I.AI Computing Power Expansion Opens Growth Space for Copper Global computing power infrastructure and data center construction have seen simultaneous explosive growth, with intensive commissioning of intelligent computing and supercomputing projects across regions, generating an entirely new incremental demand curve for copper semis. According to SMM projections, global new installations is expected to achieve a CAGR of 24% from 2025 to 2030, with the fastest pace of new deployment occurring in 2025 and 2026. New installations in 2026 are expected to grow 65% MoM, and by 2027, the growth rate of new installations is projected to pull back to 28.77%, followed by a year-by-year deceleration in 2028-2030. By region, global new installations of computing power are mainly concentrated in two major markets: the US and China. Leveraging its scale-leading cloud operators, highly efficient facility operation systems, and a well-established global AI industry ecosystem, the US continues to lead in deployment scale. In China, leading cloud producers such as Alibaba and Tencent continue to increase capital expenditure on computing power infrastructure, while the national computing power network is formally incorporated into the top-level planning of the "Six Networks" and the "East Data, West Computing" projects are being rolled out and commissioned in batches, leading to a steady rise in the market share of domestic intelligent computing centers. SMM analysis indicates that the CAGR of new copper consumption for global computing power from 2025 to 2030 is 21%, slightly lower than the growth rate of new installations. The core reason is the gradual release of medium and long-term technological effects that reduce copper usage. Looking at individual years, copper consumption growth is 54.94% in 2026, pulling back to 27.58% in 2027, and the growth of new copper consumption is also expected to exhibit a gradual slowdown trend from 2028 to 2030. II. Unit Copper Consumption in Computing Power Centers Shows a Phased Trend of First Increasing then Decreasing SMM's calculation by region shows that the comprehensive copper consumption per unit of global computing power centers will present a characteristic of first increasing then decreasing from 2025 to 2030. In the short term, new computing power is mainly through the construction of entirely new campuses, with supporting power, grounding, and other infrastructure built from scratch, coupled with high-density cabinets driving a rapid rise in the penetration rate of liquid cooling systems. Multiple factors jointly push comprehensive unit consumption upwards in 2025-2026. In the medium and long term, as 800V high-voltage DC power distribution is popularized at scale, the required thickness and cross-section of copper conductors under equivalent power scenarios will decrease significantly.Meanwhile, high-speed NVLink copper cables will face substitution by fiber optic interconnects. Coupled with the iteration of liquid cooling heat dissipation materials and technological breakthroughs in aluminum as a substitute for copper processes, the industry's comprehensive unit consumption will enter a downward trajectory. However, constrained by the pace of industrial technology penetration, SMM's calculations show no significant decline in unit consumption in 2027-2028, as consumption reduction and copper-increase factors offset each other, keeping unit consumption stable. The downward trend will only become significantly prominent after 2029. It is worth noting that comprehensive unit consumption is a weighted average calculated by SMM based on the deployment scale of computing power in the US, China, and the rest of the world. There is clear differentiation in unit consumption among data centers in different regions, with the unit copper consumption ranked as: Rest of the World > China > US, where the scale effect of large clusters effectively lowers unit copper intensity. III. Breakdown of Core Copper Usage in Computing Power Centers Computing centers fall into three categories: general-purpose IDCs, intelligent computing centers, and supercomputing centers. This article uses intelligent computing centers, which currently hold the highest market share and fastest growth rate, as the calculation sample to break down the copper consumption structure: The power supply and distribution system is the largest copper-consuming segment in a computing center, accounting for 66% of total copper consumption according to SMM calculations. It primarily handles medium- and high-voltage power conversion and ensures uninterrupted power supply for equipment rooms. Medium- and low-voltage distribution cabinets, UPS, and busways are the core copper-consuming equipment. In the short term, high-power cabinets continue to boost demand for copper semis in power distribution, while in the medium and long term, after the popularization of lithium battery UPS and high-voltage DC solutions, unit copper consumption in the distribution segment will trend steadily downward. SMM estimates that AI server hardware infrastructure accounts for 18% of copper consumption, undertaking all tasks of computing power, storage, and network interaction. It integrates core components such as GPUs, motherboards, and server power supplies, and the stable operation of the hardware directly determines the computing power output of the