Andrew Homan & Chris Miller - Redefining Semiconductor Progress - [Invest Like the Best, EP.396]

Okay, here is the analysis report generated based on the podcast transcript you provided and the strict rules.

---

At a Glance

This episode features Andrew Homan, Managing Partner of Maverick Silicon, and Chris Miller, Professor at Tufts University and author of Chip War. They delve into the AI-driven transformation of the semiconductor industry, with the core argument being: AI is triggering a paradigm shift comparable to the PC and mobile internet eras, but NVIDIA’s dominance is not unassailable. The biggest investment opportunities in the future may lie not in GPUs themselves, but in adjacent areas such as interconnect, memory, power consumption, and edge AI.

---

Thematic Section

1. Intel’s Predicament: The Innovator’s Dilemma of the Successful

Chris Miller argues that Intel’s current predicament is precisely the result of its past success. Its over-reliance on the lucrative profits from PC and data center chips made it risk-averse and slow to act when faced with paradigm shifts like mobile and AI.

• Historical Context: Intel missed the opportunity to manufacture chips for the iPhone (Steve Jobs had approached the company with the request) and largely missed the AI wave. Andrew Homan cites Andy Grove’s “Only the Paranoid Survive,” noting that Intel failed to heed its founder’s warning.
• Mechanism Breakdown: Chris Miller points out that for institutions and individuals, embracing the “innovator’s dilemma” means sacrificing the products that made them famous and promoting employees from new business lines, creating immense internal tension. It is far easier to maintain the status quo, especially when it remains highly profitable.
• Data Support: Intel’s struggle with the 10-nanometer process node (roughly 7–8 years ago) was the key inflection point where it was overtaken by TSMC.

2. The “Three-Layer Cake” of AI Investment and Value Capture

Andrew Homan compares the AI ecosystem to a “three-layer cake”: the bottom layer is chips and cloud, the middle layer is foundation models, and the top layer is applications. He believes that for the foreseeable future, the bottom layer (the chip layer) will capture the most economic rent.

• Mechanism Breakdown: Andrew argues that the ultimate winners in the middle layer (e.g., OpenAI) and the top layer (e.g., ChatGPT) remain unclear, and the path to value capture is very fuzzy. In the chip layer, however, the demand for computing power is almost certain, regardless of how AI evolves. Chris Miller adds that the moat of the chip layer comes not only from hardware but also from the “silicon-software” interface (e.g., NVIDIA’s CUDA ecosystem), which increases the stickiness of its value.
• Data Chain: Andrew cites data showing that the cost of training a frontier model surged from $10 million in 2022 to $100 million in 2023, $1 billion in 2024, and is projected to reach $10 billion in 2025. This results in a cost structure where 90% of a frontier model company’s expenses are computing costs.
• Extrapolation and Falsification: Andrew acknowledges that if “Scaling Laws” are disproven, or if increased competition in the AI accelerator market leads to sharp price declines, the value captured by the chip layer could diminish. However, he believes that even with lower prices, the massive market size could still foster great companies (e.g., Intel in the PC era, ARM and Qualcomm in the mobile era).

3. NVIDIA’s “Death Star” and Potential Weaknesses

Andrew Homan believes that directly challenging NVIDIA’s GPU core is very difficult; its true weakness may lie in the “generality” and “high power consumption” of its chips, creating opportunities for customized chips for specific scenarios and edge AI.

• Mechanism Breakdown: NVIDIA’s GPU is powerful because of its flexibility and generality, allowing it to adapt to evolving model architectures. However, if future AI model architectures (e.g., Transformers) stabilize, startups could design highly optimized application-specific integrated circuits (ASICs) for those architectures, achieving comparable or better performance at a fraction of NVIDIA’s cost.
• Data and Analogy: Andrew uses the “Star Wars Death Star” as a metaphor for NVIDIA’s dominance, while “Luke Skywalker’s arrow” represents specialized chips for specific models. He also notes that NVIDIA’s chips have extremely high power consumption (“If I plug a laptop with an NVIDIA GPU into an airplane, the entire plane would drop 200 feet”), making them uncompetitive in power-sensitive “edge” markets (e.g., smartphones, automobiles).
• Extrapolation: Andrew observes that NVIDIA’s CEO, Jensen Huang, is concentrating nearly all resources on the data center business, while its automotive business has been largely stagnant for the past eight quarters. This leaves a window of opportunity for startups in the edge AI space.

