How AI Development Is Reshaping Commodity Markets

Data centres currently account for roughly 1–2% of global electricity consumption. By 2030, depending on the pace of AI deployment, estimates from the IEA and Goldman Sachs suggest that figure could reach 3–4% — with AI inference workloads as the primary driver. Unlike traditional enterprise computing, which runs on predictable schedules, AI inference is always-on: every ChatGPT query, every API call, every real-time recommendation system draws power continuously.

The implication for natural gas is significant. Renewable capacity is being added rapidly, but it cannot be dispatched on demand. When a hyperscaler needs 500MW of guaranteed baseload power for a new data centre campus, gas turbines are the answer — not solar panels. This is already showing up in US power market data: several grid operators have reversed previous forecasts of flat electricity demand and now project sustained growth through the end of the decade.

Nuclear power is experiencing a related renaissance. Microsoft signed a deal to restart Three Mile Island. Google has contracted for small modular reactor capacity. Amazon has acquired a data centre campus directly adjacent to a nuclear plant. Uranium, after a decade of bear market following Fukushima, has entered a structural bull cycle driven in part by AI's insatiable appetite for clean, firm, baseload power.

A hyperscale data centre is an extraordinarily copper-intensive structure. Power distribution systems, server racks, cooling infrastructure, fibre runs, and the transmission lines that connect the facility to the grid all require copper. Estimates suggest a large data centre campus (100–500MW) requires 20–40 tonnes of copper per megawatt of capacity — comparable to the copper intensity of offshore wind. With hundreds of new data centre campuses announced globally through 2027, the incremental copper demand from AI infrastructure alone could represent several hundred thousand tonnes per year.

Aluminium follows a similar logic. Liquid cooling systems — increasingly preferred over air cooling for high-density AI chips — use aluminium heat exchangers and manifolds. Transmission infrastructure upgrades, required to connect new data centre campuses to the grid, are heavily aluminium-intensive in overhead lines. And the chassis of every server rack, every network switch, every UPS system is aluminium.

The timing creates a problem: copper mine supply is structurally constrained. No major new mine has come online in the past decade without significant delays and cost overruns. Chilean grades are declining. The pipeline of projects that could add meaningful supply before 2030 is thin. AI-driven demand is arriving into a market that was already structurally tight due to the energy transition — the two demand shocks are compounding.

AI accelerators — Nvidia GPUs, Google TPUs, custom ASICs — require a specific set of critical minerals at scale. High-purity silicon for wafers. Cobalt and tantalum for capacitors. Indium for displays and transistors. Gallium and germanium for compound semiconductors. These are not commodities in the traditional sense — they do not trade on exchanges with transparent price discovery — but their supply dynamics are increasingly central to AI development timelines.

China controls refining capacity for many of these materials at levels that dwarf its share of raw ore production. It refines roughly 80% of the world's gallium, 60% of germanium, and dominates rare earth processing globally. When China imposed export controls on gallium and germanium in 2023 — framed as a response to US chip export restrictions — it demonstrated that the critical mineral supply chain is a geopolitical pressure point, not just a manufacturing input.

For commodity analysts, the actionable insight is that AI chip demand creates a new class of concentrated supply risk. Unlike copper or oil, where disruptions are priced into liquid futures markets, disruptions in gallium or high-purity quartz (used in chip furnaces) are invisible until they manifest as production delays — at which point the impact is transmitted indirectly through tech earnings and capex cuts.

Data centres are thirsty. Evaporative cooling towers — the dominant cooling method for large facilities — consume between 1 and 5 litres of water per kilowatt-hour of electricity consumed. A 100MW data centre running continuously can consume 1–5 billion litres of water per year. Microsoft's data centres consumed 1.7 billion litres of water in 2021; by 2022 that figure had grown to 6.4 billion litres, driven largely by AI workloads from ChatGPT training and inference.

This is creating direct conflicts with agricultural and municipal water users in regions where data centres are concentrated — Northern Virginia, Phoenix, central Iowa, the Netherlands. Water utilities in several of these areas have begun reporting increased industrial demand from hyperscalers as a material factor in their capacity planning. In drought-prone regions, regulators are starting to impose water use restrictions on new data centre permitting.

Water is not a commodity that trades globally, but it is a constraint that increasingly affects where AI infrastructure can be built, how quickly it can be permitted, and what its operating costs will be. For investors in agricultural commodities, the competition for water between AI data centres and irrigation in arid farming regions is an underappreciated long-term risk.