cluster. High-end AI server PCBs and internal interconnection copper wires are the main sources of copper consumption in this segment. The liquid cooling system accounts for 11% of copper consumption. The closed-loop liquid cooling cycle meets the heat dissipation demands of high-power AI chips, with cold plates, CDU heat exchange units, and circulating copper pipe & tube serving as the main copper-consuming components. Liquid cooling penetration during 2025-2026 will boost demand for copper pipe & tube and copper plate/sheet and strip, and once copper-aluminum composite heat dissipation materials mature, the copper intensity for heat dissipation will gradually decline. Network communication, grounding protection, and supporting auxiliary systems together occupy the remaining 5% of copper consumption , covering sub-scenarios such as high-speed interconnection cabling and equipment room grounding and lightning protection copper grids. The current optical fiber interconnection industry chain continues to expand production, with optical fiber enterprises seeing simultaneous improvements in orders and profitability, indirectly confirming the overall high prosperity of AI computing power construction. In the long term, optical fiber will also continue to divert demand from high-speed copper cables. IV. Comprehensive Analysis of the Proportions of Different Copper Semis Breaking down the copper consumption structure of computing centers comprehensively by semi-finished copper product category, cables and copper busbars are the core consumables throughout the construction process. SMM analysis shows that cables account for 40% of the total copper consumption in a computing center, acting as the “blood vessels” permeating every link, with core applications in high-voltage access, low-voltage distribution, power transmission, high-speed communication copper cables, building wiring, as well as grounding and lightning protection cables. Copper busbar (24% of total copper consumption), the "backbone" for high-current power distribution in data centers, is mainly used in high- and low-voltage power distribution cabinets, transformer copper busbars, UPS systems, etc.Copper plate/sheet and strip (17%) is mostly used in transformer windings and liquid cooling cold plate substrates, performing dual functions of power transformation and heat dissipation. Copper pipe & tube (11% of total copper consumption) is a dedicated consumable for liquid cooling systems, mostly used in circulation piping, CDU heat exchange units, and precision air conditioning heat exchange pipes. The large-scale expansion of liquid cooling will boost demand for copper pipe & tube in the short term. Copper foil (4%) covers application scenarios including servers, switches, and various PCB circuit boards. Industry demand is concentrated on HVLP ultra-low profile high-end copper foil. Although the copper consumption per GW is relatively small, the incremental elasticity driven by AI computing power expansion is extremely strong. Currently, copper foil enterprises are accelerating the switch of capacity from ordinary electronic copper foil to high-end HVLP capacity, while copper clad laminate (CCL) producers have full order books and processing fees are being raised continuously, indicating that the prosperity of computing hardware demand has been verified across the entire industry chain. In summary, the rapid expansion of computing centers directly drives the growth in demand for related copper semis. At the same time, high-density AI computing clusters significantly raise the requirements for power supply supporting facilities, and the overall electricity consumption scale of the industry surges simultaneously. The demand for power infrastructure construction derived from computing expansion has become a key focus for long-term tracking and research in the future. While computing demand expands, the industry's development also faces external constraints. Currently, the grid connection approval process has a relatively long queuing period, and the market is generally concerned that transmission, distribution, and generation-side capacity bottlenecks may drag down the implementation pace of computing projects. However, according to SMM forecasts, no substantial power supply gap risk is expected for the industry over the next five years. It will still be necessary to closely track the approval progress and commissioning pace of various transmission and distribution supporting projects. SMM will also continue to follow the relevant industry dynamics and copper demand changes. For detailed data, please contact Cynthia Wang of the SMM Copper Research Team at 15762822325.

South Korea’s LS Group recently saw its market value rise to around KRW 64 trillion, driven by growing demand for AI-related power infrastructure. Its subsidiary LS Cable & System, a major global high-voltage cable supplier, continues to benefit from investment in AI data centres, grid upgrades, and energy infrastructure. Public information shows orders for high-voltage transformers, cables, and ESS continue to grow across North America and Asia, with particularly strong expansion in Southeast Asia. Sales from subsidiaries in Vietnam and Indonesia have also accelerated. As AI data centre construction expands, market attention remains on copper demand linked to cables, power equipment, and grid upgrades.

