On April 9, at an event hosted by SMM Information & Technology Co. , Ltd. (SMM), Shandong Aisi Information Technology Co. , Ltd. , and SMM Trading Center Co. , Ltd. , and co-organized by several industry and governmental organizations including Shandong Humon Smelting Co. , Ltd. and Zambia Development Agency (ZDA), Wu Jinkai, Head of the Metals Team at Sinolink Securities Co. , Ltd. , presented on "Computing Power – Electricity – Copper: Repricing the 'New Infrastructure Metal' in the AI Era. " 1. The Evolving Role of Copper: From Traditional Infrastructure to AI-Driven Demand Copper’s role is shifting from a traditional infrastructure metal, seen mainly as a supporting material in real estate, manufacturing, and legacy power grids, to a systemic variable in AI-era infrastructure. AI data centers require not only servers but also a reconstruction of high-density load infrastructure—such as racks, cooling systems, power distribution, substations, transmission and distribution (T&D), and integration of green electricity—resulting in copper becoming central to infrastructure investment narratives. This shift means copper pricing now depends more on the deployment of computing power and capital expenditure across the power chain than on traditional infrastructure demand recovery. Previous market consensus underestimated AI-driven copper demand due to overly conservative assumptions—relying solely on publicly disclosed projects, ignoring expansions of existing data center parks, and excluding downstream infrastructure investments. As GPU shipments gain clarity, demand models must be revised upward comprehensively. The new consensus emphasizes not just increasing copper intensity but repricing the entire guidance based on visible AI-related demand growth, particularly from 2026 onwards, when AI applications such as OpenClaw validate inference and commercialization loops. Reassessment of supply, especially around NVIDIA shipments and TSMC packaging, further supports higher copper demand forecasts. The core debate is no longer small differences in copper intensity per megawatt (e. g. , 39 vs. 45 tons/MW) but whether market participants accept that AI-driven infrastructure expansion will materialize across racks, campuses, and grids. 2. Copper Use in AI Data Centers: Extending Beyond Racks to the Power Grid Based on the 2024 United States Data Center Energy Usage Report, US data center power consumption grew significantly from 60 TWh annually in 2014–2016 to 176 TWh in 2023 (4. 4% of total US electricity). This trend accelerated as GPU servers increased share, driving an 18% CAGR in power demand from 2018 to 2023. Projected 2024–2028 consumption ranges between 325 TWh and 580 TWh, implying total data center power demands of 74–132 GW or 6. 7%–12% of US electricity by 2028, growing at 13%–27% CAGR. These projections assume high sustained AI activity and continued GPU growth rates; the upper bound estimate is considered more valid. Current racks (e. g. , NVIDIA NVL72 with 72 GPUs) consume about 15 times more power than baseline AI 8-GPU racks mainly due to cooling demands, highlighting significant power inflation. Given GPU upgrades and increased rack power, electricity consumption is estimated to reach at least 800 TWh by 2028. Electricity consumption drives copper demand significantly: with US data center electricity expected to increase by 404–624 billion kWh by 2028 (accounting for 10%–15. 6% of incremental electricity), copper and aluminum demand related to power grids could rise by 2. 1 million mt and 3. 71 million mt respectively by 2030 compared to 2025. Copper demand growth centers on wires, cables, and transformers, while aluminum mainly supports wires, cables, and substations. Copper consumption in AI data centers follows a three-layer pathway: in-rack and near-rack systems (servers and networks), on-site but out-of-rack infrastructure (power distribution and cooling), and beyond-site substations, T&D, and grid integration. Contrary to common misconceptions equating data center copper use with server bill of materials, the majority of copper lies in on-site power and cooling infrastructure. The often-cited figure of 39 tons/MW copper intensity refers to the full direct copper usage across AI hyperscale AI training centers, including power chains and cooling; China designs may reach 47 tons/MW due to higher redundancy. This breaks down as approximately 61% copper in power chains, 22% in cooling, and 17% in servers and networks.

Increasing power density means power chain copper demand scales with compute infrastructure growth, making “mt/MW” a more meaningful metric than “mt/rack. ” For a 120kW NVIDIA NVL72 cabinet, greenfield deployments consume about 7. 8 mt copper per cabinet; expansion scenarios reuse existing infrastructure, lowering consumption to about 6 mt; infill scenarios further reduce consumption to roughly 4. 25 mt. A refined core demand model traces demand from GPU count to cabinet count, power, electricity consumption, and then copper consumption. Baseline assumptions include 7 million NVIDIA and 5. 5 million Google GPUs in 2026, growing 50% annually through 2030, with cabinet configurations of 72 GPUs/120kW (NVIDIA) and 64 GPUs (Google) and 6 mt copper per cabinet for expansions. This model allows repeatable, dynamic estimation of copper demand linked directly to GPU deployment. NVIDIA alone, at 7 million GPUs by 2026, corresponds to about 97, 000 cabinets, 11. 7 GW IT load, and 14 GW facility load, sustaining around 580, 000 mt copper demand in 2026, potentially rising to nearly 3 million mt by 2030 with 50% annual growth. Similarly, Google’s 5. 5 million GPUs yield approximately 460, 000 mt copper demand in 2026, scaling to 2. 3 million mt by 2030. Together by 2030, these two companies could generate a midpoint copper demand around 5. 27 million mt annually, within a range of 4. 64–5. 94 million mt. This demand exhibits a steep upward slope after the third year, challenging market views based on older, lower estimates and underscoring AI’s systemic impact on copper markets. 3. Evidence from High-Frequency Data Confirms the Trend In China, February 2026 saw a 10% year-over-year decline in apparent copper demand and reduced cathode production, with weakening fabricated product export demand, notably in wires and cables. Globally, early 2026 demand declined modestly but with marked domestic drops balanced somewhat by external demand. In the US, October 2025 recorded a 12% YoY drop in apparent demand but a cumulative 22% rise over ten months, reflecting significant net new demand (~412, 000 mt) mainly from power sector growth (~200, 000 mt actual demand plus inventory changes). Power-related products (wires, cables, transformers) accounted for 128% of incremental imports, while auto sector imports declined. Customs data suggest US power grid copper demand increased by at least 400, 000 mt, supported by rising domestic production replacing imports post-2025 tariffs. Seven major US tech firms—Microsoft, Google, OpenAI, Amazon, Meta, xAI, and Oracle—signed a White House document committing to self-supply power infrastructure. This commitment addresses prior concerns over utilities’ capacity and willingness to build necessary grid infrastructure, clearing the path for accelerated power grid development to support AI data centers. Campus power systems include high-voltage transmission, medium-voltage campus distribution, and low-voltage server room delivery systems; all will increase copper wire and cable demand. In Europe, October 2025 appeared demand fell 4% YoY, but fabricated product imports surged (~63% YoY in ten months, 48% YoY in October), mainly driven by wires and cables. New European power grid demand is estimated at 160, 000–180, 000 mt for the year. 4. Risks • AI demand falling short of expectations. • Aluminum substituting copper at higher than anticipated rates. In summary, AI-driven computing infrastructure significantly enhances copper demand beyond traditional frameworks, with data center power consumption growth driving substantial increases in copper use across servers, power distribution, cooling, and grid integration. This trend is validated by power consumption forecasts, refined copper intensity metrics, modeling tied to GPU deployment, and corroborating physical flow data in key regions. Market participants should reprice copper accordingly, mindful of risks from demand fluctuations and material substitution.