Most analysts talk about AI stocks from the software side. We trace the hardware supply chain — from GPU die to the power transformer outside the data center — and identify who gets paid at every step.
By EcrioniX · Updated June 2026 · Educational analysis, not financial advice
⚠
Not financial advice. This page is an educational breakdown of the AI hardware supply chain written by chip engineers. Nothing here is a recommendation to buy or sell any security. Do your own research. Past supply-chain surges do not predict future performance. Consult a licensed financial advisor before investing.
The Chip Engineer's Lens on AI Investing
When ChatGPT launched in late 2022, most investors bought AI software companies. But the engineers who build the hardware asked a different question: what physical components does every AI model run on, and who makes them?
This is the supply chain question — and history shows it often identifies the biggest winners before the market does. When the internet exploded in the 1990s, Cisco (networking hardware) and Intel (server chips) outperformed most dot-com stocks. When smartphones took over in 2010–2015, TSMC (the fab manufacturing every smartphone chip) quietly 10×'d. When AI training scaled up in 2023–2024, SK Hynix (the dominant HBM memory maker) surged over 100% in 12 months.
The SanDisk pattern — how supply chain investing works
When mobile phones scaled from feature phones to smartphones, NAND flash memory demand exploded (every phone needs storage). SanDisk, Micron, and SK Hynix — the NAND flash makers — all surged because they supplied the raw material every phone OEM needed. The same pattern repeated with AI: GPUs need HBM (High Bandwidth Memory) — a specialised, expensive DRAM stacked on the GPU package. Only SK Hynix, Micron, and Samsung make it. When NVIDIA shipped millions of H100 GPUs in 2023, HBM demand went vertical, and SK Hynix's HBM revenue grew over 300% year-over-year. The SanDisk pattern struck again.
The framework is simple: map the AI hardware stack → identify who has monopoly or near-monopoly supply at each layer → find the companies with pricing power and capacity constraints. That is where the supply-chain upside concentrates.
The AI Hardware Supply Chain — Visualised
Every AI training cluster — from model to power grid
Each box in that diagram represents a supply-chain layer where specific companies have significant pricing power. Let's walk through each tier and the companies that dominate it.
Tier 1 — Direct AI Hardware (Highest Exposure)
These companies sell directly into AI data center buildout. Revenue scales linearly with the number of AI GPUs deployed.
🔴 Tier 1 — Direct, High Conviction
NVDA
NVIDIA Corporation
Designs the H100/H200/B200 GPUs that train every major AI model. Controls CUDA software ecosystem — switching costs are so high that labs don't switch even if AMD offers better hardware specs. Data center revenue went from $3B (2022) to $49B (2024).
Direct AI compute monopoly · CUDA lock-in
TSM
Taiwan Semiconductor (TSMC)
The only foundry that can manufacture NVIDIA's H100 at 4nm, Apple's A-series at 3nm, and AMD's MI300X. Every advanced AI chip in the world is made here. Also owns CoWoS (Chip on Wafer on Substrate) — the advanced packaging tech that stacks HBM onto GPU dies. Capacity is the bottleneck.
Sole manufacturer of all advanced AI silicon
000660.KS
SK Hynix
Dominant supplier of HBM3 and HBM3E — the high-bandwidth memory stacked on every H100 GPU. Each H100 needs ~80 GB of HBM3. NVIDIA-exclusive supply agreement in 2023. HBM ASP is 5–6× standard DRAM. HBM revenue grew 300%+ in 2024. Korean-listed.
Sole preferred HBM supplier to NVIDIA
MU
Micron Technology
Second major HBM supplier (after SK Hynix). Ramping HBM3E production for NVIDIA H200 and GB200 GPUs. Also supplies LPDDR5X for AI inference on-device (Apple, Qualcomm). HBM ASP is dramatically higher margin than standard DRAM, and AI demand is structural, not cyclical.
HBM ramp + LPDDR5X for edge AI inference
ASML
ASML Holding (Netherlands)
The only company in the world that makes EUV (Extreme Ultraviolet) lithography machines — the €300M machines that TSMC needs to manufacture chips at 5nm and below. No ASML machines = no advanced AI chips. A near-perfect monopoly with 10-year equipment backlogs.
