4 times. When cluster scale surpasses 10,000 cards and moves toward 100,000 cards, the communication pattern between GPUs is no longer simple point-to-point data transmission; instead, it becomes a complex system involving thousands or even tens of thousands of links operating simultaneously. Congestion or latency on any single link can slow down the entire training iteration cycle.

A paper published by IEEE in February 2026 further confirms this assessment: as AI model scale grows, interconnects have become a key bottleneck in large-scale GPU clusters, and traditional packet-switched networks face increasingly severe challenges in power consumption, cost, and scalability. Research shows that architectures based on photonic circuit switching can reduce backbone-layer power consumption by nearly 99%, and lower lifecycle costs by 76% over an eight-year period.

From industry data, this structural contradiction is driving an acceleration in the expansion of optical communications infrastructure. UBS estimates that global fiber demand has grown at an average annual rate of only about 2% over the past five years. But as AI data center construction accelerates, the industry’s demand growth in the coming years is expected to exceed 30%, and fiber demand related to data centers could potentially achieve compound growth of 75% or more. Previously, 70% to 80% of fiber demand came from telecom operators; UBS expects that by 2030, the share of demand related to enterprises and data centers will exceed 80%. The fiber industry is shifting from a traditional telecommunications industry to a core component of AI infrastructure.

Optical Interconnect: A Deterministic Technical Path to Address the Computing Bottleneck

Faced with the “scissors gap” between computing power and network, optical interconnect technology is rising from a supplemental solution to a fundamental architectural choice. AI cluster scaling typically expands along three dimensions: Scale-up (vertical expansion, corresponding to high-speed interconnects between GPUs within a rack/cabinet), Scale-out (horizontal expansion, corresponding to cluster interconnects across nodes between cabinets/racks), and Scale-across (inter-domain interconnect, corresponding to connections between geographically distributed data centers). The requirements for bandwidth, latency, power consumption, and transmission distance differ across the three dimensions, but they all point to the irreplaceability of optical interconnects.

In the Scale-up scenario, optical interconnect primarily replaces copper cables or electrical switches to achieve higher bandwidth and lower latency for intra-node communication. Taking NVIDIA’s NVL576 as an example, it uses Spectrum-X Ethernet switches based on CPO to provide a switching capacity for 512×200Gbps ports, and includes 32 1.6T silicon photonic optical engine modules, used for Scale-out and Scale-across scenarios. Huawei’s CloudMatrix 384 supernode adopts a fully peer-to-peer interconnect architecture. It builds a high-speed interconnect backbone bus using 3168 optical fibers and 6912 400G LPO modules, connecting and pooling all 384 NPU units, 192 CPU units, and resources such as storage and memory.

At the technology-path level, the “x”PO technology family represented by LPO, LRO, and CPO is accelerating its evolution. LightCounting data shows that the global Ethernet optical module market size is expected to grow 35% year-over-year in 2026 to $18.9 billion. By 2030, it may surpass $35.0 billion, and demand for high-speed optical modules such as 800G and 1.6T will dominate the market. TrendForce expects that in 2026, the shipment share of optical transceivers above 800G globally will rise from 19.5% in 2024 to more than 60%. Based on the estimated shipment of Google TPU at nearly 4 million units in 2026, demand for optical modules above 800G would exceed 6 million units.

Power consumption is one of the core challenges facing pluggable optical modules. Google’s Apollo OCS technology uses micro reflective mirrors to directly connect data optical fibers, avoiding the energy consumption and latency caused by repeated optical-electrical conversions in traditional technologies. Compared with a traditional switch, the power consumption of a single OCS switch is reduced by about 95%. In terms of latency, THine’s optical DSP chipset—designed for short-distance optical interconnect scenarios adapted to LPO or CPO—can achieve a 90% latency reduction and a 73% power savings.

In early 2026, Li Junjie, Vice President of China Telecom Research Institute, stated that optical interconnect technology is evolving from local performance compensation into a key technological capability supporting the large-scale, flexible, and highly reliable operation of AI supernodes. Whether it is addressing speed bottlenecks, power constraints, or capacity limitations, optical interconnect has become a prerequisite condition for AI infrastructure to evolve from the kilocard level to ten-thousand-card clusters.

Ciena’s Strategic Pivot: From Telecom Broadband to AI Optical Networks

When optical interconnect becomes the core proposition of AI infrastructure, the strategic choices made by leading equipment vendors in this area become an important window for understanding industry evolution. Ciena, a global leader in high-speed connectivity network systems, is undergoing a fundamental strategic adjustment.

In the third quarter of fiscal 2025, Ciena reported revenue of $1.22 billion, mainly driven by sales of optical and routing platforms. At the same time, the company announced it would stop further development of broadband PON business. It will redirect R&D investment toward core optical and data center solutions, including out-of-band management technology, and cut 4% to 5% of its employees, recording approximately $90 million of non-cash R&D expense write-offs. Ciena expects that future growth will mainly come from the AI and hyperscale cloud provider markets.

