One of Wall Street's favorite stories right now is the copper supercycle. The logic is simple: AI needs massive data centers, data centers need tons of copper, so copper demand will keep climbing. It's a neat narrative that has fueled a rally in mining stocks and commodity ETFs.

But reality, as always, is messier. Rising capital costs, declining ore grades, and a shortage of new discoveries are squeezing supply. On the demand side, the story is also more complicated than a straight line up. While S&P Global forecasts copper demand from data centers and related infrastructure could rise from 1.1 million metric tons in 2025 to 2.5 million by 2040, those projections often assume a relatively static technological environment. In practice, AI infrastructure is evolving rapidly, shifting toward less copper-intensive solutions.

Two Key Narrative Flaws

The bullish argument starts with a valid observation: AI facilities are incredibly metal-intensive. According to S&P Global, a crypto data center requires roughly 21 metric tons of copper per megawatt of installed capacity, while an AI training center can require as much as 47 metric tons. The difference reflects denser wiring, larger cooling systems, and dramatically higher power requirements.

However, the gap between announcements and actual deployment is considerable. The concept of "bragawatts" is increasingly present in 2026, as capacity figures look impressive in investor presentations yet face a substantial reality check in practice. Reuters reported that Ireland imposed a de facto moratorium on new data centers after the sector ballooned to more than 20% of the national electricity demand. Policy interference and infrastructure constraints also limit the thesis of exponential expansion for copper.

The second flaw in the copper supercycle narrative is that engineers are actively trying to reduce metal usage. AI economics reward efficiency. NVIDIA Corp. (NVDA) and other industry leaders are redesigning power architectures to move more energy with less material.

NVIDIA's solution is a transition to 800 volts, where the same wire gauge can transmit around 157% more power. With fewer conductors, smaller connectors, and simpler designs, performance would be better even with less copper per rack.

Huang's Pragmatic Transition

"You use optics wherever you must, [and] you use copper wherever you can," NVIDIA's CEO Jensen Huang said at a trade show in Taipei, according to the WSJ.

Contrary to claims that copper will suddenly become obsolete, Huang argues for a hybrid model. Copper remains a practical, inexpensive, and proven technology within certain distance and bandwidth limits. Existing infrastructure will continue relying heavily on it for years.

He noted that copper will continue to play an important role in infrastructure, even as the use of optical devices expands rapidly in the coming years.

This coexistence model creates opportunities for firms positioned across both ecosystems. Huang has repeatedly highlighted Marvell Technology, Inc. (MRVL) as a key beneficiary because it supplies networking technologies that bridge traditional copper-based systems and advanced optical interconnects.

The Copper Wall at 400 Gbps

Copper's transmission distance is inversely proportional to bandwidth. Today's leading 200 Gbps-per-channel copper systems start to lose signal after only about 2.5 meters. Since modern server racks are about 2 meters tall, the margin is already becoming rather small.

Matt Murphy, CEO of Marvell, said copper cables will be unable to support complete rack connectivity as systems migrate to 400 Gbps speeds.

As bandwidth requirements continue climbing, electrical signals weaken faster and generate more heat. At some point, physics overwhelms economics.

Fiber optics and silicon photonics have become necessary rather than optional because they enable the transmission of data over far greater distances with minimal degradation.

Eventually, it doesn't mean that the copper thesis is dead. One way or another, AI will increase copper demand. Still, the assumption of an unlimited copper supercycle ignores the realities of grid constraints, deployment delays, and relentless engineering innovation.