Monopoly on EUV — no substitute exists
AMD
Advanced Micro Devices
Second GPU vendor after NVIDIA. MI300X is strong in HPC and gaining in AI inference. ROCm (AMD's CUDA equivalent) is maturing. Main risk: CUDA moat is deep and enterprises default to NVIDIA. AMD wins when supply is constrained or when price sensitivity forces alternatives.
NVIDIA alternative — second-mover but real
Tier 2 — Picks & Shovels (Indirect but Essential)
These companies don't sell GPUs, but every AI chip in existence flows through their tools, their processes, or their silicon. They benefit regardless of which AI company wins — just from total volume of AI silicon designed and built.
🟡 Tier 2 — Indirect, Structural Benefit
SNPS
Synopsys
The dominant EDA (Electronic Design Automation) software company. Every AI chip — from NVIDIA's Blackwell to Apple's M4 to Google's TPU v5 — is designed using Synopsys tools (Design Compiler, Fusion Compiler, PrimeTime for timing signoff). Revenue scales with the number of chip tape-outs. AI chips are complex → more tool licenses needed.
Every AI chip designed in their tools · 60%+ market share
CDNS
Cadence Design Systems
Number-two EDA vendor, strong in analog/mixed-signal and custom IC design (Virtuoso). Also makes the Palladium/Protium emulation platforms that AI chip companies use to verify designs before tape-out. NVIDIA, Qualcomm, and Apple are all major customers. Duopoly with Synopsys — no credible third player.
EDA duopoly · AI verification platform growth
AVGO
Broadcom
Designs custom AI ASICs (XPUs) for Google (TPU), Meta (MTIA), and Apple. Revenue from custom AI silicon is growing rapidly. Also supplies Ethernet switching silicon (Tomahawk/Trident series) for data center networking — AI clusters need massive bandwidth between nodes. Two AI revenue streams in one company.
Custom AI ASICs + AI data center networking
MRVL
Marvell Technology
Custom AI ASIC design for Amazon (Trainium/Inferentia chips), Microsoft, and others. Also makes optical DSPs for 400G/800G data center interconnects — AI training clusters need massive low-latency inter-GPU optical connections. Management guided to over $1.5B AI revenue in FY2025, growing fast.
Custom AI ASICs + optical interconnect DSPs
AMAT
Applied Materials
Largest semiconductor equipment company by revenue. Makes deposition (CVD/PVD), etch, and CMP machines that every fab — including TSMC — uses to manufacture chips. Advanced packaging for HBM (CoWoS) requires their bonding and deposition tools. Benefits from every TSMC capacity expansion.
Fab equipment leader · TSMC capacity expansion
KLAC
KLA Corporation
Dominates process control — the inspection and metrology equipment that checks wafers for defects at every step. Advanced nodes (3nm, 2nm) require far more inspection steps. KLA has ~50% market share in this segment. Every time TSMC adds capacity for AI chips, KLA equipment revenue follows.
Process control monopoly · scales with fab capacity
Tier 3 — AI Infrastructure (Data Center Build-Out)
These companies supply the physical infrastructure that AI data centers require — power delivery, cooling, networking, and test equipment. They benefit from the construction and operation of AI clusters, not just chip manufacturing.
🟢 Tier 3 — Infrastructure, Longer Cycle
VRT
Vertiv Holdings
Makes UPS systems, power distribution units (PDUs), and liquid cooling infrastructure for data centers. An AI GPU rack (H100 × 8) consumes ~10 kW — 3× a standard server rack. Data centers need Vertiv's liquid cooling to handle this density. Stock went from $12 (2023) to over $100 (2024) — a 700% run on AI data center buildout.
AI power density forces liquid cooling adoption
ANET
Arista Networks
Dominant data center Ethernet switching company. AI training clusters connect thousands of GPUs — NVIDIA's DGX SuperPOD uses Arista 400GbE switches. Meta, Google, Microsoft are major customers building out AI networking infrastructure. Revenue grows with scale of AI cluster deployments.