SMM, April 22: The global new energy and AI industries maintained high prosperity, coupled with traditional consumer electronics entering a seasonal peak, driving a comprehensive recovery in demand for copper foil and high-speed connectivity products. In March 2026, China's copper foil exports hit a single-month record high. The power and ESS sectors maintained high prosperity with steadily growing orders; AI-related orders remained robust, with optimistic demand for copper foil across all specifications, and the operating rate of copper foil enterprises rose both YoY and MoM in March. Supported by these fundamentals, copper cable high-speed connectivity, as a core component of AI computing infrastructure, continued to see improving demand expectations. Combined with resonating capital sentiment in the sector, copper cable high-speed connectivity concept stocks strengthened overall on April 22. As of the close on the 22nd, the copper cable high-speed connectivity concept rose 2.29%. Among individual stocks, Far East Holdings hit the daily limit, while Hengtong Optic-Electric, ZTT, Shenlan Technology, Changxin Bochuang, Zhaolonginterconnect, and Kingsinno led the gains. News [State Council: Advancing Computing Power Layout and Edge Computing Power Development in an Orderly Manner, Improving Intelligent Computing Cloud Service System] The State Council issued the "Opinions on Promoting the Expansion and Quality Improvement of the Service Industry." It mentioned deepening the implementation of the Industrial Internet innovation and development project, advancing the industrial data foundation initiative, cultivating data cooperation consortia, and building a number of high-quality industry datasets. It also called for developing professional services such as data annotation and certification, exploring the establishment of classified and graded mechanisms for data rights confirmation, evaluation, and pricing. Computing power layout and edge computing power development will be advanced in an orderly manner, and the intelligent computing cloud service system will be improved. The application of urban information modeling platforms and building information modeling technologies will be accelerated. [China's Intelligent Computing Power Scale Reaches 1882 EFLOPS] Zhang Yunming, Vice Minister of the Ministry of Industry and Information Technology, stated on the 21st that computing power infrastructure has become a key foundation driving the development of artificial intelligence. As of month-end in March, China's intelligent computing power scale reached 1882 EFLOPS. At a press conference held by the State Council Information Office on the same day, Zhang Yunming said that recently, the Ministry of Industry and Information Technology has been making sustained efforts in key areas such as enabling small and medium-sized enterprises with inclusive computing power and computing-electricity coordination, and the computing power industry has shown a positive development trend. Industrial innovation has become stronger, with the in-depth implementation of the computing power foundation "open competition" initiative, the conduct of policy research and standard formulation for computing-electricity coordination, and the promotion of accelerated implementation of generation-grid-load-storage and green electricity direct connection. Network transmission has become smoother, with over 70 computing power corridors built around computing power hubs over the past two years, and the implementation of the metropolitan "millisecond computing" special initiative, among others. [National Advanced Computing Industry Innovation Center Qingdao Base Officially Launched] According to Sugon, on April 22, the National Advanced Computing Industry Innovation Center Qingdao Base was officially launched in Laoshan District, Qingdao. The National Advanced Computing Industry Innovation Center is a national-level industrial innovation platform established in 2018 with the approval of the National Development and Reform Commission (NDRC), aimed at achieving breakthroughs in core technologies in the field of advanced computing and fostering internationally competitive industrial clusters. The Qingdao base launched this time, co-built with the participation of Sugon, will leverage regional industrial advantages to create an advanced computing innovation platform integrating technology R&D, commercialization of research outcomes, and industrial incubation, further pooling industry resources, driving regional industrial intelligent upgrading, and injecting strong "computing power momentum" into the digital economy development of Qingdao and even Shandong Province. [Yuandong Holdings: Q1 2026 Net Profit of 96.6284 Million Yuan, up 110.36% YoY] Yuandong Holdings disclosed its Q1 2026 report on April 22. The company achieved total operating revenue of 5.325 billion yuan, up 9.26% YoY; net profit attributable to the parent company was 96.6284 million yuan, up 110.36% YoY. [Rickda: Achieved Small-Batch Delivery of 400G/800G Products and Is Driving Capacity Ramp-Up] Rickda released an investor relations activity record announcement. The company is a quality supplier capable of simultaneously providing comprehensive solutions for optical, electrical, microwave, high-speed data, and fluid connections. It has also deployed multiple products for AI applications, including power supply, high-speed transmission, and liquid cooling categories. High-speed copper cable connector-related products are primarily led by Suzhou Richuang Connection Technology Co., Ltd., with planned products mainly including AEC, DAC, and ACC high-speed copper cable products, high-speed board-to-board connectors, high-speed I/O connectors, and other product series. Currently, Richuang has achieved small-batch delivery of 400G/800G products and is driving capacity ramp-up, while multiple products with even higher speeds are also progressing as planned. [Tongguan Copper Foil: The Company's High-Frequency High-Speed PCB Copper Foil Can Be Widely Used in 