Ciena’s CEO Gary Smith said during the earnings call that service provider customers are concentrating network investments in areas that can achieve economies of scale to carry AI traffic growth. This is creating new system demands and interconnect opportunities, and ultimately extending into data centers. Ciena stated that hyperscale cloud providers account for about 50% of its business, and its customer mix in 2026 is expected to be similar.

Ciena has already seen results in specific deployments within AI infrastructure. The company points to a North American AI infrastructure project involving interconnection of regionally distributed GPU clusters used for training, including its RLS platform and WaveLogic 6 Nano 800-gig ZR pluggable modules. In addition, its DCOM out-of-band management solution, aimed at data center internal operations and maintenance scenarios, can help hyperscale operators simplify the installation and management of large-scale data centers, improving scalability while reducing power consumption and space usage.

From a broader industry perspective, Ciena’s strategic pivot reflects the qualitative leap in optical network demand driven by AI data centers—from “quantitative change” to “qualitative change.” Jürgen Hatheier, Chief Technology Officer of Ciena’s global partners, said that the market is clearly shifting toward higher-capacity optical connections, and it has already seen strong demand for 1.6T wavelengths, which the company expects to continue into 2026. Rob Shore, Head of Marketing for Nokia’s optical networking product portfolio, expects that in 2026, 800G coherent pluggable modules will become the standard optical connectivity solution for AI networks.

The market size for AI data center networks is growing exponentially. According to industry data, this market will grow from $10.31 billion in 2025 to $12.8 billion in 2026, with a compound annual growth rate (CAGR) of 24.2%, and is expected to reach $30.17 billion by 2030. Among them, optical cabling demand for AI applications is expected to grow 77% in 2025. Over the five years to 2029, the CAGR is expected to be 26%, far higher than non-AI applications. Ciena is positioned at the core of this structural growth curve.

From Computing Infrastructure to Financial Infrastructure: Gate’s Stock Trading Landscape

The evolution of infrastructure is happening not only at the computing layer, but also at the asset allocation layer. When optical interconnect in AI data centers becomes a key infrastructure that determines GPU cluster efficiency, the investment side’s multi-asset allocation capabilities also need to be supported by correspondingly efficient underlying infrastructure.

Gate’s footprint in the traditional finance sector is progressing steadily. In January 2026, the platform first introduced TradFi contract-for-difference (CFD) functionality, covering gold, foreign exchange, stock indexes, commodities, and popular stocks. In March, it further expanded to include stock tokens and leveraged ETFs. In June, Gate officially launched real stock trading services through a strategic cooperation with Alpaca.

At present, Gate supports more than 10,000 US stocks and ETFs. It covers companies listed on major exchanges such as the New York Stock Exchange and Nasdaq, far exceeding the hundreds of assets typically supported by most tokenized-stock platforms. Users can use USDT directly to participate in investments in the US mainstream securities market. With fractional share trading with a minimum of 0.01 shares, users can participate in investments in top US stocks with an initial amount as low as $1.

On the technical and partnership side, Gate connects with compliant brokers holding US Broker-Dealer licenses and clearing qualifications, with underlying access to major exchanges such as the NYSE and Nasdaq. Each share is backed by real assets independently held in custody through the DTC system—not on-chain derivatives or RWA-mapped products. Users with open positions can automatically enjoy full shareholder rights such as dividends, rights offerings, and stock splits.

From industry trends, integrating stock trading into leading crypto platforms has become a clear direction. Data shows that 73% of crypto traders also hold traditional assets. Gate’s approach ensures users receive genuine price discovery and settlement by conducting real stock trading through regulated infrastructure rather than synthetic or tokenized representations. Combined with the platform’s CFD products, Gate is evolving from a single crypto-asset exchange into a multi-asset center integrating crypto, traditional finance, and derivatives.

This evolution aligns with the macro trend of tokenizing RWA assets. In September 2025, Gate officially launched the Ondo zone, introducing on the first batch tokenized stocks and ETFs from well-known companies such as Apple, Tesla, and Microsoft. The total value locked (TVL) of the RWA sector has exceeded $15.7 billion, and Ondo Finance, with approximately $1.66 billion in locked value, ranks third globally. From real stocks to tokenized stocks, and then to stock CFDs, Gate is building a multi-layer asset allocation channel that covers multiple asset forms.

Conclusion

The evolution path of optical interconnect technology points clearly to a basic fact: the competitiveness of AI data centers is shifting from a single metric—GPU computing power—to a system-level efficiency metric. The network is no longer merely an auxiliary support layer for compute clusters; it is the prerequisite condition for whether hundred-thousand-card clusters can truly realize their theoretical computing power. Under this logic, the strategic value of optical network infrastructure companies is being reassessed by the market—the fact that Ciena is fully pivoting to AI optical networks is the most direct illustration of this trend.

At the same time, the evolution of infrastructure on the investment side also cannot be overlooked. As AI computing power becomes a core productivity factor in the digital age, the platform that effectively connects this productivity with global capital is also undergoing systematic shifts. From computing to networking, and from hardware to assets, the intersection of technological evolution and financial innovation is often where structural opportunities become concentrated and emerge.