Preferred AI cluster Ethernet switching vendor
COHR
Coherent Corp
Makes optical transceivers for high-speed data center interconnects (400G, 800G, 1.6T). AI clusters need massive inter-rack and inter-data-center optical bandwidth. Coherent and II-VI (now merged) dominate 800G optics. Every GPU cluster expansion increases optical transceiver demand.
800G optical transceivers for AI cluster interconnect
TER
Teradyne
Makes automated test equipment (ATE) used to test every chip before it ships. AI chips (H100, M4, TPU) are large, complex dies that require extensive testing — burn-in testing, memory testing, SoC testing. Revenue scales with volume of AI chip production. AI chips also require more test time per die, increasing revenue per unit.
More AI chips shipped = more test revenue
ONTO
Onto Innovation
Makes optical metrology and inspection equipment for advanced packaging — the CoWoS and HBM stacking processes that connect GPUs to HBM memory. Advanced packaging is the fastest-growing segment in semiconductor equipment, driven almost entirely by AI chip packaging requirements.
AI data centers consume enormous amounts of electricity. A large AI cluster (100,000 H100 GPUs) draws ~1 GW — equivalent to a small city. Power companies near major data center corridors (Virginia, Texas, Arizona) are seeing structural demand increases. Nuclear-adjacent power companies are also re-rating upward.
Power demand from AI data centers · longer cycle
Full Comparison: AI Supply Chain Position
Company (Ticker)
AI Revenue Link
Monopoly / Moat
Tier
Risk
NVIDIA (NVDA)
Direct — GPU sales
CUDA software lock-in
1
AMD ROCm, custom ASICs, export controls
TSMC (TSM)
Direct — every AI chip fab'd here
Only 3nm/4nm EUV fab
1
Taiwan geopolitical risk
SK Hynix (000660)
Direct — HBM3/HBM3E to NVIDIA
HBM supply exclusivity
1
Competitor ramp (Micron), DRAM cycle
Micron (MU)
Direct — HBM3E + LPDDR5X
Second HBM supplier
1
DRAM commoditisation, memory cycle
ASML (ASML)
Indirect — every advanced fab needs EUV
Absolute EUV monopoly
1
Export controls to China, long backlog cycles
Synopsys (SNPS)
Indirect — every AI chip designed in EDA tools
60%+ EDA market share
2
Ansys acquisition regulatory risk
Cadence (CDNS)
Indirect — EDA + emulation
Duopoly with Synopsys
2
Limited — EDA duopoly is stable
Broadcom (AVGO)
Direct — custom XPUs + Ethernet switching
Custom ASIC + networking silicon
2
Customer concentration (Google/Meta)
Marvell (MRVL)
Direct — custom ASICs (AWS) + optical DSP
Custom silicon + optical
2
Execution risk on custom programs
Applied Materials (AMAT)
Indirect — fab equipment
CVD/PVD market leader
2
Cyclical capex, China export controls
Vertiv (VRT)
Indirect — power/cooling for AI racks
Liquid cooling expertise
3
Competition, build cycle lags chip cycle
Arista (ANET)
Indirect — AI cluster networking
Software-defined networking
3
Cisco competition at enterprise level
Coherent (COHR)
Indirect — 800G optical transceivers
Scale in high-speed optics
3
Price erosion in transceivers
The Patterns That Repeat in Semiconductor Investing
Pattern 1 — The Bottleneck Bet
In every technology wave, one component becomes the bottleneck. In AI, HBM memory is the bottleneck for GPU training performance — not compute. NVIDIA and Google would buy more GPUs if they could get more HBM. The bottleneck company (SK Hynix for HBM, ASML for EUV machines) has the most pricing power. Find the bottleneck; that's where the margin concentrates.
Pattern 2 — The Monopoly Multiplier
Companies with near-monopoly supply of an essential component see their revenues scale proportionally with the entire industry, without the winner-takes-all risk of the application layer. ASML sells to TSMC, Samsung, AND Intel. Synopsys sells to NVIDIA, Qualcomm, Apple, AND their competitors. Every AI chip designed anywhere flows through their software. This is structurally better than picking which AI model wins.
