An AI equity analyst's view on the most critical and most undervalued companies across the AI data center supply chain.
Disclaimer: This is for educational and informational purposes only. It does not constitute investment advice, a recommendation, or solicitation to buy or sell any securities. Always conduct your own research and consult a licensed financial advisor before making investment decisions.
Critical Infrastructure
10 Most Important Companies
Ranked by how much the AI data center supply chain depends on them. If any of these companies disappeared tomorrow, the entire buildout stalls.
1
TSMC
TSMNYSE · Taiwan
Wafer FabAdvanced Packaging
~84% advanced logic foundry · Sole CoWoS provider at scale
The single most irreplaceable company in the AI supply chain. TSMC fabricates over 90% of the world's advanced-node chips — every NVIDIA GPU, AMD MI accelerator, Apple Silicon, Google TPU, and Amazon Trainium flows through their fabs. They also operate the only CoWoS advanced packaging line at scale. There is no substitute: Samsung's yields trail significantly at comparable nodes, and Intel Foundry is years from being competitive. If TSMC stopped producing tomorrow, the entire AI revolution halts within weeks.
~80% AI accelerator share · NVLink/NVSwitch · Spectrum-X
NVIDIA's dominance spans both the chip and the network. Their GPU architecture and CUDA software ecosystem create switching costs that no competitor has overcome — AMD's ROCm is years behind, and custom silicon (TPU, Trainium) only works for internal hyperscaler workloads. Beyond GPUs, NVIDIA's NVLink interconnect and Spectrum-X networking fabric are becoming the default for AI cluster networking. They define the AI data center architecture that everyone else builds around.
100% EUV lithography monopoly · 5+ year order backlog
The most absolute monopoly in the technology industry. ASML is the sole manufacturer of EUV lithography systems — the $380M machines required to print transistors below 7nm. No chip in any NVIDIA GPU, AMD accelerator, or Apple processor can be made without ASML's scanners. The technology took 30 years and $10B+ to develop, requires sole-source optics from ZEISS and lasers from TRUMPF, and no competitor is even attempting to replicate it. Their backlog stretches years into the future.
~57% HBM market share · Sole initial HBM3E supplier for NVIDIA
High Bandwidth Memory is the critical memory bottleneck for AI. Every GPU needs HBM stacks, and SK Hynix makes more than half of the world's supply. They were the sole initial supplier of HBM3E for NVIDIA's H200 and B200, and are leading the transition to HBM4 for the Rubin generation. HBM production requires advanced die stacking and TSV technology that only three companies in the world can do — and SK Hynix is years ahead of Samsung and Micron in yield and qualification.
#1 in UPS, switchgear, PDUs, busbars · Appears across 7+ subsystem cards
No single company touches more parts of the data center than Schneider Electric. They hold the #1 position in UPS systems (APC), switchgear, power distribution units, and busbar systems — plus significant presence in precision cooling, liquid cooling, racks, and modular data centers. Their EcoStruxure DCIM platform ties it all together. When a hyperscaler builds a new AI data center, Schneider is on virtually every purchase order. They are to data center infrastructure what TSMC is to chip fabrication.
The broadest semiconductor equipment company in the world. Applied Materials holds leadership positions across more process steps than any competitor — deposition, etch, CMP (near-monopoly at ~70%), ion implantation, and inspection. A single TSMC fab uses Applied Materials tools in dozens of different process steps. While ASML gets headlines for EUV, Applied Materials' equipment touches a wafer far more times during fabrication. They are the picks-and-shovels backbone of semiconductor manufacturing.
Broadcom is the hidden infrastructure layer of AI networking. Their Memory switch ASICs power ~75% of all data center Ethernet switches — meaning nearly every packet in every AI cluster flows through Broadcom silicon. They also design Google's TPU custom AI accelerators and hold the #2 position in NICs/DPUs. While NVIDIA gets credit for AI computing, Broadcom is equally critical for AI networking. The switch ASIC monopoly alone makes them irreplaceable.
