The next AI bottleneck

Key takeaways

• The AI investment opportunity is broadening beyond GPUs as investors look further down the supply chain for new bottlenecks.
• A surge in AI-driven demand has driven a “super cycle” in memory markets, with producers delivering exceptional share price performance.
• The buildout of AI infrastructure is increasing demand for commodities, including copper, specialty metals and base materials.
• Energy availability is emerging as a primary constraint, with data centre power requirements placing increasing strain on electricity grids.
• Years of underinvestment in mining and infrastructure, combined with accelerating demand from AI and the energy transition, are creating structural supply constraints and supporting higher commodity prices and miner profitability.

The AI trade broadens out down the supply chain

Two things have happened in May which I think speaks to a telling trend in the AI investment story: After hours on 20 May Nvidia reported quarterly revenue which was 85% higher than a year prior – the market shrugged off this news and, as at the time of writing, the shares are off -3% from results day; On the 26 May Micron Technology smashed through the $1 trillion market capitalisation barrier and is now the fastest company to ever go from $500bn to $1trn (71 days from mid-March to the end of May) – the memory company was worth just c.$70bn at the depths of liberation day chaos last year.

This is part of a broader “super cycle” in memory markets which my colleague Ben Hall covered in detail in a recent article. Memory chip producers such as Micron, SK Hynix and Samsung Electronics have seen their share prices rise between 400-950% in the last year alone (not a typo – SK Hynix is up nearly 10x in the last 12 months and over 20x in the last 3 years). This is the latest leg in the Artificial Intelligence investment story which began with the release of ChatGPT in November 2022 and now, as adoption becomes clearer and the promise of agentic capabilities requires ever increasing investment, market participants are starting to look further down the AI supply chain to identify supply bottlenecks and parts of the economy which could benefit from supranormal profitability. This has seen long-established memory producers become multi-baggers in the space of just 12 months – today we argue that as investors look even further down the supply chain they will run into another part of the economy which is crucial for the build out of AI – commodities.

Memory stocks total return (rebased to 100)

Source: Bloomberg, 30 May 2025 to 29 May 2026

Commodities are the ultimate bottleneck

The commodity complex includes the raw materials for the build out of data centres. These range from the specialty metals crucial for the production of high-performance chips, the copper required for the wiring and inter-connectivity within the chips and data centres and the basic materials such as steel and aluminium to build the physical infrastructure. The increased demand for AI capabilities has a direct increase on the demand for raw materials – every MW of data centre capacity requires at least 20 tonnes of copper. On top of this the strategic nature of Artificial Intelligence has exposed fragility in the supply chains of rarer but nonetheless crucial minerals e.g. China controls 98% of Gallium supply and 60-67% of Germanium refining.

However, after the sourcing of computer chips, the biggest bottleneck for the buildout of data centres is not the materials used to build them, but the energy required to power them. It is informative that the size of a data centre is not measured in square-footage or compute, but in Megawatts (MW) and, increasingly, Gigawatts (GW). In June 2025 Mark Zuckerberg’s Meta announced plans for a 5 GW data centre in Louisiana called Hyperion; 5GW is equivalent to the output of 5 nuclear reactors – enough output to power over 4 million homes continuously.

This demand for 24/7 baseload power is unprecedented. Data centre power demand is expected to rise by 165% by 2030 and account for 20% of global electricity demand growth this decade. At a local level this can be destabilising – in Dublin 80% of power consumption came from data centres and in 2021 a de facto moratorium on new projects was introduced by the grid operator. In an environment where AI deployment cycles are being measured in months (with multiple new and improved models released each year), the reality of 5-10 year grid upgrade cycles will become a bigger and bigger problem.

Hyperscalers are starting to take this problem into their own hands, and we’ve seen Microsoft announce a 20-year power purchase agreement to restart the nuclear energy facility at Three Mile Island and Meta engage with multiple partners to build a gigawatt-scale nuclear portfolio. Nuclear energy is particularly well suited to solving the both the AI energy problem and the energy transition given its ability to generate utility-scale baseload power. In this context Cameco, the largest western producer of uranium (c.8,000 tonnes p.a.) is well set up, particularly given its 49% stake in Westinghouse which provides nuclear design and buildout services and gives the company exposure to the upswing in nuclear development. The Real Assets Fund has a 2.3% position in the company.

Cameco total return (rebased to 100)

Source: Bloomberg, 30 May 2025 to 29 May 2026

AI is only the start (or end) of the problem

Prior to the AI thematic other parts of the economy were already having to deal with the under-investment in materials and mining sector.

The world is going through an energy transition which sees energy sources switching from fossil fuels to technologies such as solar and wind which bring with them significant challenges related to the intermittent nature of their energy production – grid systems have to fundamentally change to cope with the more localised, dispersed, non-baseload electricity generation.  

The challenges that the energy transition and data centre growth poses to the grid come at a time when the existing infrastructure was already creaking. As covered before, the average age of the installed transmission network in the US is 40 years old with 25% of those copper cables over 50 years old. 100,000 miles of cable will need replacing by the end of the next decade and infrastructure services firms such as Quanta Services (2.2% of the Real Assets Fund) stand to benefit from this investment spend.

15 years of under-investment

So, with the biggest global megatrends, from AI to Energy Transition, reliant on the sourcing and refining of key raw materials one would hope that there is ample supply of these key elements. Unfortunately this is not the case. Since the commodity super-cycle prior to the 2008 Global Financial Crisis the sector has been starved for capital in an environment which has favoured intangible assets and seen the rise of the hyperscaler companies which dominate equity markets today. This lack of investment has also translated into a real slowdown of mining activity – management teams across the mining industry have been scarred by the lessons of 2008 and have shifted to much more conservative capital allocation policies (i.e. returning cash flow to shareholders via dividends or buybacks as opposed to reinvestment). A recent report from Berstein highlighted that even through the first half of 2025 when the mining sector had seen EBITDA grow +35% in aggregate, capex spend was flat versus the same period in the prior year; the miners remain disciplined.

MSCI USA Sector % Weight: Tech vs. Resource Sectors

Source: Factset as at 31 March 2026

The practical output of this is that supply of raw materials remains constrained at a point when demand is inflecting upwards. The IEA is projecting that copper supply shrinks from c.25Mt today to less than 20Mt in 2035. Given the lead time on new mines is 10-20 years, the result is that there could be a supply deficit of 30% within the next 10 years.

However, in a free market supply must equal demand (accounting for inventory changes and recycling) with the balancing variable being the commodity price, having to rise to a price which leads to demand destruction. In a world where governments, hyperscalers and utilities are scrambling to meet the needs of AI models and renewable energy development there is significant upside risk across the commodity complex.

To that end the Real Assets Fund has positions across crucial raw materials including energy, uranium, copper, PGMs, gold and critical minerals. Including both exposure to mining and materials companies and the physical commodities themselves, over 25% of the fund is exposed to a commodity complex which is well set up to benefit from significant tailwinds in the economy.

Glossary

Hyperscalers
Large technology companies that build and operate data centres at scale. They are increasingly investing directly in energy and infrastructure to support AI growth.

Baseload power
A continuous, always-on supply of electricity required to run data centres. AI infrastructure depends on stable 24/7 power rather than intermittent energy sources.

Critical minerals
Rarer but essential materials used in advanced technology. Their supply chains are often concentrated and fragile.

Super-cycle
A period of exceptionally strong growth in demand and pricing, as seen recently in memory chips driven by AI adoption.