The Metal Hiding in Plain Sight: Why Copper Miners Could Be Your Best AI Trade

Copper and COPX ETF Vulcan analysis as part of the GNG Research team . February 3, 2026. While everyone chases semiconductor stocks, the real AI infrastructure bottleneck is piling up in warehouses from Shanghai to Chicago. The $50 Million Problem Nobody’s Talking About In a keynote explanation of the GB200 NVL72 design, Jensen Huang said the NVLink spine uses roughly 5,000 cables totaling about 2 miles, and that going optical would have added roughly 20 kW of transceiver and retimer power draw. Not fiber optics. Copper. The same metal that still does the heavy lifting in power and cooling hardware worldwide. I’ve spent 15 years watching investors chase shiny objects while ignoring the obvious plays. And right now, the AI trade has become so obsessed with chips that it’s completely missing the infrastructure bottleneck forming underneath. A conventional data center might use 5,000 to 15,000 tons of copper (CSIS estimate). A hyperscale AI facility running NVIDIA’s latest hardware? Try 50,000 tons per site (CSIS estimate). That’s not a typo. That’s a 10x jump in copper intensity, and it’s happening across every major tech campus being built. My read: the copper miners aren’t just a derivative play on AI. They might be the trade with the clearest risk/reward in the entire sector. Why NVIDIA Couldn’t Use Fiber (And What That Means for Your Portfolio) Here’s the part that surprised me. When NVIDIA engineers designed the GB200 system, they faced a brutal tradeoff. They could use fiber optic connections between GPUs and keep the heat down. Or they could use copper and deal with the thermal nightmare. They chose copper. Why? Because switching to fiber would have added another 20 kilowatts of power draw just from the transceivers and retimers. Think about that for a second. NVIDIA, the company building the most sophisticated AI hardware on the planet, looked at every material science option available and concluded that copper was irreplaceable for their highest performance systems. The metal that humans have been mining for 10,000 years is still the optimal solution for connecting AI processors in 2026. This isn’t just an NVIDIA story either. Every AI accelerator generates intense heat in a tiny space, and copper remains one of the best practical conductors for power delivery and heat transfer at scale. The power delivery systems require thick copper busbars to carry hundreds of megawatts across campus. Even the grounding systems protecting billions in AI hardware rely on copper. When I pull up the supply data, the math gets uncomfortable. AI data center construction in North America hit 6.3 gigawatts of capacity under development in 2024. Industry estimates suggest AI facilities alone could push data center copper demand to more than half a million metric tons annually by 2030 (Forbes). And that’s before you count EVs, grid upgrades, and everything else fighting for the same constrained supply. The 29-Year Problem That Nobody Can Fix Here’s where the thesis gets interesting. Unlike semiconductors, where capacity can scale relatively quickly, copper has a fundamental timing problem. It takes 17 to 29 years to permit and develop a new copper mine in Western countries. You read that right. The copper that will supply AI data centers in 2035 needed to start permitting before smartphones existed. The industry spent the 2010s underinvesting after commodity prices collapsed. Now ore grades are declining at existing mines, and the giant Chilean operations that supply much of the world are aging. S&P Global projects that even with mines under construction, global supply will peak by 2030 and fall behind demand, potentially creating a 10 million ton annual shortfall by 2040. BHP, the world’s largest mining company, has been screaming about this for months. Their CFO publicly stated that while data centers represent only 1% of copper demand today, they could reach 6 to 7% by 2050. They projected AI application