AI Infrastructure Enters a Three-Dimensional Leap: CPO, CPU, and Space-Based Computing Converge
AI Computing Infrastructure Enters a Three-Dimensional Leap Phase: The Triple Resonance of CPO Optical Interconnects, CPU Supply Restructuring, and Space-Based Computing Policies
Global AI infrastructure development is undergoing a quiet yet profound paradigm shift—from its early focus on isolated GPU performance breakthroughs toward a holistic, three-dimensional upgrade spanning “chips–networks–space.” Three pivotal signals have recently converged with striking intensity: NVIDIA’s Spectrum-X switches—powered by Co-Packaged Optics (CPO)—have entered mass production; Intel CEO Dr. Lip-Bu Tan publicly confirmed an explosive global surge in CPU demand, prompting urgent calls from CEOs of technology firms across multiple countries seeking immediate supply; and Beijing Yizhuang has launched construction of a Space-Based Computing Innovation Center and convened a dedicated enterprise symposium. These are not isolated, coincidental events—but rather the inevitable outcome of exponential AI compute demand forcing systemic expansion across the entire infrastructure stack. This marks a definitive transition in AI infrastructure investment logic—not just in China but globally—from “securing individual strategic nodes” to “orchestrating full-stack synergy.”
CPO Optical Interconnects Enter Mass Production: Breaking the “Bandwidth Wall” in AI Clusters, Boosting Energy Efficiency Fivefold
Traditional AI training clusters have long been bottlenecked by “copper-cable limitations”: In GPU clusters scaling to tens of thousands of cards, multiple copper-cable hops link PCIe devices and Ethernet switches—causing severe signal attenuation, soaring power consumption, and complex thermal management. As disclosed by NVIDIA, its Spectrum-X platform integrates 800G optical engines directly into the switch ASIC package for the first time, achieving true co-packaging of optical modules and application-specific integrated circuits (ASICs). This architecture slashes end-to-end latency by 40%, reduces per-bit power consumption by 75%, and improves overall system reliability fivefold—effectively doubling effective compute density within the same rack footprint. More critically, the CPO architecture natively supports future evolutionary paths toward 1.6T and even 3.2T optical interconnects, delivering deterministic low-latency networking essential for ultra-large-scale training clusters exceeding 1,000 GPUs. Domestic optical module leaders—including Zhongji Xunlai (ICT) and New Bright—have already entered NVIDIA’s supply chain, while high-speed laser diode vendors such as Changguang Huaxin and Yuanjie Technology are accelerating R&D on CPO-compatible chips. Notably, CPO imposes exceptionally stringent requirements on advanced packaging technologies (e.g., silicon photonics wafer-level packaging), driving leading OSATs—including JCET and Tongfu Microelectronics—to rapidly deploy hybrid Chiplet + optical interconnect packaging lines.
CPU Supply Restructuring: The “New Foundation” of the AI Era Exceeds Expectations, Driving Upstream Pressure and Upside
While market attention remains fixed on GPUs, CPUs are quietly emerging as the invisible backbone of AI infrastructure. Dr. Tan’s remarks at Computex Taipei were highly revealing: “Over the past four weeks, numerous company CEOs have called me directly requesting more CPUs.” This reflects acute, structural demand for high-frequency, large-cache, multi-core parallel CPUs—driven by AI inference services, vector databases, real-time recommendation systems, and other workloads. Particularly under the impetus of open-source large models like Llama 3 and Qwen2, deployments of edge AI servers and hybrid-cloud inference nodes have surged—far outpacing IDM capacity planning for Intel Xeon 6 and AMD EPYC 9004 series processors. Supply-chain data shows lead times for premium server CPUs now exceed 32 weeks, with spot premiums reaching 15% for select SKUs. This CPU shortage fundamentally stems from evolving AI workload structures: GPUs handle model training, whereas CPUs manage critical tasks—including data preprocessing, API orchestration, security encryption, and multimodal fusion—making them indispensable partners in the “golden duo.” Domestically, companies including Cambricon and Hygon Information are accelerating the launch of custom CPU + AI accelerator solutions tailored for inference workloads, while memory interface chipmakers (e.g., Montage Technology) and DDR5 module suppliers (e.g., G-CHEN) benefit concurrently from the broader server memory bandwidth upgrade wave.
Space-Based Computing Policy Implementation: Extending the Compute Frontier from Ground Clusters to Low Earth Orbit
The Space-Based Computing Enterprise Symposium held in Beijing Yizhuang on June 1 marks the entry of China’s first regional space-based computing infrastructure initiative into the implementation phase. Distinct from conventional satellite communications, space-based computing emphasizes real-time onboard processing of massive data streams—including remote sensing imagery, IoT telemetry, and high-frequency financial transactions—thereby minimizing costly, latency-prone downlink transmission to ground stations. For example, a low-Earth-orbit (LEO) satellite equipped with an AI inference chip can identify crop pests and diseases across thousands of acres in seconds—without needing to transmit raw imagery back to terrestrial data centers. Yizhuang’s planned Innovation Center will prioritize core technologies including satellite-mounted heterogeneous computing architectures, high-speed inter-satellite and satellite-to-ground laser links (≥25 Gbps), and onboard model lightweighting and compression. It will also collaborate with commercial aerospace firms—including GalaxySpace and SpaceTime DaoYu—to build a closed-loop ecosystem encompassing “onboard training → inter-satellite coordination → ground distribution.” At the policy level, the National Space Infrastructure Mid- and Long-Term Development Plan explicitly designates “intelligent satellite networks” as a key national project; LEO constellation launch frequency is projected to reach a historical peak in 2026. This domain delivers direct upside to satellite internet terminal chips (e.g., ZORO’s RF front-end), satellite-mounted phased-array antennas (e.g., Chengchang Technology), and radiation-hardened domestic FPGAs (e.g., Unisplendour Microelectronics).
Full-Stack Synergy Logic Solidified: From “Compute Anxiety” to “Infrastructure Dividend”
These three concurrent developments collectively validate a broader trend: AI infrastructure investment has transcended single-hardware dimensions and entered the era of systems engineering. CPO solves “interconnect efficiency,” CPUs ensure “orchestration intelligence,” and space-based computing expands the “geographic frontier”—together forming a novel digital foundation for the AGI era. Capital markets have responded swiftly and decisively: over the past month, the optical module index rose 23%; server CPU-related stocks gained an average of 18%; and assets under management (AUM) in commercial aerospace ETFs grew 41% month-on-month. Yet short-term overheating risks warrant caution—the recent suspension of new domestic account openings by Tiger Securities, though driven by cross-border regulatory compliance adjustments, reflects tightening liquidity pressures. Looking ahead, the true long-term beneficiaries will be enterprises possessing cross-layer technical integration capabilities: communications giants capable of delivering both CPO optical engines and satellite-to-ground protocol stacks—or tech conglomerates simultaneously building CPU ecosystems and orbital AI frameworks. When computing power is no longer confined to data center server racks—but extends into ocean depths, mountain peaks, and low Earth orbit—the very definition of “new infrastructure” has been fundamentally rewritten: it is no longer a physical edifice of steel and concrete, but an intelligent neural network woven from photons, electrons, and orbital trajectories.