KLA is the quality gatekeeper of semiconductor manufacturing. Their inspection and metrology tools verify that every layer on every wafer meets spec — without them, yield collapses and fabs produce expensive scrap. At ~57% of the process control market, they dominate a segment that grows in importance as nodes shrink (more layers = more inspection). They also hold a key position in EUV mask inspection. No advanced fab runs without KLA's approval at dozens of process checkpoints.
Tokyo Electron holds a near-total monopoly on coat-develop track systems — the equipment that applies and develops photoresist before and after every lithography step. With ~90% share (approaching 100% for EUV tracks), they are as essential to lithography as ASML itself: every EUV exposure requires TEL's coater and developer on either side. Beyond track, they hold strong #3 positions in etch and deposition. Their tools touch a wafer at virtually every stage of fabrication.
Vertiv is the most concentrated pure-play on AI data center infrastructure. While Schneider is larger overall, Vertiv derives a higher percentage of revenue from data centers. They hold top-2 positions across UPS, precision cooling, liquid cooling (increasingly critical for AI GPU racks), PDUs, and server racks. As AI power density per rack climbs from 20kW to 100kW+, Vertiv's thermal and power delivery solutions become non-negotiable. They appear across more infrastructure subsystem cards than any company except Schneider.
Ranked by market concentration and lack of viable substitutes. These companies have pricing power, multi-year visibility, and competitors that are years from catching up.
1
ASML
ASMLNASDAQ · Netherlands
EUV Lithography
100% EUV lithography · 100% High-NA EUV · Zero competitors
The most absolute monopoly in all of technology. There is no second source for EUV lithography — not a distant competitor, not an emerging startup, nothing. The technology required 30 years, $10B+, and convergence of three sole-source suppliers (ASML systems, ZEISS optics, TRUMPF lasers) to create. No nation, company, or consortium is attempting to replicate it. Every advanced chip in the world flows through ASML's machines.
100% actinic EUV mask inspection · Only tool using 13.5nm wavelength
A monopoly as absolute as ASML's but far less well-known. Lasertec is the sole manufacturer of actinic EUV mask inspection tools — the only equipment that can detect defects on photomasks at the actual 13.5nm EUV wavelength. Every EUV mask in the world must be inspected by a Lasertec system before it can be used in production. KLA offers non-actinic optical approaches but cannot see the same class of defects. A tiny Japanese company with an outsized chokehold on the entire EUV ecosystem.
~90% advanced node (<7nm) fabrication · Sole manufacturer of AI accelerator dies
Every leading AI chip — NVIDIA's Blackwell and Rubin, AMD's Instinct, Amazon's Trainium, Google's TPU — is fabricated by TSMC. At advanced nodes below 7nm, TSMC holds ~90% market share, with Samsung's distant ~8% plagued by yield issues. There is no realistic path to diversify away from TSMC for cutting-edge AI silicon. The entire $600B AI supply chain funnels through their fabs in Hsinchu. A single earthquake, a single geopolitical event, and the world's AI buildout stops.
~90% coat-develop track · Near-100% for EUV tracks
Every lithography exposure — and there are dozens per wafer — requires photoresist to be coated before and developed after. Tokyo Electron controls ~90% of the track systems that perform these steps, rising to near-100% for EUV-compatible tracks. SCREEN's ~7% share exists only in older nodes. TEL's tracks are so tightly coupled to ASML scanners that they're effectively sold as a pair. A monopoly that hides in ASML's shadow.
~80% AI accelerator share · CUDA ecosystem lock-in
NVIDIA's ~80% share of AI accelerators understates their actual monopoly power. The real moat is CUDA — a software ecosystem with 4M+ developers, millions of lines of optimized libraries, and a decade of ML framework integration. Switching from NVIDIA to AMD or custom silicon means rewriting and re-optimizing entire software stacks. This software lock-in makes NVIDIA's hardware monopoly self-reinforcing in a way that pure market share numbers don't capture.
