Chip Shortages, Foundry Strain — and Why Copper Will Be the Hidden Bottleneck
Share
1) The supply shock: fabs are stretched thin
The concentrated nature of advanced manufacturing — a handful of fabs (TSMC, Samsung Foundry, Intel) that produce most advanced logic — is now colliding with explosive demand from cloud AI, autonomous vehicles and robotics. In 2025 coverage and company reports show TSMC’s growth rate easing (yet still large), Samsung’s foundry business encountering headwinds and reports that leading foundries are running near capacity.
At Tesla’s November 2025 shareholder meeting, Elon Musk publicly said Tesla is likely to need a “gigantic chip fab” (and is exploring partnerships, e.g., Intel) because existing fabs will not supply the volumes Tesla expects for Cybercab, Optimus and in-car AI. Multiple outlets reported Musk’s “Terafab / Gigafab” comments and his concern that the current supplier base cannot meet future demand.
Put bluntly: if demand keeps doubling for AI accelerators and automotive SoCs, wafer starts and packaging capacity become a hard physical constraint — not just a logistics problem.
2) Where copper already sits in the chip supply chain (and why it matters)
Copper is used across the semiconductor value chain in three distinct buckets:
- On-chip interconnects (BEOL — back-end-of-line): modern ICs use copper metallization for interconnect lines inside the die because copper lowers resistance vs. aluminum and enables higher performance. This use is technologically critical (though the mass of copper in the silicon die itself is tiny).
- Packaging, substrates and PCBs (module level): chip packages, substrate interposers, GPU and accelerator PCBs contain measurable copper in multiple layers. Printed circuit boards commonly use copper weights measured in oz/ft² (1 oz/ft² ≈ 28.35 g copper per ft² per layer); typical multilayer boards and heavy-current power traces drive copper mass into the tens of grams per module. Studies of e-waste and PCB composition show copper can be ~20–30% of a PCB’s metal content.
- Fab & data-center infrastructure: the physical fab (power distribution busbars, transformers, chillers, pumps, cooling loops, process equipment) and the data centers that host training jobs use large quantities of copper — hundreds of kilograms to tonnes per facility. Industry reports and copper factbooks document that industrial users (including electronics manufacture and heavy electrical equipment) are major copper consumers.
Bottom line: although copper mass inside a single silicon die is small, the packaging + PCB + plant + power/infrastructure uses dominate real copper demand tied to chip scale-up.
3) Typical copper quantities — reference points (sources + assumptions)
- PCB copper mass: a 1 oz/ft² copper layer = ~28.35 g copper per ft²; multilayer boards add up. Typical complex PCBs used on GPU accelerators and server cards commonly add up to tens of grams of copper per board, and heavy power distribution traces push that higher. (PCB manufacturing references and copper weight tables).
- PCB metal composition: PCs / phone PCB studies report copper as ~20–30% of PCB metal weight (useful for extrapolating copper per kg of PCB).
- On-die copper: copper interconnects are essential but represent milligram → gram scale copper per die (very small mass compared with packaging/PCB). Technical reviews of copper interconnects document the process and benefits.
- Fab & infra copper: major equipment (transformers, busbars, chillers, cabling) is copper-intensive; industry copper factbooks and fabrication studies document large industrial copper consumption (detailed tonnages vary by fab size).
4) A worked example: estimating chips needed and implied copper demand (transparent assumptions)
I’ll show a transparent, conservative calculation method you can reuse. These are scenario estimates; I list assumptions so you can tweak them.
Assumptions (example / conservative)
- Tesla (vehicles + Optimus):
Vehicles: 2,000,000 vehicles/year × 10 years = 20M vehicles.
AI chips per vehicle (autonomy + redundancy): 3 chips (SoC/accelerator + backup).
Optimus robots: assume 1,000,000 robots over 10 years × 2 chips each = 2M chips.
Total Tesla chips (10 yrs) = (20M × 3) + 2M = 62M chips.
- NVIDIA / Hyperscalers (GPUs/AI accelerators):
Conservatively assume 10 million AI accelerator modules (data-center / edge) over 10 years (this is illustrative; actual demand could be higher).
X.AI & other AI firms: combined additional demand ~1–3 million chips over 10 years (smaller players).
Copper-per-chip assumptions (by module type)
- Automotive SoC module (packaged board + connectors): ~30 g Cu (PCB layers, small power traces, small copper plating). — conservative estimate based on PCB copper weights and small board area.
- High-end AI accelerator / GPU module (full card + power PCB + connectors + some heatsink piping): ~100 g Cu (multilayer board area, heavy current traces, power connectors; heatsink pipes/copper components) — mid-range engineering estimate; real cards vary widely. (See PCB copper weight refs and packaging discussion.)
Example math (conservative scenario)
- Tesla chips copper (vehicles): 20M vehicles × 3 chips × 30 g = 1,800,000,000 g = 1,800 tonnes Cu
- Tesla robots copper: 2M robots × 2 chips × 30 g = 120 tonnes Cu
Total Tesla (10 yrs) ≈ 1,920 tonnes Cu
- NVIDIA / datacenter GPUs: 10M accelerator modules × 100 g = 1,000,000,000 g = 1,000 tonnes Cu
- X.AI / others: 2M × 100 g = 200 tonnes Cu
Aggregate (example): Tesla + datacenter + other = ~3,120 tonnes Cu over 10 years (conservative illustrative calculation).
Important: this does not include copper used by fabs and data-center infrastructure (busbars, generators, chillers, process equipment), which can add hundreds to thousands of tonnes per new fab or large data center. Adding one large gigafab or megafab could easily require many thousands of tonnes of copper when you count power distribution, transformers, heavy cabling, chillers and site infrastructure. World copper factbooks and regional fabrication studies document industrial copper tonnages at that scale.
5) Interpretation & implications for traders / infrastructure planners
- Per-chip copper is small but scale multiplies: a few dozen to a few hundred grams per module becomes thousands of tonnes once multiplied by millions of units.
- If major vertically integrated firms (Tesla, X.AI or hyperscalers) move to build fabs, copper demand jumps dramatically — both from packaging/boards and from fab + plant infrastructure. Reuters/Tom’s Hardware coverage of Musk’s Terafab comments highlights the scale of the ambition (and the implied industrial footprint).
- Foundry constraints already push up lead times & capex: TSMC and Samsung utilization signals mean customers may need to vertically integrate or secure long-term supply contracts. That in turn raises demand for copper across the stack (boards → manufacturing plants → data centers).
6) Takeaway (actionable summary)
- Short-term: expect continued tightness in advanced chips and longer lead times. That tightness increases demand for copper in packaging and boards.
- Mid/long term: if big players build fabs or major new data centers, copper demand will jump both for equipment (kg per chip module) and infrastructure (tonnes per fab/data center). Traders should model both streams separately: (A) component copper (tens–hundreds g/chip) and (B) infrastructure copper (tonnes/facility).
- For a conservative planning example, supplying chips for one large automaker’s 10-year rollout (tens of millions of chips) can require low-thousands of tonnes of copper just for modules — plus much more if the company builds fabs or data centers.