Over the past two years, discussions around AI computing power have focused almost exclusively on GPUs: supply shortages of the H100, performance specs of the B200, and architectural roadmaps for next-generation GPUs have dominated industry narratives. However, as AI training clusters scale from thousands to tens of thousands, and even hundreds of thousands of GPUs, a deeper structural constraint is emerging—the efficiency of data movement between GPUs is becoming the ultimate ceiling for overall cluster performance.
At the start of 2026, Tencent optical network architect Si Dong Fu highlighted that from the Pascal architecture in 2016 to the Blackwell architecture in 2024, AI computing power has increased roughly 1,000-fold in eight years. Inference performance has grown 32 times in the past four years, while training power has expanded 16 times. Meanwhile, network bandwidth has only risen from 200G to 800G—a fourfold increase. This imbalance, where "computing power rockets ahead while networking walks," has made inter-node data transfer speeds the critical bottleneck for large-scale clusters, severely impacting overall efficiency and resource utilization.
This reality is reshaping both the investment logic and technology choices for AI infrastructure. As optical interconnect technology evolves from a localized performance enhancement to a foundational capability supporting large-scale AI cluster operations, understanding its technical rationale, market landscape, and industrial value has become essential for evaluating the AI computing sector. At the same time, the investment side is undergoing a similar structural shift—from single-asset allocation to multi-market synergy, forming a value chain that links computing infrastructure with financial infrastructure.
The Communication Challenge of 100,000-GPU Clusters: The Widening Gap Between Compute and Network
The efficiency of a GPU cluster is not determined by the peak performance of a single GPU, but by the time required for all GPUs to complete collaborative computation. In large-scale distributed model training, frequent parameter synchronization and gradient exchange mean that inter-node communication directly determines overall training efficiency. According to H3C’s CPO technology white paper, recent years have seen single-GPU performance increase far faster than network interconnect bandwidth. Most clusters keep adding GPUs on the compute side, but communication bandwidth expansion lags behind. As a result, communication time is taking up an ever-larger share of total training time, with GPUs waiting for data to arrive and overall effective computing power failing to scale proportionally with GPU count.
This phenomenon is backed by clear quantitative evidence. Tencent’s presentation shows that over the past four years, training power has grown 16 times, inference power 32 times, while network bandwidth has only increased from 200G to 800G—a fourfold rise. As clusters scale beyond ten thousand GPUs and move toward one hundred thousand, communication between GPUs shifts from simple point-to-point transfers to a complex system with thousands or tens of thousands of simultaneous links. Congestion or latency on any single link can slow down the entire training cycle.
A paper published by IEEE in February 2026 further confirms this assessment: as AI models grow, interconnects have become the key bottleneck in large GPU clusters, with traditional packet-switched networks facing mounting challenges in power consumption, cost, and scalability. Research shows that architectures based on optical circuit switching can reduce backbone layer power consumption by nearly 99% and lower eight-year lifecycle costs by 76%.
Industry data indicates that this structural imbalance is accelerating the expansion of optical communication infrastructure. UBS estimates that global fiber demand grew at an average annual rate of just 2% over the past five years, but with the rapid buildout of AI data centers, industry demand is expected to grow by more than 30% annually in the coming years, with data center-related fiber demand potentially achieving a compound growth rate of over 75%. Previously, 70% to 80% of fiber demand came from telecom operators; UBS projects that by 2030, enterprise and data center demand will account for more than 80%. The fiber industry is shifting from traditional communications to become a core component of AI infrastructure.
Optical Interconnects: The Definitive Solution to Compute Bottlenecks
Facing the widening gap between compute and network, optical interconnect technology is moving from a supplemental option to a foundational architectural choice. AI cluster expansion typically unfolds along three dimensions: Scale-up (vertical expansion, high-speed interconnects within a cabinet), Scale-out (horizontal expansion, interconnects across cabinets and nodes), and Scale-across (cross-domain interconnects, linking geographically dispersed data centers). Each dimension has different requirements for bandwidth, latency, power consumption, and transmission distance, but all point to the irreplaceable role of optical interconnects.
In Scale-up scenarios, optical interconnects mainly replace copper wires or electrical switches, enabling higher bandwidth and lower latency for intra-node communication. For example, NVIDIA’s NVL576 uses Spectrum-X Ethernet switches based on CPO technology, delivering switching capacity for 512×200Gbps ports and incorporating 32 1.6T silicon photonic engines for Scale-out and Scale-across scenarios. Huawei’s CloudMatrix 384 supernode adopts a fully peer-to-peer interconnect architecture, building a high-speed bus with 3,168 optical fibers and 6,912 400G LPO modules to pool and interconnect 384 NPUs, 192 CPUs, plus storage and memory resources.
On the technical front, the "x"PO technology family—represented by LPO, LRO, and CPO—is evolving rapidly. LightCounting reports that the global Ethernet optical module market will grow 35% year-on-year to $18.9 billion in 2026, and could exceed $35 billion by 2030, with demand for high-speed modules like 800G and 1.6T dominating. TrendForce forecasts that the share of global shipments for optical transceiver modules above 800G will rise from 19.5% in 2024 to over 60% in 2026. Based on Google’s projected shipment of nearly 4 million TPUs in 2026, demand for optical modules above 800G will exceed 6 million units.
Power consumption is a core challenge for pluggable optical modules. Google’s Apollo OCS technology uses micro-reflectors to directly connect data fibers, avoiding repeated conversions between optical and electrical signals that cause energy loss and latency; a single OCS switch consumes about 95% less power than traditional switches. In terms of latency, THine’s DSP-free chipset, designed for short-range LPO or CPO optical interconnects, can reduce latency by 90% and save 73% on power consumption.