4. TSMC’s Moat and Geopolitical Factors

Chris Miller points out that TSMC’s moat lies not only in its technological leadership but also in Taiwan’s unique, highly concentrated semiconductor manufacturing ecosystem, making it extremely difficult to replicate its model.

• Mechanism Breakdown: TSMC’s competitive advantage stems from its focus on a “single business model” (pure-play foundry) and the highly developed manufacturing ecosystem on the island of Taiwan—from chemical suppliers to equipment maintenance engineers, all within a 1.5-hour high-speed rail ride. This cluster effect generates extremely high efficiency.
• Data and Analogy: Andrew Homan uses the analogy of “hitting a golf ball from Earth and getting a hole-in-one on the Moon” to describe the precision of EUV lithography. Chris Miller notes that chip manufacturing is the most complex manufacturing process in human history, with a fault tolerance one-thousandth that of an automobile (nanometers vs. millimeters).
• Geopolitical Extrapolation: The three main goals of the U.S. government are: 1) reducing reliance on Taiwan for chip manufacturing; 2) maintaining U.S. technological leadership; and 3) preventing China from obtaining advanced AI chips. Chris believes that while the $40 billion in subsidies from the CHIPS Act is substantial, it is only equivalent to TSMC’s annual capital expenditure, so it can only change the landscape “at the margin.” China’s attempts at self-sufficiency are extremely difficult because it cannot access the world’s most advanced equipment and materials.

5. Investment Perspective: Cycles, Bubbles, and “Edge” Opportunities

Andrew Homan believes that the semiconductor industry is a “growth cyclical stock,” that the current investment environment is fundamentally different from the 2000 telecom bubble, and that the biggest investment opportunities lie in the “bottleneck” areas around GPUs and in edge AI.

• Historical Comparison: Andrew contrasts current AI infrastructure investment with the 2000 telecom bubble. He notes that telecom companies (e.g., Global Crossing) had capital expenditures amounting to 200% (or even 500%) of their operating cash flow, with fiber optic utilization at only 2%. In contrast, the current capital expenditures of the four major cloud providers (Meta, Alphabet, Amazon, Microsoft) account for only 50% of their operating cash flow, consistent with the past five years, and GPU utilization is close to 100%.
• Investment Strategy: Andrew argues that directly challenging NVIDIA’s GPU core is a “daunting task.” Better opportunities lie in the “bottleneck” areas surrounding GPUs: interconnect, memory, and power. Additionally, he believes the “edge AI” story is “far from being fully written” and is in “the first inning,” while the data center is already in “the third or fourth inning.”
• Extrapolation and Risks: Andrew’s concern is that if a single company like NVIDIA captures all the economic value from the entire AI transformation, it would be detrimental to a healthy startup ecosystem. However, he also sees that large cloud providers (e.g., Microsoft, Amazon, Meta) have strong incentives to support AMD, develop their own chips, and invest in startups to build a diversified supplier base.

