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

Episode Summary

Executive Summary: Patrick O’Shaughnessy speaks with Andrew Homan and Chris Miller about semiconductors as the backbone of AI, why Intel and legacy players struggle with disruption, where profits may accrue across chips, cloud, models, and apps, and how geopolitics, capital intensity, and supply-chain bottlenecks will shape the next decade.

Main Topics: Intel as a case study in disruption (Priority: 5/5): Intel shows how success in one era can make firms miss mobile and AI transitions. AI’s semiconductor value chain (Priority: 5/5): The guests map AI into chips, models, and apps, arguing the infrastructure layer is still most durable. TSMC, manufacturing, and yield (Priority: 5/5): Taiwan’s manufacturing ecosystem, specialization, and extreme precision explain TSMC’s advantage. Geopolitics and industrial policy (Priority: 4/5): US, China, Taiwan, and the Middle East all shape chip supply, access, and investment flows. Emerging bottlenecks around power and memory (Priority: 4/5): Memory, interconnect, packaging, and power are becoming constraints as GPUs scale. Semis as an investing opportunity (Priority: 4/5): Private semi investing is seen as undercapitalized, cyclical, and increasingly attractive. Edge AI and new chip markets (Priority: 3/5): Intelligence is moving to phones, cars, wearables, and industrial devices, creating fresh demand.

Key Arguments: Intel’s incumbency made it risk-averse, helping it miss mobile and AI shifts. AI likely shifts huge value to infrastructure because compute demand is exploding. Hyperscalers are designing custom chips to lower cost and reduce dependence on NVIDIA. TSMC’s moat comes from focus plus Taiwan’s dense manufacturing ecosystem. Semis are cyclical, but this cycle looks like a growth cycle with durable demand. Memory, packaging, networking, and power may bottleneck GPU scaling before demand fades. China struggles because no country can build the whole stack alone; specialization matters. Private semis offer attractive returns because capital is scarce in Series B and later rounds.

Data Points: Current AI frontier model training cost progression: "$10 million" in 2022, "$100 million" in 2023, "$1 billion" in 2024, "$10 billion" in 2025, "$100 billion" in 2026 - Illustrates the explosive rise in model-training spend Hyperscaler CapEx as share of operating cash flow: about "50%" - Meta, Alphabet, Amazon, and Microsoft are spending at a sustainable level Telecom bubble CapEx as share of operating cash flow: about "200%" on average; some years "500%" - Shows why the late-90s fiber buildout differed from today’s AI buildout Global crossings / Level 3 / Quest utilization: around "2%" lit up - Fiber in the ground during the telecom bubble was massively underutilized China’s GDP-scale AI spend compared to US big tech: not matched except maybe ByteDance - China is seen as a downside mover in AI investment intensity TSMC CapEx: "almost that much" as the CHIPS Act’s "$40 billion" - Used to show how small government subsidies are versus industry-scale investment Semiconductor industry size today: about "$700 billion" - Current market size estimate Software industry size today: about "$800 billion" - Used for comparison with semiconductors Semiconductor index CAGR: "25% CAGR over the last decade" - Evidence that semis have already performed strongly despite being viewed as a backwater Taiwan’s manufacturing proximity: "within an hour and a half" of each other on the high-speed rail - Part of the ecosystem advantage TSMC benefits from GPU utilization in hyperscaler clouds: "nearly 100%" - Counters fears of idle AI infrastructure High-end chip precision: nanometers vs millimeters; "a thousand times more precise" - Explains why semiconductor manufacturing is uniquely difficult EUV process heat: "100 times the temperature of the sun" - Describes the extreme complexity of advanced lithography Semiconductor startup middle-stage capital need: "$50 million" - TAPE-out, IP, EDA, and emulation costs can add up quickly Public-market leader scale: NVIDIA is now larger than some hyperscalers in scale - Used to explain bargaining power shifts in the ecosystem Number of people to know to stay informed: three key contacts: hyperscaler, TSMC, NVIDIA - The guests emphasize the importance of direct industry access

Pivotal Quotes: "I would describe it as a Hong Kong action film" — Andrew Homan: Used to characterize Intel’s chaotic transition and multiple moving pieces "The semis are challenging. They're technically complex. They are capital intensive." — Andrew Homan: Explains why semiconductors historically deter many investors "The real goal should be finding ways to economize, to bring new technology to bear, to find ways to bring down that cost." — Chris Miller: Focuses on lowering compute costs as a key AI objective

Implications: The next winners will likely be determined by who can ease compute bottlenecks fastest while navigating geopolitics, power, and capital constraints.