Li Junjie, Deputy Director of the China Telecom Research Institute, noted at the start of 2026 that optical interconnect technology is evolving from a localized performance boost to a key capability supporting scalable, flexible, and highly reliable AI supernodes. Whether addressing bandwidth bottlenecks, power constraints, or capacity limits, optical interconnects have become the prerequisite for AI infrastructure’s evolution from thousands to hundreds of thousands of GPUs.
Ciena’s Strategic Shift: From Telecom Broadband to AI Optical Networks
As optical interconnects become central to AI infrastructure, the strategic choices of leading equipment vendors provide valuable insight into industry evolution. Ciena, a global leader in high-speed network systems, is undergoing a fundamental strategic transformation.
In the third quarter of fiscal 2025, Ciena reported revenues of $1.22 billion, driven mainly by sales of optical and routing platforms. At the same time, the company announced it would stop further development of its broadband PON business, redirecting R&D investment toward core optical and data center solutions, including out-of-band management technology, and cutting 4% to 5% of its workforce, with about $90 million in non-cash R&D write-offs. Ciena expects future growth to come primarily from the AI and hyperscale cloud markets.
CEO Gary Smith stated during the earnings call that service provider customers are focusing network investments on areas that enable scale to support AI traffic growth, creating new system requirements and interconnect opportunities that ultimately extend into data centers. Ciena said hyperscale cloud customers account for about 50% of its business, and expects a similar customer mix in 2026.
Ciena has already made tangible progress in AI infrastructure. The company highlighted a North American AI infrastructure project involving regional GPU cluster interconnects for training and geographic distribution, featuring its RLS platform and WaveLogic 6 Nano 800-gig ZR plug-ins. Its DCOM out-of-band management solution targets data center operations, helping hyperscale operators simplify large-scale data center installation and management, improve scalability, and reduce power and space requirements.
From a broader industry perspective, Ciena’s strategic pivot reflects a qualitative leap in AI data center optical network demand. Jürgen Hatheier, Ciena’s Chief Technology Officer for Business Development and Global Partnerships, noted a clear market shift toward higher-capacity optical connections, with strong demand for 1.6T wavelengths expected to continue through 2026. Rob Shore, Nokia’s Optical Networks Portfolio Marketing Lead, predicts that 800G coherent pluggable modules will become the standard optical connection for AI networks in 2026.
The AI data center network market is growing exponentially. Industry data shows the market will expand from $10.31 billion in 2025 to $12.8 billion in 2026, a compound annual growth rate of 24.2%, and is projected to reach $30.17 billion by 2030. Demand for optical cables for AI applications is expected to grow 77% in 2025, with a five-year compound annual growth rate of 26% through 2029—far outpacing non-AI applications. Ciena is positioned at the heart of this structural growth curve.
From Compute Infrastructure to Financial Infrastructure: Gate’s Equity Trading Landscape
Infrastructure evolution is happening not only at the compute level but also in asset allocation. As optical interconnects in AI data centers become the critical infrastructure determining GPU cluster efficiency, the investment side’s multi-asset allocation capabilities require equally efficient supporting infrastructure.
Gate’s expansion into traditional finance is progressing steadily. In January 2026, the platform launched TradFi CFD functionality, covering gold, forex, stock indices, commodities, and popular equities. In March, it expanded to include stock tokens and leveraged ETFs. By June, Gate, through a strategic partnership with Alpaca, officially rolled out real stock trading services.
Gate now supports over 10,000 US stocks and ETFs, covering companies listed on major exchanges such as the NYSE and Nasdaq, far surpassing most tokenized stock platforms that typically support only a few hundred assets. Users can invest directly in the US securities market using USDT, with fractional share trading starting at just 0.01 shares, allowing participation in leading US stocks with as little as $1.
On the technical and partnership front, Gate connects with compliant brokers holding US Broker-Dealer licenses and clearing qualifications, integrating with major exchanges like NYSE and Nasdaq. Each share is backed by real assets independently custodied via the DTC system, not by on-chain derivatives or RWA-mapped products. Holders automatically enjoy full shareholder rights, including dividends, stock splits, and rights issues.
Industry trends show that leading crypto platforms integrating stock trading is a clear direction. Data reveals that 73% of crypto traders also hold traditional assets. Gate’s approach enables real stock trading through regulated infrastructure, not synthetic or tokenized representations, ensuring users get genuine price discovery and settlement. Combined with its CFD products, Gate is evolving from a single crypto asset exchange to a multi-asset hub spanning crypto, traditional finance, and derivatives.
This evolution aligns with the broader trend of RWA asset tokenization. In September 2025, Gate launched the Ondo zone, introducing tokenized stocks and ETFs from well-known companies like Apple, Tesla, and Microsoft. The RWA sector’s total value locked has surpassed $15.7 billion, with Ondo Finance ranking third globally at about $1.66 billion. From real stocks to tokenized stocks to equity CFDs, Gate is building a multi-layered allocation channel covering various asset forms.
Conclusion
The development path of optical interconnect technology points clearly to a fundamental truth: the competitiveness of AI data centers is shifting from single GPU performance metrics to system-level efficiency indicators. Networks are no longer just a supporting layer for compute clusters—they are now the prerequisite for realizing the theoretical compute power of 100,000-GPU clusters. In this context, the strategic value of optical networking infrastructure companies is being reassessed by the market—Ciena’s decisive pivot to AI optical networks is the most direct illustration of this trend.
Meanwhile, infrastructure evolution on the investment side is equally significant. As AI computing power becomes the core productive force of the digital era, platforms that can effectively connect this productivity with global capital are seeing their value anchors shift systemically. From compute to network, from hardware to assets, the intersection of technological advancement and financial innovation is often where structural opportunities are born.