5G Communication Equipment, High-Computing-Power Servers, and Other Network Equipment and Network Connectors] Tongguan Copper Foil stated on the investor interaction platform on March 23 that the company's high-frequency high-speed PCB copper foil can be widely used in 5G communication equipment, high-computing-power servers, data centers, switches, and other network equipment and network connectors. [Xinhongye: Dedicated Technical Research Has Been Conducted in the High-Speed Copper Cable Connection Field for Intelligent Computing Scenarios] Xinhongye stated on the interaction platform on March 17 that the company pays close attention to the development trends of computing power data centers and the computing-power-and-electricity synergy industry, and will leverage its core business and technological advantages to continuously track industry dynamics and prudently evaluate opportunities for penetration and deployment in related fields. The current AI intelligent computing industry is trending toward the parallel development and synergy of optical and copper solutions. Leveraging its core copper cable technology advantages, the company has launched dedicated technical research in the high-speed copper cable connectivity domain for intelligent computing scenarios. Meanwhile, in line with the industry's technological evolution, the company maintains forward-looking technical research and market tracking in fiber optic-related fields, and is expected to steadily advance related deployments based on industry demand, technological development, and its own strategic planning. AI-Related Orders Remain Red-Hot with Optimistic Demand across All Copper Foil Specifications According to SMM, the operating rate of Chinese copper foil enterprises rose both YoY and MoM in March 2026. Downstream demand recovered rapidly after Chinese New Year, with both production and sales of lithium battery copper foil and electronic circuit copper foil climbing in March, bringing the overall industry operating rate close to 90%. End-use demand side, the power and ESS sectors maintained high prosperity with steady order growth. The traditional consumer electronics industry transitioned smoothly into peak season, AI-related orders remained red-hot, and demand across all copper foil specifications performed optimistically . Global New Energy and AI Industries Show High Prosperity — March Copper Foil Exports Hit a Single-Month Record High According to data from the General Administration of Customs, China's copper foil (HS codes: 74101100, 74102190) imports in March 2026 totaled 8,220.04 mt, up 11.88% YoY and up 27.59% MoM. China's copper foil exports in March 2026 totaled 6,663.48 mt, up 56.19% YoY and up 38.23% MoM. Statistics showed that both China's copper foil imports and exports rose to elevated levels in March 2026, with exports hitting a single-month record high. Global new energy and AI industries exhibited high prosperity, coupled with optimistic demand during the seasonal consumption peak. Voices from Various Parties A Huaxi Securities research report stated that external geopolitical risks persist at the current juncture, and market risk appetite may affect short-term market fluctuations. AI remains a key investment theme in the near term, with attention on domestically produced computing power and computing power leasing driven by supply-demand imbalance, as well as optical module and fiber optic cable segments with strong earnings certainty. Computing power supply-demand imbalance is intensifying, with computing power consumption expanding sharply and the three major cloud providers collectively raising prices. On one hand, the current supply-demand imbalance reflects robust demand in upstream chips and computing power leasing industries, with enhanced bargaining power driving price increases, benefiting domestic AI chip and computing power leasing producers. On the other hand, rising AI service costs will be transmitted to downstream application development, potentially accelerating downstream industry consolidation and improving computing power utilization efficiency. Changjiang Securities believed that geopolitical conflicts were accelerating the restructuring of global supply chains, enhancing China's irreplaceability as the "world's factory," and that related export chain and manufacturing leading assets may command a "scarcity premium." In terms of allocation, regardless of whether the Middle East conflict fluctuated, three main themes should be firmly pursued. First, the AI trend that conflicts could hardly alter — focusing on AI infrastructure such as computing power, storage capacity, and power equipment, as well as "HALO" asset opportunities. Second, the urgency of conflicts accelerating the energy revolution — focusing on new energy sectors such as lithium battery and hydrogen energy, as well as resource commodities including non-ferrous metals, oil, and coal. Third, sectors at cyclical bottoms where earnings were expected to gradually improve, such as chemicals, steel, and condiments. Data showed that the CSI A500 Index, closely tracked by A500 ETF South China, covered high-quality large- and mid-cap leading enterprises in A-shares, evenly distributed across core areas such as emerging manufacturing and consumption upgrading, with a focus on new quality productive forces. Investors could gain exposure through A500 ETF South China and its feeder funds in one step. A Huatai Securities research report believed that, with the rapid increase in demand for 800G and 1.6T optical modules in recent years and the approaching 3.2T era, the development opportunities for upstream core materials of optical modules were promising. It systematically reviewed the growth logic of two major industries — InP substrates and thin-film lithium niobate. InP substrates, as upstream core raw materials for optical chips, benefited from the rapid boost in demand from optical chip producers, with the industry showing an undersupply trend. Modulators made from thin-film lithium niobate, leveraging advantages