Two Layers Deeper: The AI Supply Chain's Second Chokepoint per Theme

Every large power transformer (LPT) above 100 megavolt-amperes (MVA — the standard unit for transformer rating) requires a laminated grain-oriented electrical steel (GOES) core. A typical unit at or above that threshold uses 50 to 150 tons of GOES.^[4]

For US-domicile transformer supply, Cleveland-Cliffs' Butler, Pennsylvania mill is the single named domestic GOES producer currently operating. Imports arrive from Nippon Steel, JFE Steel, POSCO, and Thyssenkrupp — four named non-US alternatives — but import logistics add lead time and regulatory exposure under any trade-restriction scenario. Steel Dynamics (STLD) announced a competing US GOES mill in 2026; commissioning is multiple years out.^[4]

"30-month transformer lead time" is structural, not a supply-chain glitch policy can fix quickly. The Defense Production Act (DPA) cannot expand GOES capacity on a relevant timeline — a greenfield GOES mill takes five or more years to commercial operation. The LPT lead-time crisis hyperscalers are now navigating is, at its floor, a GOES capacity constraint propagating upward.^{[5] [6]}

Under the buyer-constrained substitutability framework — where Federal Energy Regulatory Commission (FERC) LPT availability requirements, CHIPS Act-funded fab build-out, and hyperscaler "made in America" power purchase agreements create domicile-sourcing constraints — the effective concentration is elevated for a significant share of US demand.

Cleveland-Cliffs (CLF) is a roughly $8 billion small/mid-cap steel producer whose Butler PA mill represents the primary pure-play domestic GOES exposure, blended with broader steel-cycle economics. The chokepoint thesis is graphable; the single-asset dependence is the unusual feature.

An H-class combined-cycle gas turbine (CCGT) can ship in two to three years today. A 765-kilovolt large power transformer (LPT) can take four. That inversion is the structural point that the public framing misses when it focuses on turbine lead times.^[7]

Qualified suppliers with US-deliverable LPT capacity include Hitachi Energy (South Boston, Virginia plant, a Hitachi subsidiary operating as a standalone private entity), HD Hyundai Electric (Korea, approximately $26 billion), Hyosung Heavy (Memphis, Tennessee plant, approximately $20 billion), LS Electric (Korea), GE Vernova, Siemens Energy, Pennsylvania Transformer Technology, and Virginia Transformer Corp — at least seven named suppliers with US-deliverable presence. The structure is a qualified-supplier concentration with multi-year lead times, not a sole-source situation.

NVIDIA unveiled its 800-volt direct current (800V DC) datacenter architecture at GTC 2025 in March 2025. Hitachi Energy formally announced its NVIDIA collaboration in October 2025. Korean industry reporting (Asia Business Daily) named HD Hyundai Electric, LS Electric, and Hyosung Heavy as probable Korean partners NVIDIA had approached for transformer supply — partner identities have not been confirmed publicly by NVIDIA. Hyperscalers pre-position transformer capacity three to four years ahead of gigawatt-scale campus energization. Hitachi Energy's South Boston expansion, announced at $457 million, confirms that the capacity constraint is structural, not temporary.^{[7] [8] [9]}

Two of the named Korean qualified partners are listed companies with disclosed AI-buildout exposure in their order backlogs. HD Hyundai Electric is approximately $26 billion; Hyosung Heavy currently trades at approximately $20 billion (down from a peak of roughly $27 billion in 2025 on transformer demand), reflecting the normalisation of the initial AI-infrastructure re-rating cycle. GE Vernova is a large-cap. Hitachi Energy, Pennsylvania Transformer, and Virginia Transformer are private. Investability tracks accordingly.

Inside a data center building, medium-voltage (MV) to low-voltage (LV) step-down uses dry-type cast-resin transformers — physically and regulatorily distinct from the oil-filled grid transformers that sit upstream on the utility side.^[10]

Hammond Power Solutions (HPS-A.TO, a Canadian listing, approximately $2.3 billion) is the largest North American dry-type producer. Eaton, Hubbell, HD Hyundai Electric, and Siemens compete in the segment. Hammond is the cleanest pure-play in North America.

Hammond's Q3 2025 disclosure noted that post-Q3 orders received shortly after the quarter close were sized at 53% of Q3 closing backlog, primarily for 2026 data center delivery. That figure is direct disclosure of net new data center demand absorbed in a single quarter, not historical concentration.^[10] The company's North American competitive set is named; the disclosure is direct; the listing is mid-cap. Hammond is the cleanest North American dry-type pure-play in this segment, with disclosed multi-year visibility into AI-buildout demand.