---

Mentioned Positions

Position	Guest's View	Key Data
NVIDIA	Bullish on its data center dominance, but flags potential weaknesses	Data center business has seen remarkable growth over the past 8 quarters; automotive business is flat; GPU utilization is near 100%
TSMC	Bullish on its moat, which the report argues is difficult to replicate	Annual capital expenditure of approximately $40 billion (comparable to the total CHIPS Act funding); leads in advanced nodes such as 3nm
Intel	Risk warning, the report argues it is caught in an "innovator's dilemma"	Missed two paradigm shifts: mobile and AI; was overtaken by TSMC at the 10nm node
AMD	Neutral to slightly positive, seen as a potential challenger to NVIDIA	Considered an important alternative supplier by major cloud providers; its new chips are catching up in performance
Broadcom	Bullish, seen as a key player in AI infrastructure	Listed by Andrew as one of the most important companies besides NVIDIA and TSMC
SK Hynix	Bullish, seen as a key player in AI memory	Listed by Andrew as one of the most important companies besides NVIDIA and TSMC
ASML	Bullish, seen as the most critical and sensitive link in the supply chain	Its EUV lithography machines are essential for manufacturing advanced chips, with extremely complex technology
CoreWeave	Bullish, seen as an emerging "GPU cloud" representative	Offered H100 instances 6 months earlier than AWS/Azure, demonstrating the agility of a startup
Apple	Neutral, used as a case study	Captured all the economic value from AT&T during the mobile era (market cap from $70 billion to $3.5 trillion vs. AT&T from $250 billion to $150 billion)
Huawei	Neutral, used as a geopolitical case study	A leader in China's AI space, but its chip manufacturing is constrained by SMIC's inability to access the most advanced ASML equipment

---

Judgments Worth Remembering

1. (Andrew Homan) The "Three-Layer Cake" Model of AI Investment: The AI ecosystem is divided into chips/cloud (bottom layer), foundation models (middle layer), and applications (top layer). Currently, the winners in the middle and top layers are unclear, while the bottom layer (chips) will capture significant economic rents in almost any AI development scenario.

• Support: The cost of training frontier models surged from $10 million in 2022 to $1 billion in 2024, and 90% of a frontier model company's costs are computing power.

2. (Chris Miller) Intel's Struggles Are a Byproduct of Its Success: Intel became risk-averse due to its overwhelming success in PCs and data centers, causing it to miss two paradigm shifts: mobile and AI.

• Support: Steve Jobs once asked Intel to manufacture chips for the iPhone, but the request was rejected.

3. (Andrew Homan) NVIDIA's "Death Star" Weakness Lies in Its Generality and High Power Consumption: Directly challenging its GPU core is difficult, but application-specific chips (ASICs) targeting specific model architectures and the power-sensitive "edge" market represent potential breakthroughs.

• Support: NVIDIA's automotive business has been largely stagnant for the past eight quarters, while its data center business has exploded, indicating a heavy concentration of resources on the latter.

4. (Chris Miller) TSMC's Moat Is Taiwan's "Ecosystem": Its advantage lies not only in technology but also in the highly concentrated and efficient manufacturing cluster on the island of Taiwan, which is harder to replicate than the technology itself.

• Support: TSMC's "high-speed rail" can transport personnel and materials between its factories on the island within a single day, a key source of its efficiency.

5. (Andrew Homan) The Current AI Infrastructure Investment Differs Fundamentally from the 2000 Telecom Bubble: Capital expenditures back then were unsustainable (200%-500% of operating cash flow) with extremely low utilization rates (2%). In contrast, current cloud providers' capital expenditures are sustainable (50% of operating cash flow), and GPU utilization is near 100%.

• Support: The capital expenditure/operating cash flow ratio for the four major cloud providers is currently in line with the past five years.

6. (Andrew Homan) Edge AI Is "Game One," While Data Centers Are Already "Game Three or Four": The story of edge AI is far from written, offering a significant window of opportunity for startups.

• Support: Low-power chips are not in NVIDIA's DNA, leaving room for other companies.

7. (Chris Miller) The True Bottleneck for AI Development Is "Computing Cost": The industry is currently "admiring" the scale of massive investments, but the real goal should be to reduce computing costs through technological innovation, which is key to AI's sustainable development.

• Support: If computing costs do not come down, AI development will face fundamental constraints.

8. (Andrew Homan) The "Upgrade" Strategy for Semiconductor Startups: Startups face three major pain points: EDA tools, IP, and tape-out. Maverick Silicon's strategy is to help them connect with key players in the industry chain (e.g., Synopsys, Cadence, TSMC), upgrading them from "economy class" to "first class."

• Support: These major players have an incentive to support startups, as the latter may become their large customers or part of their ecosystem in the future.