such as low power consumption and high bandwidth, were expected to enter an adoption window in 3.2T pluggable solutions in the future, with broad growth potential across the industry chain. A Shanxi Securities research report indicated that NVIDIA made a $2 billion strategic investment in Marvell to cooperate on expanding the NVLINKFusion ecosystem. On March 31 local time, NVIDIA and Marvell announced a strategic partnership to connect Marvell with NVIDIA's AI factory and AI-RAN ecosystem through NVIDIA NVLINKFusion. Marvell would provide customized XPUs and NVLINKFusion-compatible expansion networks, while NVIDIA would provide supporting technologies, including Vera CPU, ConnectX NICs, Bluefield DPU, NVLINK interconnects, and SpectrumX switches, as well as rack-scale AI computing. Marvell is one of the world's leading ASIC custom service providers, with clients including AWS, Microsoft, and Google. It is a major global supplier of optical module DSP and EIC, and has focused on CPO deployment through the acquisition of CelestialAI. We believe that NVIDIA's strategic investment in Marvell is expected to enhance Marvell's design capabilities in memory semantics, high-speed SERDES, and super-node systems, and to facilitate the expansion of the NVLINKFusion ecosystem. NVIDIA's NVLINK Scaleup technology encompasses an overall solution covering NVLINK SERDES, NVLINK chiplets, NVLINK switches, and rack-scale expansion architecture (including NVLINK SPINE, copper cabling systems, innovative mechanical architecture, power supply, and liquid cooling technologies). NVIDIA unveiled a complete copper-connected rack-scale solution at GTC 2026, including Rubin NVL72, Rubin Ultra NVL144, LPX 256, and ETL 256. Shanxi Securities believed that Marvell's participation is expected to expand the addressable market for copper connectivity. NVIDIA has positioned CPO as one of the most important transformations in Feynman-generation Scaleup technology, and combined with Marvell's silicon photonics technology, we believe the CPO penetration rate is expected to gradually increase. Soochow Securities' research report commenting on Fujida noted: a leading player in China's RF connector industry. The company specializes in the R&D, sales, and services of connectors, cables, cable assemblies, and microwave components, holding a leading position in China and non-China markets. Since its establishment in 1998, the company has been deeply engaged in the RF interconnect field. Backed by AVIC Optronics (its controlling shareholder), it has gradually expanded from traditional general-purpose RF connectors to high-end new product categories, including RF cables (aerospace applications), advanced ceramic products (chip integrated packaging applications), and RF links (active and passive microwave components). In 2025, the company achieved revenue of 881 million yuan and net profit attributable to the parent company of 78 million yuan, up 15.5% and 52.0% YoY, respectively. The company has positioned itself in five core tracks, with broad prospects for high-end connectors: 1) Demand in the defense informatization sector remained strong, with new products showing considerable potential. 2) The civil aerospace satellite sector has become a new growth engine. 3) The semiconductor equipment industry urgently needs high-end RF/electrical connectors. 4) The domestic supply chain for high-end electronic measurement instruments is embracing opportunities. 5) High-speed copper cables and quantum communication cable products are benefiting from the data center construction boom. Large-scale growth in data center infrastructure is boosting high-speed transmission products to gradually evolve toward system-level solutions. The company's high-speed copper cables have achieved product category expansion around 400G components, reaching an internationally leading level; meanwhile, in the quantum communication field, the company has deployed cryogenic superconducting cables, achieved initial small-batch supply, and is expected to gradually achieve commercialization. Recommended Reading:

On April 9, at the , hosted by SMM Information & Technology Co., Ltd. (SMM), Shandong Aisi Information Technology Co., Ltd., and SMM Trading Center Co., Ltd., and co-organized by Shandong Humon Smelting Co., Ltd., Zambia Development Agency (ZDA), Chalco Luoyang Copper Processing Co., Ltd., and Hetian Commerce and Logistics Group Co., Ltd., Wu Jinkai, Head of the Metals Team at Sinolink Securities Co., Ltd., delivered a presentation on the topic "Computing Power – Electricity – Copper: Repricing the 'New Infrastructure Metal' in the AI Era." 1. From "Traditional Infrastructure Metal" to "AI New Infrastructure Metal": The Shifting Role of Copper The Shifting Role of Copper: From Supporting Material to Systemic Variable • Traditional infrastructure phase: In the era of conventional real estate, manufacturing, and legacy power grids, copper was largely viewed as a supporting metal that followed demand expansion; it was important, but rarely became a core variable driving infrastructure investment narratives. • AI new infrastructure phase: AI data centers are not simply about purchasing more servers, but about rebuilding high-density load infrastructure: as GPU counts rise, they simultaneously boost demand for racks, cooling, power distribution, substations, T&D, and green electricity integration, making copper a systemic variable. • Implications of the role shift: The marginal pricing anchor for copper is gradually migrating from demand recovery in traditional infrastructure to whether computing power deployment materializes and whether capital expenditure across the power chain steps up to a new level; this means the market should no longer understand AI-driven copper demand solely through a traditional commodity framework. Why the Old Consensus Underestimated AI-Driven Copper Demand: Overly Conservative Base Assumptions • Commonality 1: In the past, mainstream research preferred to start from publicly disclosed projects/GW or annual construction