Nearly every data center Ethernet switch in the world — from Arista, Cisco, and white-box vendors alike — runs on a Broadcom switch ASIC. Their Tomahawk and Jericho chip families define the industry's switching capacity roadmap. Marvell is the only alternative at ~20%, but Broadcom's ecosystem of SDK, drivers, and qualification makes displacement extremely difficult. A silicon monopoly that sits beneath every brand name in networking.
~70% CMP equipment · Ebara (~27%) is the only alternative
Chemical-mechanical planarization — flattening the wafer surface between deposition steps — is performed dozens of times per wafer. Applied Materials controls ~70% of CMP equipment, with Japan's Ebara the sole meaningful alternative at ~27%. At advanced nodes, CMP uniformity directly determines yield, giving AMAT significant pricing power. This near-monopoly is just one of several dominant positions AMAT holds across different process steps.
~57% process control · Share growing from 50% in 2010
KLA's dominance in process control is not just large — it's expanding. Their share has grown from ~50% in 2010 to ~57% today because advanced nodes require more sophisticated inspection that only KLA's platforms can deliver. At EUV nodes, the number of inspection steps per wafer increases dramatically, and KLA captures most of that growth. Applied Materials at ~17% is a distant second. When fabs set yield targets, KLA's tools are the arbiter.
~55% wafer dicing · ~47% wafer grinding · Tokyo Seimitsu is only real competitor
Every semiconductor die must be singulated from its wafer, and DISCO's precision dicing saws and laser cutting systems do over half of it globally. In wafer grinding (thinning wafers for 3D stacking and advanced packaging), they hold ~47%. Tokyo Seimitsu is the only real competitor. As advanced packaging drives thinner wafers and finer-pitch dicing, the precision requirements increase — playing directly to DISCO's technological lead.
While ASML dominates front-end lithography, Canon quietly controls ~55% of the packaging lithography market — the steppers that pattern redistribution layers, bumping, and fan-out structures on interposers and packages. Every CoWoS interposer, every fan-out package requires Canon's i-line or KrF systems. As advanced packaging volumes surge for AI, Canon's packaging lithography business is growing rapidly while their lost front-end business fades from relevance.
Smaller companies actively gaining market share with a credible path to becoming much bigger. These are the disruptors and fast-followers across the AI supply chain.
1
Technoprobe
TPROBIT · Italy
Probe Cards
~22% probe cards (#2) · Won 30% of TSMC 2nm qualification sockets
The fastest-growing probe card maker in the world. Technoprobe has surged from a niche Italian supplier to #2 globally at ~22%, closing the gap on FormFactor's ~28%. Their breakthrough moment: winning 30% of TSMC's 2nm probe card qualification sockets — a dramatic gain at the most important foundry in the world. Their vertical MEMS probe technology is proving competitive at the most advanced nodes, and TSMC qualification is a beachhead that expands with every new customer tape-out.
ACM Research's proprietary SAPS and TEBO megasonic cleaning technologies are genuinely differentiated — they clean at frequencies and energy levels that conventional spray-based systems cannot match. Growing from ~10% against entrenched incumbents SCREEN (~45%) and TEL (~20%). Their China-heavy customer base (~70% revenue) is both their growth engine and biggest risk, but the technology itself is winning evaluations on merit. If geopolitical tensions ease, their addressable market expands dramatically.
~17% TC bonding (#3) · Gaining share in HBM and advanced packaging
The Korean challenger in the critical thermocompression bonding market. Hanmi has grown to ~17% share behind Besi (~37%) and ASMPT (~22%) by offering competitive bonding tools at lower price points with strong SK Hynix and Samsung relationships. As HBM production scales exponentially for AI, bonding tool demand outstrips supply from the top two — creating an opening for Hanmi to capture overflow capacity. Their domestic Korean customer base gives them a structural advantage for HBM bonding specifically.