Every extreme-ultraviolet (EUV) exposure mask must be protected during use by a pellicle — a thin transparent membrane that prevents particle contamination from landing on the mask surface and printing as defects on the wafer. Pellicles are an ASML-tool-adjacent layer: every EUV exposure step at TSMC, Samsung, and Intel depends on them, but the pellicle itself is a distinct, separately sourced component that gates EUV throughput independently of the scanner tool.^[11]

Mitsui Chemicals (a Japanese listing, approximately $4.2 billion) is the named primary commercial supplier of EUV pellicles. Shin-Etsu Chemical is a meaningful second supplier. Samsung-backed S&S Tech (a Korean small-cap, approximately $1 billion) is an emerging third source. Mitsui Chemicals, Shin-Etsu Chemical, and S&S Tech together hold an estimated 70 to 80 percent global pellicle share per industry-analyst aggregates. Mitsui leads, Shin-Etsu is the meaningful second supplier, and S&S Tech is the emerging Korean qualifier — a three-supplier concentration.

TSMC was reported in September 2025 to be building in-house pellicle capability at its repurposed Fab 3 facility — a development that confirms the dependency at scale and simultaneously threatens external supply on a three-or-more-year horizon.^[11] Mitsui Chemicals is the cleanest small-mid cap pure-play in this layer. S&S Tech is a Korean small-cap worth noting for investors focused on the Korean supply chain.

Every flip-chip ball-grid-array (FC-BGA) substrate used in Chip on Wafer on Substrate (CoWoS) packaging — and in NVIDIA Blackwell and Rubin GPU modules, AMD MI-series accelerators, and Intel server CPUs — is built on a dielectric film produced by the chemical materials division of a Japanese seasoning company.^[12]

Ajinomoto Build-up Film (ABF), produced by Ajinomoto Co., is that dielectric. All major outsourced semiconductor assembly and test (OSAT) laminators that produce high-end FC-BGA substrates — including Ibiden, Unimicron, AT&S, Kinsus, and SEMCO — source ABF from Ajinomoto's chemicals division. Competing alternatives include Sekisui Chemical, Resonac (formerly Showa Denko Materials), and Taiyo Holdings. None has matched Ajinomoto's qualified share at the most advanced FC-BGA process nodes as of 2025; both Sekisui and Resonac are working toward qualification at high-bandwidth memory and CoWoS-class processes, with multi-year qualification cycles ahead.^[12]

Ajinomoto Co. is a roughly $28 billion mid-cap; food is more than 40% of revenue. ABF is the highest-margin segment but does not make the equity a pure-play. Sekisui Chemical is a smaller, more diversified Japanese listing. A dominant qualified supplier with named challengers and a plausible medium-term relaxation timeline — concentration risk, not a permanent monopoly.

High-bandwidth memory (HBM) stacking is the rate-limiter on AI accelerator output. The bonder equipment is the rate-limiter on HBM stacking. HBM3E and HBM4-class die stacks require either thermocompression bonders (TCB) or hybrid bonders to join memory dies with sub-micron precision.^[13]

The qualified Western producers in this layer are BE Semiconductor (BESI, a Dutch listing, approximately $24 billion), ASMPT (a Hong Kong listing, approximately $9 billion), Kulicke & Soffa, and Korea's Hanmi Semi — at least four named competitors with overlapping capabilities. BESI holds the largest qualified share in hybrid bonding for HBM4-class flows as of 2025. Applied Materials acquired a 9% stake in BESI in April 2025 and co-developed the Kinex hybrid bonder — a structural bet on this layer being supply-constrained that would not make commercial sense if it were not. BESI's 2025 Investor Day projected 476 million euros in hybrid-bonding revenue by 2026.^{[13] [14]}

BESI and ASMPT appear on some AI-infrastructure equipment lists. The framing here is novel: this is the equipment for the equipment that gates HBM4, not simply a "packaging stock." The investment thesis tracks to HBM4 volume ramp timing, not generic packaging capex.

The 2025 National Defense Authorization Act (NDAA) Section 4872 prohibits Chinese-origin tungsten in Department of Defense supply chains effective January 1, 2027. The Defense Logistics Agency (DLA) issued a 2026 request for information (RFI) for 1,700 metric tonnes of tungsten for the National Defense Stockpile.^{[16] [17]}

Almonty Industries (AII, approximately $5 billion) commissioned the Sangdong tungsten mine in South Korea in March 2026 — a Western-aligned, NDAA-compliant production asset now in active ramp.^[18] Additional named non-Chinese producers include EQ Resources (Australia and Spain), Masan High-Tech Materials (Vietnam), Group 6 Metals (Australia), Plansee Group (private Austria), and Global Tungsten & Powders (a Plansee subsidiary with US operations). At least five named non-Chinese producers with meaningful output streams exist; this is a concentration risk with a structural NDAA-driven domicile constraint, not a sole-source situation.