volumes. The advantage was clear and verifiable metrics, but the drawback was that it easily overlooked expansions of existing parks, undisclosed projects, and spillover copper consumption extending from racks to the power grid. • Commonality 2: When the market did not believe AI deployment would materialize at scale, research naturally opted for more defensive parameters: only recognizing project pipelines, only counting the data center facilities themselves, only accounting for capital expenditure already incurred, and refusing to price in downstream supporting infrastructure. • Commonality 3: Therefore, the market's past underestimation of AI-driven copper demand was not essentially about "copper intensity being off by a few percentage points," but rather about overly conservative front-end assumptions; once GPU shipments demonstrated greater certainty, the model's starting point had to be revised upward across the board. The Starting Point of the New Consensus: Not Simply Revising Copper Intensity Upward, but Repricing the Guidance • Demand validation moving forward: Since 2026, agent applications represented by OpenClaw have gone mainstream, enabling the market to see that inference demand, token usage, and commercialization loops are migrating toward real deployment; this has caused the narrative that "AI has no demand" to lose its explanatory power. • Reassessment of supply guidance: Once demand-side momentum becomes visible, the aggressive guidance on NVIDIA shipments and TSMC's advanced packaging expansion should begin to be priced in, prompting a reassessment of previous copper demand estimates. • The real expectations gap: Therefore, the dividing line between old and new consensus lies not in 39 t/MW versus 45 t/MW, but in whether the market is willing to price in that expansion will truly materialize in racks, campuses, and the power grid; as soon as the answer shifts from disbelief to belief, copper demand will undergo a systematic reassessment. 2. How AI Data Centers Use Copper: Beyond Racks, Extending to the Power Grid US Data Center Power Consumption Trends: Power as an Inflationary Factor • The most widely cited report for AI data center estimates is the *2024 United States Data Center Energy Usage Report*, published in December 2024. We focus our discussion on this report. • Historical trends (2014–2023): 2014–2016: Power consumption remained stable at approximately 60 TWh per year, continuing the low-growth trend since 2010. • 2017 turning point: As server installations grew—particularly with GPU-accelerated servers for AI accounting for a significantly larger share of the data center server fleet—data center power consumption began to rebound; consumption reached approximately 76 TWh in 2018, accounting for 1.9% of total annual US electricity consumption. • 2018–2023: Growth accelerated, with power consumption reaching 176 TWh in 2023, accounting for 4.4% of total US electricity consumption, representing an 18% CAGR from 2018 to 2023. • Future scenario projections (2024–2028): 2028 projected range: Power consumption ranges from a low of 325 TWh to a high of 580 TWh. Assuming an average capacity utilization rate of 50%, this corresponds to total data center power demand of 74–132 GW, accounting for 6.7%–12.0% of projected total US electricity consumption in 2028, with a CAGR of 13%–27% expected from 2023 to 2028. • From the perspective of assumptions: Upper bound: Based on the IDC 2024b report, assuming sustained AI activity, GPU shipments continuing at H2 2024 growth rates, and manufacturers being able to meet demand; Lower bound: Based on the IDC 2023a report, assuming AI activity enthusiasm pulls back, with GPU shipment growth reverting to pre-2024 historical average levels (e.g., 70%–80% of the 2021–2023 growth rate). Currently, using the upper bound as the basis for estimation appears more reasonable. • The report's estimates are based on AI 8-GPU racks, whereas current racks are NVL72, containing 72 cards. Theoretically, power consumption should be 9 times that of AI 8-GPU racks, but actual NVL 72 rack power consumption is approximately 15 times higher, indicating severe power inflation on the cooling side. • Considering GPU upgrades and increased power consumption, we estimate electricity consumption will reach at least 800 TWH by 2028. Calculation results: US power grid copper consumption to increase by nearly 2.1 million mt compared to 2025 by 2030 • AI's boost to copper demand is primarily driven through electricity consumption. US electricity consumption in 2023 was 4,000 billion kWh. According to the aforementioned Berkeley Lab, data center electricity consumption in 2023 was 176 billion kWh, accounting for 4.4% of total US electricity consumption. Based on our revised forecast, by 2028, electricity consumption is expected to range from a low of 580 billion kWh to a high of 800 billion kWh, contributing incremental electricity consumption of 404–624 billion kWh, accounting for 10%–15.6% of incremental electricity consumption in total. • Under this scenario, we calculated the copper and aluminum demand for the US power grid. By 2030, the boost to copper and aluminum from data centers, manufacturing reshoring, and new energy is expected to increase by 2.1 million mt and 3.71 million mt respectively compared to 2025. From a product structure perspective, copper demand is primarily boosted by wires & cables and transformers, while aluminum is primarily boosted by wires & cables and substations. How AI Data Centers Use Copper: Three-Layer Pathway • Three-layer framework: Before discussing AI copper usage, it is essential to first clarify the scope: in-rack/near-rack primarily covers servers and proximity networks; out-of-rack but on-site primarily covers power distribution and cooling; further out covers substations, T&D, and green electricity connections supporting incremental loads. • Most common misconception: The market tends to equate "server BOM" with "total copper consumption of AI data centers"; however, for high-density training clusters, in-rack accounts for only a small portion, with the real bulk lying in on-site power chain and cooling. • Research requirements: All subsequent calculations must simultaneously answer three questions: whether the scope is in-rack or the entire campus, whether it is greenfield or expansion, and whether it covers only the facility itself or also incorporates the external power chain. Only then can different reports be truly comparable. 