AI Compute Interconnect: Industry Trends and Investment Logic

    This article examines the shift in AI infrastructure spending from experimental to strategic, the evolution of large models and agents, and five key technology themes across the AI compute supply chain: accelerators, optical interconnect, fiber supply and demand, liquid cooling, and advanced thermal materials. It also outlines evaluation criteria and architecture planning recommendations for technical teams and business decision-makers.

    1. Industry Background: Compute Investment Enters a High-Intensity Cycle

    In recent years, capital expenditure by leading global cloud service providers has continued to climb, and investment in AI infrastructure has remained intense. Public industry data shows that the combined quarterly capex year-over-year growth rate of mainstream cloud providers has exceeded 80%, with the vast majority flowing into data centers, accelerators, high-speed interconnect networks, and other compute-related areas. This trend indicates that AI compute has moved from experimental spending to strategic infrastructure investment.

    At the same time, capital market pricing for the compute sector has become more rational. For example, the valuation level of the communications industry index has fallen back to around its historical midpoint, and market focus has shifted from concept expectations to earnings delivery and growth sustainability. This means the core criterion for judging value in the AI compute supply chain is shifting from whether a company is associated with the AI concept to whether it has real orders, capacity, and delivery capability.

    2. Large Model Industry: From Capability Breakthroughs to Deeper Application

    2.1 Model Iteration and Explosive Token Usage

    Over the past year, leading overseas frontier models have maintained frequent iteration, while domestic models have continued to improve competitiveness through architecture innovation and inference efficiency optimization. Global token usage has expanded rapidly, and inference costs have continued to fall, laying the foundation for an application-layer breakout. In this process, AI coding tools were among the first to form a relatively clear commercialization path: developers are willing to pay for capabilities that significantly improve coding efficiency, and enterprises are willing to pay for quantifiable productivity gains. This has made code generation one of the most successful commercialization scenarios for large models.

    2.2 Scaled Agent Adoption

    Looking ahead, the large model industry is expected to move from rapid model capability breakthroughs and explosive token usage into a stage of scaled agent applications, improved commercialization efficiency, and accelerated expansion of new scenarios. The value of enterprise agents is no longer limited to conversational Q&A; they are embedded in business processes and handle complete workflows such as task planning, tool invocation, and result verification. The core metrics for evaluating agent value should focus on whether they deliver real business value and sustainable revenue, rather than demo effects alone.

    2.3 New Scenarios Expand Long-Term Space

    Beyond coding, design, modeling, simulation, and other broadly programming-related domains, as well as AI for Science and robotics and other physical AI directions, are opening up longer-term growth space. These scenarios share common traits: they are data-intensive, compute-intensive, and have verifiable results, making them well suited for deep integration with large models and compute infrastructure. In addition, if frontier research directions such as recursive self-improvement (RSI) achieve breakthroughs, they could drive new leaps in model capability and further amplify compute demand.

    3. Five Technology Themes in the AI Compute Supply Chain

    3.1 Investment Opportunities Spread from Models to the Full Stack

    Investment opportunities in the large model sector are spreading from the model layer to the application layer, compute layer, and ecosystem layer. For technical decision-makers, this means establishing a full-stack view: the application layer should focus on real business value; the compute layer should focus on effective compute supply and cluster delivery capability; and the ecosystem layer should focus on the value of developer toolchains and user entry points. The risk of betting on a single link is rising.

    3.2 Domestic Accelerator Chips: Delivery and Share Gains

    Domestic GPU and other accelerator chip companies are entering a period of earnings delivery. With software-hardware co-optimization, maturing cluster interconnect solutions, and a rising share of inference workloads, the cost-performance advantages of domestic chips on some workloads are gradually emerging, and there is still significant room for future market share gains. When evaluating domestic compute, the focus should be on effective compute supply rather than paper peak compute, the stability and delivery capability of large-scale clusters, and the adaptation maturity of mainstream frameworks.

    3.3 Optical Interconnect: A Core Incremental Growth Area with Rising Share

    As cluster scale expands, interconnect bandwidth has become a key bottleneck constraining compute efficiency, and the share of optical interconnect in overall investment is expected to continue rising:

    • High-speed optical modules: 800G demand continues to grow rapidly, 1.6T is entering a volume ramp, and the pace of specification upgrades has clearly accelerated.
    • Scale-up interconnect: Optical interconnect demand corresponding to scale-up supernode architectures is becoming an important incremental area, and the trend toward optical intra-rack and inter-rack interconnect is clear.
    • New packaging and architectures: Near-package optics, co-packaged optics, and other solutions will reach scale, with optoelectronic integration deepening.
    • Emerging directions: Optical switching, data center interconnect, and new display and light-source technologies are accelerating, bringing more possibilities to network architecture.

    3.4 Fiber Supply and Demand: AI Demand Sustains Strong Conditions

    Demand for high-density fiber connections in AI data centers, combined with limited capacity for some high-end categories, is keeping fiber prices high. In the near term, the supply-demand balance is expected to remain tight, and vendors with high-end fiber capacity and delivery capability will continue to benefit.

    3.5 Liquid Cooling and Advanced Thermal Materials

    As per-rack power density rises rapidly, air cooling can no longer meet the cooling needs of high-density compute clusters, and liquid cooling has moved from an optional solution to a mandatory one. Relevant liquid cooling vendors will enter a key earnings delivery period; cold plate solutions are scaling first, while immersion solutions are gradually penetrating ultra-high-density scenarios. At the same time, the use of high-thermal-conductivity materials represented by diamond in chip-level cooling is accelerating, and their engineering progress is worth watching.

    4. Conclusion and Recommendations

    Overall, the long-term logic of the AI compute supply chain is clear: improvements in model capability and deeper application drive continued growth in token demand, which in turn drives systemic upgrades in compute, interconnect, cooling, and other infrastructure. Technical teams and business decision-makers should grasp five main themes: at the model layer, focus on commercialization revenue and profitability; for enterprise agents, focus on real business value and recurring revenue; for domestic compute, focus on effective compute supply and cluster delivery; for AI for Science and physical AI, focus on commercial partnerships and scenario deployment; and for personal agents, focus on user entry points and ecosystem value. In architecture planning, high-speed interconnect and liquid cooling capabilities should be laid out in advance to address the engineering challenges brought by rising compute density.