BofA: Top 10 Material Bottlenecks in AI Supply Chain as Narrative Shifts to Raw Materials
N.R. Finch
Bank of America identifies ten foundational material bottlenecks across the AI supply chain, arguing that the binding constraint on compute delivery has shifted from chips to low-profile raw materials — helium, high-purity quartz, gallium and others — whose supply elasticity now sets the hidden ceiling on AI's expansion pace.
Why are raw materials — not chips — suddenly AI's bottleneck?
BofA says the AI supply-chain narrative is shifting from "whose model is better" to "who can actually deliver the compute."
The binding factors are no longer just GPUs, HBM and advanced nodes — they include a set of basic materials with low unit value but extreme systemic importance.
This means → when these materials run short, the pain does not show up as a simple price spike; it shows up as longer lead times, slower capacity ramps and lower equipment utilization across the entire AI supply chain.
In plain terms = chips are the star players, but if the stadium's turf, lighting and plumbing break down, the game still cannot be played.
What keeps a fab "always on"? Two gases hold the lifeline
Helium is used for wafer-fab cooling, purging and leak detection — essentially the "uptime layer" that keeps production lines running. The US and Qatar together supply roughly 70% of global helium; after the Qatar hub lost about one-third of global supply due to Middle East conflict disruptions, Russia imposed temporary export controls in 2026, raising allocation risk further.
This means → fabs hold several months of inventory in the short term, but if tightness persists, high-margin AI chips get priority while consumer-electronics chips are deferred, and lead-time extensions ripple outward.
Neon is the core gas for DUV lithography lasers — BofA calls it "the off switch for the lithography tool." Neon is extracted from large air-separation units tied to the industrial-gas and steel complex, not to semiconductor demand cycles.
Leading fabs can recycle over 70%, but qualifying a new supplier or gas source typically takes 6 to 18 months — supply cannot ramp in sync when AI demand surges.
From wafers to packaging — which materials are gating delivery?
High-purity quartz is used to make crucibles — the vessels for growing single-crystal silicon ingots — and sits at the very top of every AI accelerator's supply chain. Most semiconductor-grade ultra-high-purity natural quartz comes from the Spruce Pine region of North Carolina; Hurricane Helene flooding in 2024 already disrupted mining and logistics there.
In plain terms = nearly every AI chip in the world traces its "first drop of silicon" back to one small town — if that town has a problem, the entire chain waits.
ABF substrate (Ajinomoto Build-up Film) — an insulating dielectric used in advanced packaging to connect chips to circuit boards — is another chokepoint. A single high-end GPU consumes 10 to 18 times the ABF area of a mainstream PC processor, and Ajinomoto holds over 90% market share.
BofA projects AI will account for 50% of ABF demand by 2027, with shortfalls of 3% in 2026 and 10% in 2027; new capacity takes 12 to 24 months from construction to qualification.
Power density is surging — who is under pressure?
Gallium is used in gallium-nitride and other compound semiconductors — materials that convert electrical power more efficiently and cut heat loss in AI servers. BofA's semiconductor team projects the GaN market in AI will grow at a 69% CAGR to $1.6 billion by 2030. China supplied roughly 98% of global primary gallium in 2024 and has tightened export controls in stages since July 2023; as of May 2025, Rotterdam gallium prices are up over 150% from pre-control levels.
Tungsten, with a melting point of roughly 3,400 °C, is used in semiconductor interconnects and contact structures where stability under high heat and current density is critical. China accounts for about 80% of global tungsten mine output; tungsten prices have risen 438% since early 2025 to $1,775 per metric ton of tungsten oxide, driven by export restrictions, tariffs and defense demand.
This means → both gallium and tungsten are heavily dependent on Chinese supply and are already being used as policy instruments — AI is entering an environment where it competes with defense for critical materials.
What is the "invisible dependency" in cooling and motion control?
Rare-earth permanent magnets — especially NdFeB (neodymium-iron-boron) — are used widely in server fans, cooling pumps and lithography-tool motion control. The denser the GPU cluster, the higher the thermal load, and the greater the reliance on fans, pumps and liquid-cooling systems.
NdFeB magnets account for roughly 50% of rare-earth demand by volume but over 95% by value. This reflects a core position in the rare-earth system far exceeding their physical share.
China dominates light and heavy rare-earth oxide processing, metallization and magnet manufacturing; the dysprosium and terbium heavy-rare-earth segments are almost entirely China-controlled.
What is BofA's core takeaway?
As AI capital expenditure enters large-scale deployment, the first pressure points to surface are typically not the highest-profile chip segments but the foundational materials that carry no standalone valuation and no dedicated analyst model.
This means → investors tracking the AI supply chain need to extend their view from GPUs and HBM to gases, minerals, films and magnetic materials — the "unvalued" links.
In plain terms = AI's expansion speed is ultimately set not by the fastest chip but by the slowest batch of raw materials.
Content is for reference only, not financial advice.