The Semiconductor Value Chain: Where AI Money Really Flows

Understanding the $1 trillion semiconductor industry and where the leverage points are

Following the Money

The semiconductor industry is complex, global, and deeply interdependent. When someone says “I’m investing in chips,” they could mean equipment makers, foundries, designers, memory manufacturers, or packaging specialists. These are fundamentally different businesses with different economics.

To invest intelligently in the AI infrastructure build-out, you need to understand where money flows in the value chain—and where it gets captured.

The Value Chain Map

Let’s trace how a cutting-edge AI chip gets made:

1. Design Tools → 2. Equipment → 3. Manufacturing → 4. Materials → 5. Packaging → 6. Memory → 7. Integration

Each layer has different competitive dynamics, margin profiles, and investment characteristics.

Layer 1: Semiconductor Equipment – The True Monopoly

ASML: The Most Important Company You’ve Never Heard Of

ASML is the only company in the world that makes extreme ultraviolet (EUV) lithography machines. Every cutting-edge chip—whether from Apple, Nvidia, AMD, or Qualcomm—must be made using ASML’s machines.

Here’s why this matters:

Technical monopoly: EUV lithography uses 13.5nm wavelength light to pattern chips. ASML spent 17 years and €6 billion developing this technology. The physics are so complex that no competitor has succeeded in replicating it.
Pricing power: EUV machines cost $150-200 million each. High-NA EUV (the next generation) will cost $400+ million. ASML has pricing power because alternatives don’t exist.
Multi-year backlog: Lead times for EUV machines exceed 18 months. ASML knows exactly what revenue they’ll book years in advance.
Recurring revenue: Each machine requires constant servicing, upgrades, and consumables. ASML captures revenue long after the initial sale.

Investment implication: ASML is the ultimate picks-and-shovels play. Every AI chip depends on their machines. They have no competition, massive margins (50%+ gross margins), and visibility extending years into the future.

Real-world validation: In 2025, ASML is gearing up for “huge expansion” specifically to meet AI-driven demand. When the company with a monopoly needs to expand capacity, you know demand is real.

The Supporting Cast

While ASML dominates lithography, other equipment makers control critical steps:

Applied Materials: Deposition and etching equipment
Lam Research: Specialized etching systems
Tokyo Electron: Coating and developing equipment
KLA Corporation: Inspection and metrology

These companies don’t have ASML’s monopoly, but they have oligopoly positions. The top 3-5 players control 80%+ of their respective markets.

Investment thesis: Equipment makers are early-cycle plays. They get paid years before chips generate revenue. They have multi-year order books, limited competition, and essential products. When Nvidia announced a deal with Intel’s foundry, ASML, Applied Materials, and Lam Research all surged—the equipment makers win regardless of which chip company succeeds.

Layer 2: Foundries – Where Chips Are Actually Made

The global foundry industry is effectively a duopoly for advanced nodes:

TSMC: The 900-Pound Gorilla

Taiwan Semiconductor Manufacturing Company (TSMC) manufactures over 90% of the world’s most advanced chips. Here’s what makes them dominant:

Process leadership: First to 3nm, first to 2nm, consistently 12-18 months ahead of competitors
Customer lock-in: Designing a chip for TSMC’s process takes 2-3 years and costs hundreds of millions. Customers don’t switch lightly.
Scale advantages: TSMC’s R&D spending ($30B+ annually) dwarfs competitors. This gap compounds over time.
Manufacturing excellence: Industry-leading yields mean customers get more working chips per wafer

The Taiwan risk: TSMC’s dominance creates a single point of failure. Any conflict in Taiwan would cripple global chip production. This geopolitical risk is both TSMC’s biggest concern and a moat—governments worldwide are desperate to reduce dependence but haven’t succeeded.

Recent development: The U.S. is spending billions to bring TSMC fabs to Arizona, but these will be 1-2 generations behind Taiwan’s leading-edge facilities. Geography is TSMC’s moat.