39 mt/MW Is Not "Rack BOM" but the Full Electrical System Intensity of AI Training Data Centers • Intensity meaning: The 39 mt/MW figure from S&P does not mean "39 mt of copper installed in a single rack," but rather represents the comprehensive intensity under the direct copper scope for AI training hyperscale data centers; high-redundancy designs in China can even reach 47 mt/MW. • Structural meaning: Breaking down by sub-item median values, the power chain accounts for approximately 61% of direct copper, cooling approximately 22%, and server + network only 17%; in other words, most copper is not in the GPU itself but in the systems that supply power to and dissipate heat from the GPU. • Investment meaning: This is also why we later switch from "mt/rack" back to "mt/MW": as long as power density continues to rise, power chain copper consumption will scale accordingly; even if fiber optics replace some copper cables, it is only a partial offset and does not change the overall logic. 120kW cabinet: greenfield 7.8 mt, expansion 6 mt, infill 4.25 mt •Using the NVIDIA NVL72 ~120kW cabinet as a reference, based on the greenfield new-build average, the total system copper consumption per cabinet is approximately 7.8 mt; this scope includes on-site direct usage and a rough allocation of off-site power-side consumption. •However, current mainstream AI deployments do not always start from scratch at every campus. More commonly, expansion and partial reuse of existing grid connections, main power distribution, and cooling trunk lines occur within existing campuses. Therefore, the expansion scope is better represented by 5.5–6.5 mt/cabinet, with a midpoint of approximately 6 mt. •For pure infill scenarios, where existing infrastructure is extensively reused, copper consumption per cabinet can be further reduced to 3.5–5 mt/cabinet. Core model: deriving copper demand directly from GPU count •This version of the core model no longer works backward from aggregate figures, but instead models sequentially from "GPU count → cabinet count → power → electricity consumption → copper consumption." This approach directly translates advanced packaging, cabinet power, and expansion pace into copper demand. •Baseline assumptions: NVIDIA 7 million GPUs and Google 5.5 million GPUs in 2026; GPU counts for both companies growing at +50% per year from 2027 to 2030. Cabinet assumptions are NVIDIA 72 GPUs/120kW and Google 64 GPUs per cabinet; the deployment scope uses a campus expansion midpoint of 6 mt/cabinet. •Under this framework, "how much copper per 1 million high-end AI GPUs corresponds to" becomes a repeatable and updatable paradigm, rather than an abstract macro discussion. NVIDIA: the first independent main thread •2026 starting point: under the base case, NVIDIA is projected to have approximately 7 million GPUs in 2026. Converting at 72 GPUs/120kW NVL72-equivalent cabinets, this roughly corresponds to 97,000 cabinets, 11.7 GW of IT load, and approximately 14 GW of facility load; even looking at just one company, the scale is already very substantial. •Copper consumption path: under the campus expansion scope, NVIDIA alone has a midpoint annual copper demand of approximately 580,000 mt in 2026; if deployment continues at a 50% annual growth rate from 2027 to 2030, this could rise to approximately 2.95 million mt/year by 2030 for this single company. •Research implication: the most important significance of these figures is not their precision down to the last digit, but rather that they demonstrate: as long as the market believes the guidance will be delivered, NVIDIA alone is sufficient to elevate AI-driven copper demand into a balance-sheet-level variable, rather than merely a thematic investment. Google: the second independent main thread •Not a supplementary item: in our base case, Google is not a "supplementary item" but rather the second independent main thread. Based on 5.5 million GPUs in 2026 at 64 GPUs/cabinet, this translates to approximately 86,000 cabinets, 9.2 GW of IT load, and 11 GW of facility load — a scale already not far from NVIDIA's first-year deployment. • Copper consumption path: Since we applied "per-card copper consumption equivalence" under the expanded capacity framework, Google's annual copper demand midpoint in 2026 was approximately 460,000 mt; if similarly advancing at a 50% annual growth rate, by 2030 a single company would correspond to approximately 2.32 million mt/year of copper demand. • Conclusion significance: Therefore, Google should not be viewed as a minor adjustment outside of NVIDIA; as long as hyperscaler in-house ASIC roadmaps also expand in parallel, AI-driven copper demand is not a single GPU leader narrative, but a narrative where multiple computing power routes collectively push up power capex. Two companies combined: ~5.27 million mt midpoint by 2030 at 50% annual growth rate • Combined result: After merging NVIDIA and Google, under the 2026 midpoint scenario, annual copper demand already reached approximately 1.04 million mt; if growing at 50% annually thereafter, the midpoint could reach approximately 5.27 million mt by 2030, with the low-to-high range roughly at 4.64–5.94 million mt. • Release pace: More importantly, the release pace: this is not a one-time story of "using up all the copper in the first year," but rather a front-low-back-high deployment curve; as card counts, rack counts, and MW levels rise in tandem, the slope of annual incremental copper demand will steepen progressively. • Market significance: Precisely because of this, if the market still uses the old framework of "a few hundred