The Hidden Materials Layer Beneath the AI Infrastructure Trade

Rare earth supply chain concentration adds a materials-science dimension to AI infrastructure investing — and exposure varies sharply across the stack.

Prepared by Richstorm.co

KEY TAKEAWAYS

• Rare earth materials show up at multiple points in the AI infrastructure stack — chips, chipmaking equipment, and power/cooling systems — not just one layer.

• Dependency varies significantly across the stack: semiconductor manufacturers and equipment makers carry the most direct materials dependency, power and cooling infrastructure the least.

• Tungsten, used in advanced-node chip interconnects and manufacturing equipment, has no practical substitute at scale and deserves as much attention as the more commonly discussed lanthanide elements.

• Building rare earth processing capacity outside existing hubs is underway (MP Materials, allied-nation initiatives) but is expected to take years to meaningfully diversify the supply base.

• Spreading AI infrastructure exposure across the stack — including power and cooling names — is one way to reduce concentration risk tied to any single layer.

Our companion piece, “Inside the Rare Earth Supply Chain,” covers the broader market structure — why processing (not mining) is the concentrated step, and the global diversification effort underway. This piece assumes that background and goes straight to a specific application: how rare earth materials show up across the AI infrastructure stack, and why exposure differs so much company to company.

Where Rare Earth Materials Actually Show Up in AI Infrastructure

Rare earth materials appear at several distinct points in the AI buildout — not just one layer. They're in the chips themselves, in the equipment used to manufacture those chips, and in the power and cooling systems that keep data centers running. The table below maps the elements most relevant to AI infrastructure specifically.

Processing concentration figures reflect current industry estimates and may shift as new capacity comes online globally.

Tungsten Deserves More Attention Than It Gets

Most coverage of rare earth supply chains focuses on the lanthanide elements — neodymium, dysprosium, terbium. Tungsten gets less attention but is arguably just as structurally important for advanced semiconductor manufacturing. Every advanced-node chip below 7 nanometers uses tungsten in its contact and via structures. Every EUV lithography machine and several other fab tools use tungsten in high-temperature components. At scale, there's currently no practical substitute, which makes tungsten's supply concentration worth tracking alongside the more commonly discussed elements.

Company Dependency Varies by Layer of the Stack

The materials-concentration factor doesn't affect every AI infrastructure company equally. It's most direct for companies manufacturing chips and chipmaking equipment, and least direct for companies further from the physical fabrication process.

Dependency levels are a qualitative assessment based on where each company sits in the physical supply chain, not a prediction of specific outcomes.

TSMC and ASML: Two Different Kinds of Structural Importance

TSMC sources a meaningful share of its consumable materials for advanced-node production (roughly 30% for chips at 7nm and below) from processing hubs that are concentrated in a small number of locations globally. These aren't easily substituted inputs — the purity specifications and process integration involved have been built up and qualified over many years, so building alternative supply relationships takes real time and capital, not a quick swap.

ASML occupies a different kind of structural position: it's the only global manufacturer of the extreme ultraviolet lithography machines needed for advanced-node chipmaking, and those machines include rare-earth-based lasers and optical components. Because TSMC, Samsung, Intel, and every other advanced fab depends on ASML's equipment, any disruption to ASML's own supply chain would ripple across the entire industry simultaneously — a structural characteristic of the semiconductor equipment market generally, independent of any single input.

Why Power and Cooling Infrastructure Looks Different

Vertiv, GE Vernova, and Eaton — the power and cooling companies supporting AI data centers — do have rare earth dependencies, mainly through permanent magnets in motors, fans, and power equipment. But this dependency differs from semiconductor manufacturing in two useful ways.

First, industrial power equipment doesn't require the same purity and precision specifications as advanced chip fabrication, so alternative sources and substitute materials are more available at this level of the stack. Second, the underlying demand driver — AI data center power and cooling needs — is structural and largely independent of any single materials constraint: data centers need power and cooling regardless of what happens elsewhere in the chip supply chain. That makes power infrastructure revenue somewhat more insulated from this particular factor than pure semiconductor revenue, which is worth factoring into how AI infrastructure exposure gets allocated across a portfolio.

Building Diversified Supply Chains Takes Time

Rare earth deposits exist outside of any single country — in Australia, the US, Brazil, and elsewhere — but the specialized processing infrastructure that converts ore into semiconductor-grade material is concentrated in relatively few facilities. Building that capability elsewhere requires specialized chemical engineering infrastructure, environmental permitting, workforce development, and years of process optimization — the same buildout timeline discussed in our companion piece.

MP Materials in California is the most advanced Western rare earth mining operation, though its processing capability still lags behind established facilities for some of the highest-purity applications. The US CHIPS Act and Inflation Reduction Act, along with initiatives in Japan, South Korea, the Netherlands, and Australia, are directing meaningful capital toward diversified semiconductor and materials supply chains — genuine progress, though on a timeline measured in years, not months. Intel's Ohio fab complex, for instance, isn't expected to reach full production until the late 2020s, and materials processing capacity generally takes longer still to scale than fab construction itself.

Portfolio Implications

None of this argues against AI infrastructure exposure — the demand growth underlying the theme is real and durable. It argues for being deliberate about where in the stack that exposure sits, since dependency on concentrated materials processing varies meaningfully by company type.

A qualitative framework for thinking about materials-concentration exposure across AI infrastructure investment categories.

Three Practical Adjustments Worth Considering

First, within semiconductor exposure, weigh the relative position of fabless designers with strong pricing power — Nvidia in particular — against manufacturers with more direct materials-process dependencies. Nvidia's scale and customer priority may offer more flexibility in managing supply variability than manufacturers further down the physical supply chain.

Second, consider power and cooling infrastructure — Vertiv, GE Vernova, Eaton — as a genuine complement to semiconductor-heavy AI infrastructure exposure, not just an adjacent theme. These companies capture the same structural AI demand tailwind with a meaningfully different materials-dependency profile.

Third, be aware that broad semiconductor ETFs (SMH, SOXX) concentrate heavily in the highest-dependency names by design — they diversify across companies within the sector, but not across this particular structural factor. Pairing that exposure with power infrastructure names is one straightforward way to diversify along this dimension specifically.

The Takeaway

The 'shovel sellers win the gold rush' framing for AI infrastructure holds up well overall — but not every shovel seller carries the same supply chain profile. Chip manufacturers and equipment makers sit closer to a concentrated materials layer than power and cooling infrastructure companies do, and that's a genuinely useful distinction for building a diversified AI infrastructure allocation, not a reason to avoid the theme altogether. Understanding where in the physical supply chain a company sits — from raw materials, through processing and fabrication, to the finished data center — is the kind of grounded, structural detail that a materials-science lens adds to an otherwise financially-driven investment thesis.