Samsung: The Eternal Second Place

Samsung has the technology and capital to compete with TSMC but consistently trails in process technology and yields. They’re relevant for two reasons:

Supply diversification: Customers like having a TSMC alternative, even if it’s inferior
Memory integration: Samsung leads in HBM, which is increasingly critical for AI chips

Investment thesis: TSMC is the foundation of modern computing. Every AI chip from Nvidia, AMD, Apple, and Qualcomm is made there. Unlike equipment makers who sell tools, TSMC provides irreplaceable manufacturing capacity. The risk is geographical and geopolitical, not competitive.

Intel: The Foundry Wildcard

Intel’s foundry business (Intel Foundry Services) is attempting a comeback after years of manufacturing missteps. The Nvidia partnership announced in 2025 is significant—Nvidia agreeing to use Intel’s process validates the technology.

But Intel’s foundry is still bleeding money, and trust must be rebuilt after years of delays and execution failures. This is a multi-year turnaround story, not a current investment thesis.

Position: Hold Intel if you already own it, but new capital should go to TSMC until Intel proves consistent execution.

Layer 3: The Custom Silicon Revolution

Something fundamental changed in chip design over the past 5 years: hyperscalers stopped buying off-the-shelf processors and started designing their own.

Why Custom Chips Matter

Google’s TPUs: Purpose-built for TensorFlow, 10x more efficient than GPUs for specific workloads
Amazon’s Graviton/Trainium: Custom ARM chips saving 40% on compute costs
Meta’s MTIA: Optimized for recommendation models that drive Facebook/Instagram
Microsoft’s Maia: Designed specifically for GPT training and inference

These aren’t science experiments—they’re production chips running at massive scale.

Broadcom: The Custom Silicon Enabler

Here’s where Broadcom becomes critical. They don’t make chips for sale—they help hyperscalers design custom accelerators. Broadcom provides:

Custom ASIC design: The IP blocks and expertise to build specialized chips
Networking silicon: The interconnects that let thousands of chips work together
Optical networking: The technology to move data between data centers at scale

Investment thesis: Broadcom is a second-order beneficiary of AI infrastructure. As hyperscalers spend hundreds of billions on data centers, Broadcom captures a percentage through custom chip design and networking infrastructure. They have high switching costs (customers are deeply integrated) and pricing power (few alternatives exist for custom accelerators at scale).

Real-world validation: Broadcom’s AI revenue is projected to exceed $15 billion annually, making them one of the largest AI semiconductor companies despite not making GPUs.

Layer 4: The Memory Bottleneck

Here’s a dirty secret about AI chips: memory bandwidth, not compute power, is often the real bottleneck.

High-Bandwidth Memory (HBM): The Hidden Constraint

Modern AI accelerators use High-Bandwidth Memory—essentially stacks of DRAM positioned extremely close to the processor. HBM provides 10-20x more bandwidth than traditional memory.

The problem: HBM is in severe shortage.

Supply constraints: Only SK Hynix, Samsung, and Micron can produce HBM3, and manufacturing capacity is limited
Complex manufacturing: HBM requires advanced packaging (TSV technology) with yields below 50%
Explosive demand: Every new AI chip generation requires more HBM

Current market: HBM prices are 3-5x higher than standard DRAM, and lead times extend 40+ weeks. This is a textbook supply shortage.

Investment implication: Memory manufacturers like Micron have unprecedented pricing power. HBM will be constrained for years, not quarters. This is a structural advantage, not a cyclical uptick.

Consider: A single Nvidia H100 GPU requires 80GB of HBM3. A training cluster might have 16,000 GPUs. That’s 1.28 million GB of HBM just for one cluster. Now multiply by hundreds of clusters worldwide. The math doesn’t work without massive capacity expansion.

Layer 5: Advanced Packaging – The Next Frontier

As transistors can’t shrink much further, the industry is turning to advanced packaging—placing multiple chiplets in a single package.