Pattern 3 — The Infrastructure Lag
Infrastructure buildout (power, cooling, networking) lags chip deployment by 6–18 months. When data centers start buying GPUs, the power transformers and cooling systems need to be ordered and installed. This lag creates a longer, more predictable revenue ramp for infrastructure companies like Vertiv — but it means they peak later in the cycle than chip companies.
Pattern 4 — The Capex Flywheel
When hyperscalers (Microsoft, Google, Meta, Amazon) commit $100B+ annual AI capex, that money flows into: (1) GPU purchases → NVIDIA, AMD; (2) Memory → SK Hynix, Micron; (3) Fab capacity → TSMC → Applied Materials, KLA; (4) Data center build-out → Vertiv, Eaton; (5) Networking → Arista, Broadcom. Each dollar of AI capex gets distributed across the entire supply chain.
Risks and What Could Break the Thesis
1
AI spending consolidation. If hyperscalers cut AI capex (e.g., a recession, a better compute paradigm, or ROI disappointment), the entire supply chain contracts simultaneously. This is the main systemic risk.
2
Export controls. US export restrictions on advanced chips and semiconductor equipment to China reduce TAM. ASML can't sell EUV to China. NVIDIA sells downgrade chips (H20). This constrains growth.
3
Custom ASIC disruption. If every hyperscaler builds its own AI chip (Google's TPU, Amazon's Trainium, Meta's MTIA, Apple's ANE), NVIDIA's GPU market share could shrink. Custom ASICs are growing. However, custom ASIC volume goes through Broadcom/Marvell and is still made at TSMC, so Tier-1 fabs and EDA still win.
4
Taiwan geopolitical risk. TSMC manufactures ~90% of the world's advanced chips. Any Taiwan Strait conflict would be a global supply chain catastrophe. TSMC is diversifying to Arizona and Japan fabs, but capacity is years away from replacing Taiwan.
5
Valuation. Many of these names trade at 30–60× forward earnings. Good supply-chain positioning doesn't protect against overpaying. High valuations mean the AI thesis must continue to outperform expectations for returns to be positive.
Frequently Asked Questions
Is NVIDIA still a buy after its 10× run?
This page doesn't make buy/sell recommendations. From a supply chain perspective, NVIDIA's CUDA moat is still intact — no hyperscaler has fully replaced NVIDIA GPUs with custom ASICs for training workloads. But at a market cap above $3T, the stock already prices in significant continued AI spending growth. Whether it's "a buy" depends entirely on your assumptions about future AI capex, competitive dynamics, and valuation tolerance — all of which are outside the scope of this technical analysis.
Why did Vertiv (VRT) go up 700% — it just makes power equipment?
AI GPU racks generate 10 kW of heat — three times a standard server rack. A large AI data center (100MW) needs liquid cooling loops, precision power distribution, and UPS systems that standard air-cooled data centers don't need. Vertiv was the market leader in thermal management and power distribution when this need suddenly exploded. The market re-rated the company from "boring data center vendor" to "AI infrastructure critical supplier." The pattern is identical to the memory story: an unsexy component becomes suddenly essential.
What about smaller AI chip startups like Groq, Cerebras, or SambaNova?
They're all private. You can't directly invest in them in public markets today. Their chips are still manufactured at TSMC and designed with Synopsys/Cadence EDA tools — so even if one of them disrupts NVIDIA, the EDA and fab supply chain still wins. When these companies eventually IPO, the supply-chain picks-and-shovels plays will already have captured years of growth.
What's the difference between AI training and AI inference stocks?
Training is compute-intensive (building AI models) — dominated by large NVIDIA H100/B200 GPU clusters in cloud data centers. Inference is deployment-intensive (running AI models in production) — distributed across edge devices, smaller cloud instances, and specialised inference chips. Training demand benefits HBM memory and high-end GPUs most. Inference at scale benefits mobile SoC makers (Qualcomm, Apple) and networking companies more. Both are growing simultaneously.
⚠ Final Disclaimer: This page is written by hardware/chip engineers to explain the AI semiconductor supply chain. It is educational content only. Nothing on this page constitutes financial advice, investment advice, or a recommendation to buy or sell any security. All investments carry risk, including the possible loss of principal. Past supply-chain performance does not guarantee future returns. Always consult a qualified financial professional before making investment decisions.