The structural thesis: tungsten has both defense and semiconductor demand (WF6 for CVD tungsten in leading-edge logic) pulling on the same non-Chinese supply pool. The NDAA constraint is statutory and dated. Almonty's Sangdong commissioning is the production proof-point that Western-aligned supply is not just a paper projection.

EQ Resources (EQR.AX, approximately AUD 1 billion (USD ~660M-990M depending on FX)) operates the Mt Carbine tungsten project in Queensland, Australia, and the Barruecopardo project in Salamanca, Spain — two operating Western-domicile mines with output accessible to NDAA-compliant procurement frameworks.^{[19] [15]}

EQ Resources is the smaller of the two cleanest Western pure-plays in this theme. It has the same NDAA structural tailwind as Almonty and a similarly direct exposure to the defense-plus-semiconductor dual demand. Two operating mines rather than one in active ramp; the Mt Carbine restart and Barruecopardo operations together provide supply diversification within a single issuer.

Every EUV exposure step at TSMC, Samsung, and Intel uses photoresist from a small set of Japanese chemical houses. JSR Corporation — historically one of the four leading EUV photoresist producers — was acquired by Japan Investment Corporation (JIC) in June 2024 and is now state-backed and private, removing the cleanest pure-play from the public investable list. Most public AI-infrastructure exposure lists have not refreshed for this structural change.^[20]

The remaining publicly investable EUV photoresist producers are Tokyo Ohka Kogyo (TOK, a Japanese listing, approximately $8.5 billion), Shin-Etsu Chemical (approximately $84 billion), Fujifilm Electronic Materials (within Fujifilm Holdings, approximately $22 billion market capitalization at the group level), and Sumitomo Chemical (approximately $5.5 billion). TOK is the cleanest small-mid cap pure-play in this layer. Sumitomo Chemical is also small-cap-tier with focused semiconductor materials exposure. The competitive landscape did not narrow when JSR went private; removal of JSR from the public list changed which names carry the concentrated investable exposure.^{[20] [21]}

Every semiconductor fab consumes hydrofluoric acid (HF) for wet etch and wafer cleaning. The 2019 Japan-Korea export-control crisis — in which Japan imposed licensing requirements on exports of ultra-high-purity hydrofluoric acid (UHP-HF) to Korea, disrupting Samsung and SK Hynix supply chains — confirmed how concentrated this layer is and how quickly disruption propagates.^[22]

The three named global UHP-HF producers are Stella Chemifa (a Japanese listing, approximately $0.6 billion), Daikin Industries (a Japanese large-cap, approximately $35 billion), and Solvay (a Belgian listing, approximately $3 billion). In the specialty fluorinated etch gases layer — nitrogen trifluoride (NF3) and tungsten hexafluoride (WF6) — Kanto Denka Kogyo (a Japanese listing, approximately $1.5 billion) is a named small-cap producer. A Kanto Denka plant fire in August 2025 sent TSMC, Samsung, and Rapidus scrambling for NF3 alternative supply, providing real-world confirmation of the qualification concentration in this layer.^[23]

Stella Chemifa (~$0.6 billion) and Kanto Denka Kogyo (~$1.5 billion) are small-cap Japanese listings with focused semiconductor materials exposure — the strongest small-cap pure-play fit in this theme. Commodity HF exists at scale; semiconductor-grade material does not.

Every silicon-carbide (SiC) or gallium-nitride (GaN) power module in a hyperscaler power supply unit (PSU), electric vehicle inverter, or 800-volt direct current (800V DC) distribution rail uses a direct-bond-copper (DBC) substrate on either a silicon-nitride or aluminum-nitride ceramic base. The DBC substrate bonds copper electrodes to the ceramic with the precision required for high-power thermal cycling — a process specification that limits the qualified supplier set.^[24]

The dominant Western-aligned DBC producers are Kyocera (a Japanese listing, approximately $31 billion) and Rogers Corporation (ROG, a US listing, approximately $2.3 billion). CoorsTek (private, US) and Tokuyama (a Japanese listing, approximately $2 billion, aluminum nitride powder) are named in the ceramic substrate materials layer. Rogers is the small-mid cap pure-play with focused power-module substrate exposure — the pattern of a smaller, specialised producer holding qualified share in a critical sub-layer of AI power infrastructure.^[24]

The substrate sits at the junction of the power theme and the advanced materials theme, and neither theme's public framing typically surfaces it.