~28% ion implantation (#2) · Gaining share from Applied Materials
Axcelis is the pure-play David to Applied Materials' Goliath in ion implantation. While AMAT spreads R&D across 15+ product lines, Axcelis focuses 100% on implant — and it shows. Their Purion platform has been winning competitive evaluations and gaining share, growing from ~20% to ~28% over recent years. Every advanced logic wafer requires multiple implant steps for transistor junction formation. Focused pure-plays with improving technology tend to be underestimated until they're not.
Micron went from a distant third in HBM to a credible challenger ramping aggressively. Their HBM3E is now qualified with NVIDIA, and they're investing billions in HBM4 capacity. As the only US-based HBM manufacturer, they carry strategic importance under the CHIPS Act and for supply chain diversification away from Korean concentration. If SK Hynix hits capacity limits — which is likely given AI demand — Micron is the primary beneficiary of overflow orders.
Marvell is the only company chipping away at Broadcom's ~75% switch ASIC dominance, and they're simultaneously building a custom AI silicon business with major hyperscaler design wins. Their Teralynx switch family and custom ASIC platform give hyperscalers an alternative to Broadcom's lock-in. The custom silicon revenue is still early but growing rapidly — Amazon, Microsoft, and others want alternatives to both NVIDIA (for compute) and Broadcom (for networking). Marvell is positioned to be that alternative.
~17% optical transceivers (#3) · Largest Western transceiver supplier
The transition from 400G to 800G and 1.6T optical transceivers is a massive upgrade cycle, and Coherent (formerly II-VI + Finisar) is the largest Western supplier at ~17% share. While Chinese vendors InnoLight and Eoptolink lead on volume, Coherent's vertical integration — they manufacture their own laser chips, not just assemble modules — gives them a technology and quality edge for the highest-speed applications. As AI clusters demand 1.6T optics, Coherent's vertically integrated model becomes increasingly valuable.
Fuel cell leader · 1 GW AEP supply agreement · On-site DC power pioneer
Grid power is becoming the #1 bottleneck for new AI data centers — utilities can't deliver enough megawatts fast enough. Bloom Energy's solid oxide fuel cells offer on-site power generation that bypasses the grid entirely. Their landmark 1 GW supply agreement with AEP validates the model. While nuclear SMRs get more hype, Bloom's fuel cells are deployable today, not in 2030. As hyperscalers grow desperate for power, behind-the-meter generation goes from niche to necessity.
~4% process control · ~15% packaging lithography (#2) · Dual growth vectors
ONTO has a unique dual position spanning front-end metrology and back-end packaging lithography — two segments both growing with AI. Their packaging lithography business (~15% share, #2 behind Canon) rides the advanced packaging wave directly: more CoWoS, more fan-out, more RDL layers means more packaging litho steps. Meanwhile, their front-end optical metrology grows with node complexity. Few companies have two independent growth vectors tied to AI chip production.
As AI GPU power density climbs past 100kW per rack, air cooling becomes physically impossible — liquid cooling transitions from optional to mandatory. nVent Electric is growing into this transition with rear-door heat exchangers, coolant distribution units, and thermal management solutions specifically designed for high-density GPU racks. They're smaller than Vertiv and Schneider but more focused, and the liquid cooling TAM is expanding at 25%+ annually. Early positioning in a market that's about to become essential.
Companies where the market hasn't fully priced in their AI supply chain positioning — often due to foreign listings, conglomerate structures, or "boring industrial" perception.
1
DISCO Corp
6146TSE · Japan
Die Preparation
~55% wafer dicing · ~47% wafer grinding
Near-monopoly in precision wafer dicing and grinding, yet trades on a Japanese industrial multiple. Every AI chip must be singulated from its wafer using DISCO's blades or laser saws. Advanced packaging trends — thinner wafers, more chiplets, finer pitch dicing for CoWoS interposers — are a direct tailwind that increases both the number of cuts per wafer and the ASP of each tool. The market treats this as a boring industrial equipment company, not the monopoly gatekeeper it actually is.