Why this matters:

Chiplets: Breaking large chips into smaller pieces improves yields and costs
3D stacking: Stacking chips vertically (like HBM) maximizes bandwidth
Heterogeneous integration: Combining chips made on different process nodes optimizes cost and performance

Key players:

TSMC: CoWoS (Chip-on-Wafer-on-Substrate) packaging for Nvidia’s chips
Amkor: Independent packaging and test services
ASE Technology: High-volume advanced packaging

Investment thesis: Advanced packaging is becoming as critical as the chips themselves. TSMC’s packaging capacity is sold out years in advance. This is a structural growth story as the industry shifts from 2D scaling to 3D integration.

Putting It All Together: Portfolio Construction

Based on this value chain analysis, here’s how to think about semiconductor exposure:

Tier 1 (Core Holdings – 40-50% of semiconductor allocation)

ASML: The monopoly. Every advanced chip depends on their machines. Multi-year backlog, 50%+ margins, no competition.

TSMC: The foundry leader. Makes 90%+ of advanced chips. Geographic risk, but no manufacturing alternative.

These are the “must-own” positions. They capture value regardless of which AI companies win at the application layer.

Tier 2 (Strong Holdings – 30-40%)

Broadcom: Custom silicon and networking. High switching costs, growing AI revenue, strategic to hyperscalers.

Micron (or Samsung/SK Hynix): HBM shortage creates multi-year tailwind. Structural advantage, not cyclical.

Nvidia: Still the dominant GPU maker. Expensive but executes consistently. Consider as a core holding if you missed the earlier run.

Tier 3 (Selective – 10-20%)

AMD: Credible GPU alternative to Nvidia. Growing MI300 traction but still distant second.

Applied Materials / Lam Research: Equipment oligopoly plays. Less dominant than ASML but still essential.

Intel: Foundry turnaround story. High risk, high reward if they execute. Not a core position.

What to Avoid

Trailing-edge foundries: Companies stuck on 28nm or 40nm nodes. AI chips require leading edge.

CPU-focused companies: Traditional processors are losing relevance in AI workloads. ARM is the exception due to power efficiency.

Chinese chipmakers: Geopolitical risk too high given export controls. SMIC and others are cut off from advanced equipment.

Validation Metrics

Track these indicators quarterly:

ASML order book: Should extend 12+ months. Cancellations are red flags.
TSMC capacity utilization: Above 90% = supply constrained (good). Below 80% = demand concerns.
HBM pricing: Sustained high prices confirm memory bottleneck thesis.
Hyperscaler custom chip announcements: More custom silicon = more Broadcom revenue.
Foundry capex: TSMC, Samsung spending should stay elevated. Cuts signal demand weakness.

The Geopolitical Overlay

No semiconductor discussion is complete without addressing geopolitical risk:

Taiwan risk: TSMC’s concentration creates systemic risk. Any conflict would devastate global chip supply. This is both a concern and an opportunity—governments worldwide are desperate to reduce dependence, creating subsidies and investment incentives.

Export controls: U.S. restrictions on advanced chip equipment to China reshape competitive dynamics. ASML can’t sell EUV to China. This limits China’s ability to make leading-edge chips but also limits ASML’s addressable market.

Reshoring initiatives: CHIPS Act in the U.S., European Chips Act, and similar programs globally are bringing fabs back home. This is a multi-year tailwind for equipment makers and foundries.

Investment implication: Geographic diversification matters. Favor companies with global manufacturing footprints or those enabling reshoring (equipment makers benefit regardless of where chips are made).

Conclusion: Where to Place Your Bets

The semiconductor value chain has clear winners:

ASML: Monopoly on critical technology. This is the single highest-conviction play.
TSMC: Foundry leader with 2-3 year process advantage. Geographic risk is real but alternatives don’t exist.
Broadcom: Custom silicon enabler positioned at the intersection of hyperscalers and AI infrastructure.
Memory manufacturers: HBM shortage creates multi-year pricing power. Structural, not cyclical.

These companies capture value regardless of which AI applications succeed. They’re the infrastructure—and infrastructure gets built first.

Next in this series: “The Cloud Platform Wars: Data Infrastructure in the AI Age” – How hyperscalers are spending $320B and which data platforms are positioned to capture value.