thousand mt" to understand AI-driven copper demand, it can easily underestimate the non-linearity in the mid-to-late stages; for a commodity like copper with a tight balance sheet, what truly matters is often not the first year, but the slope after the third year. 3. AI High-Frequency Data Continues to Deliver: Not Just a Narrative, but Reality Copper Monthly Physical Flow Tracking: China • In Feb 2026, China's apparent demand declined YoY, -10% YoY. Feb China copper cathode production was 1.09 million mt, down 100,000 mt YoY, with imports remaining at low levels. For the first 2 months, apparent demand fell 131,000 mt YoY, of which domestic demand declined 324,000 mt and external demand added 303,000 mt. Global demand for the first 2 months of 2026 was estimated to have declined 19,000 mt, with domestic demand falling significantly both YoY and MoM, while external demand support slowed down. • On fabricated product exports, auto and transformer export growth maintained high YoY increases. On a MoM basis, only transformer exports remained at elevated levels, while exports in other segments all declined MoM, with wire and cable exports declining notably MoM. Fabricated product exports fell 1,000 mt in the first 2 months, indicating weak copper fabricated product export demand. Copper Monthly Physical Flow Tracking: US • In Oct, US apparent demand was -12% YoY, with cumulative first 10 months at +22% YoY, adding 412,000 mt of new demand. Considering the solid performance of power sector demand, an estimated 200,000 mt of the 374,000 mt in new demand was actual demand growth, with cumulative hidden inventory of 212,000 mt. Combined with Oct COMEX inventory of 340,000 mt, total hidden + visible inventory in the US region was estimated at 552,000 mt. • In October, regarding US net imports of fabricated products, wires & cables, transformers, and computers were the main contributors to incremental growth, while auto net imports continued to decline. Power equipment (wires & cables + transformers) accounted for 128% of the incremental demand relative to total import demand, with autos being a significant drag. AI's boost to copper usage in power grids continued to materialize, as already reflected in high-frequency data. Wires & cables plus transformers combined added 281,000 mt, with annualized US power-related copper imports estimated at approximately 350,000 mt in 2025. • The decline in power equipment imports was likely attributable to two factors: 1) The rush to import and restock in H1 ended, with weakened urgency to front-load imports in H2; 2) Rising domestic market share in the US — after the 2025 tariffs, US domestic companies saw notable increases in sales volume, with AKTR/HUBB/POWL showing significant rises, expected to substitute for imports. Combining the above analysis, with domestic demand increase plus incremental imports, US copper demand for power grids was no less than 400,000 mt. Seven US Tech Giants Sign Self-Supply Power Commitment • Representatives of seven companies — Microsoft, Google, OpenAI, Amazon, Meta, xAI, and Oracle — signed relevant documents at the White House. US President Trump stated that many Americans were concerned that data centers would push up power demand and potentially raise electricity bills, but this document would resolve the issue. Meanwhile, the bearer of tariff costs also became clearer — the AI giants themselves. • Previously, there were two key questions about US power grid construction: utilities lacked the capability to build; and utilities lacked the willingness to build. With the signing of this document, the pathway for US power grid construction has been cleared. The seven giants have both the capability and willingness to invest in power supply construction, and self-built power plants would not affect copper demand deployment. High voltage/extra-high voltage (HV/EHV): responsible for long-distance power transmission and large-scale backbone grids; typical equipment: power transmission lines, main substations (step-up/step-down), and large switching stations. • Medium voltage (MV): responsible for campus-level power distribution, delivering electricity to each zone; typical equipment: distribution stations, ring main units/switchgear, MV cables, and distribution transformers (MV→LV). • Low voltage (LV): responsible for the last segment within server rooms, delivering electricity to loads; typical equipment: LV switchgear, busways/cables, PDUs, UPS (mostly on the LV side), and server power supplies. • Following the signing of this document, we expect demand for wires & cables to accelerate. Copper Monthly Physical Flow Tracking: Europe • In October, European apparent demand was -4% YoY. Fabricated product imports for the first 10 months were +63% YoY, with October alone surging +48% YoY. The core incremental growth came from wires & cables, while the share of auto exports increased. •In the first 10 months, new demand from European wire & cable and transformers totaled 140,000 mt, with full-year new power grid demand in Europe estimated at 160,000-180,000 mt. 4. Risk Warnings Risk Warnings •AI demand falls short of expectations •Aluminum as a substitute for copper exceeds expectations

[Xinya Electronics: High-Frequency and High-Speed Copper Cable Assembly Enters Supply Chain of Renowned Server Manufacturers such as Oracle via Amphenol] Xinya Electronics stated on the interactive platform that the company's high-frequency and high-speed copper cable assembly has entered the supply chain of renowned server manufacturers including Dell, HP, Google, Amazon, Microsoft, Oracle, Meta, Inspur, H3C, and Sugon via Amphenol. As of the first half of 2025, this product generated revenue of 93.8752 million yuan, up 87.28% YoY, accounting for 4.83% of the company's total revenue.

Despite extreme supply-side volatility, U.S. copper demand in August–September showed resilience. Manufacturing overall remained soft, but sectors such as power infrastructure, data centers, and electric vehicles (EVs) continued to support copper use.