Everyone knows SCREEN as the distant #2 in coat-develop track (~7% vs TEL's ~90%). Almost nobody talks about their ~45% dominance in wet cleaning — the single most repeated step in semiconductor fabrication, performed 100+ times per wafer. As nodes shrink, particle requirements get exponentially tighter, making cleaning more critical and more valuable. SCREEN's wet cleaning monopoly is a hidden crown jewel obscured by their tiny track share.
The Showa Denko + Hitachi Chemical merger created the world's broadest semiconductor packaging materials company, but post-restructuring complexity has kept the stock depressed. Underneath the corporate noise, they hold the #1 position in epoxy molding compounds — a material used to encapsulate every packaged AI chip. As advanced packaging complexity grows (more layers, larger packages, tighter reliability specs), Resonac's materials become more critical. Classic turnaround with leading market positions hiding underneath.
Besi gets all the investor attention for advanced packaging equipment, but ASMPT has the broadest assembly equipment portfolio and the #2 bonding position. The Hong Kong listing creates a structural valuation discount — Western investors either can't or won't buy HKEX stocks, despite ASMPT having nearly identical AI packaging exposure as Besi. As CoWoS and HBM bonding capacity expands, ASMPT benefits at a fraction of Besi's multiple.
~15% ABF substrates (#2) · Critical for GPU packaging
Every NVIDIA GPU sits on an ABF substrate — the high-layer-count organic board that connects the chip to the rest of the system. Ibiden is the #2 maker of these substrates and a key supplier for high-performance computing packages. ABF substrates have been a genuine bottleneck with multi-quarter lead times, yet Ibiden trades as a sleepy Japanese industrial. The market underprices how essential and supply-constrained their product is as AI chip packages grow larger and more complex.
#1 electronics manufacturer globally · Building GB200 NVL72 racks at scale
The largest electronics manufacturer in the world perpetually trades at 8-12x earnings. The market still prices Foxconn as a low-margin iPhone assembler, but their revenue mix is rapidly shifting toward high-value AI server integration — they're the primary manufacturer of NVIDIA's GB200 NVL72 racks. AI server assembly is higher-margin and higher-growth than consumer electronics. The valuation gap between "iPhone assembler" perception and "AI infrastructure builder" reality is significant.
The #2 fiber optic cable maker behind Corning, trading on an Italian industrial multiple. AI data center builds are driving a multi-year fiber demand supercycle — a single 10,000-GPU cluster requires millions of feet of fiber. Prysmian also holds the #1 position in submarine cables, benefiting from global data center interconnection. The European listing and "cable company" perception mask what is increasingly an AI infrastructure growth story at a value multiple.
~28% ion implantation (#2) · Gaining share from Applied Materials
The #2 ion implant equipment maker and actively gaining share from Applied Materials. Every advanced logic and memory wafer requires multiple implant steps to create transistor junctions. Axcelis is a focused pure-play in a growing equipment segment — while AMAT spreads across dozens of product lines, Axcelis concentrates all R&D on implant. Under-followed because ion implant is a "boring" process step that doesn't make headlines. Small-cap with improving margins and a clear competitive trajectory.
Legrand appears across multiple data center infrastructure cards — intelligent PDUs via their Raritan brand, busbar systems, server racks, and cable management. The market prices them as a French building electrical fittings company, but their Raritan intelligent PDUs are deployed in virtually every major hyperscale facility for per-outlet power monitoring and management. European listing and "electrical accessories" perception create a valuation discount on meaningful AI data center exposure.
~28% probe cards (#1) · Critical for wafer-level testing
The world's largest probe card maker at ~28% share. AI chips are getting larger (reticle-scale), more complex (chiplets, 3D stacking), and require more test points at finer pitches — all of which drives probe card ASPs higher. Every AI chip must pass wafer sort before packaging, and FormFactor's advanced MEMS probes are the physical interface that makes that testing possible. A pure-play on semiconductor test complexity growth that the market under-follows because probe cards aren't glamorous.