What Investors Are Overlooking in AI & Semis | Val Zlatev

Episode Summary

Executive Summary: Val Zlatov argues that hard tech—especially semiconductors—is underappreciated, structurally more investable than in past decades, and central to the AI boom. He explains how Moore’s law, fab complexity, and constrained capacity have reduced old-style cyclicality, while hyperscaler-led AI spend and adjacent supply-chain niches create durable opportunities and clearer short setups.

Main Topics: What Hard Tech Is and Why It Matters (Priority: 5/5): Hard tech is defined as non-software, non-internet technology: semiconductors, comms equipment, and manufacturing tools. Zlatov frames semis as the dominant and most innovative part of a vast TMT universe. Why Investors Underweight Semiconductors (Priority: 5/5): He says the industry shifted from hard tech to software/internet over the 2000s, leaving the semis knowledge base shallow among generalists and many TMT funds despite long-term outperformance. Moore’s Law, Fab Costs, and Reduced Cyclicality (Priority: 5/5): The end of Moore’s law made chip manufacturing far more complex and expensive, with fabs now requiring massive capital and long build times, which dampens historic boom-bust behavior. AI as a Multi-Decade Capex Cycle (Priority: 5/5): Zlatov compares AI more to cloud buildout than the dot-com era, arguing that hyperscalers are funding AI from free cash flow because customers are pulling capacity into production. DeepSeek, AI Demand, and Capacity Constraints (Priority: 4/5): He argues DeepSeek was misread as a revolutionary cost shock; instead, it accelerated broader inference cost declines and increased usage, leaving the market still chip-constrained. Government Involvement: US and China (Priority: 4/5): He criticizes both US restrictions and subsidies, while noting China has made real progress in memory and some domestic semis but still lacks scale at the leading edge. Investment Process and Portfolio Construction (Priority: 5/5): Analog Century uses a four-part framework—growth inflection, winner quality, balance sheet strength, and acceptable valuation—plus a long history database on cycles to build long and short books.

Key Arguments: Semiconductors are the core of hard tech and represent a massive portion of TMT market cap, yet investor knowledge is still unusually shallow relative to software and internet. The sector shifted from secular growth in the 1990s to a more cyclical trading vehicle in the 2000s, which pushed talent and capital away; that dynamic is now reversing. Moore’s law has not disappeared, but the old easy version of it has; atomic-scale manufacturing and quantum effects make leading-edge fabs vastly more expensive and slow to replicate. Because leading-edge fabs now take years and cost tens of billions, companies cannot add capacity as freely, which reduces the severity of classic semiconductor cycles. AI spending is being driven by large hyperscalers with strong balance sheets and real customer demand, making it more analogous to cloud infrastructure buildout than to a speculative bubble. DeepSeek did not prove AI spending was unnecessary; it largely represented another step down in inference costs, which ultimately increases usage when ROI is attractive. The AI supply chain is broader than NVIDIA; networking, assembly, CPUs, ASICs, and other infrastructure providers matter, though many are overlooked. Analog semiconductors are entering a new upcycle after a long inventory digestion period, with automotive and industrial demand increasingly important due to content growth. Many companies claim AI exposure without true product or revenue change; the market can separate real beneficiaries from marketing-driven 'AI' narratives. China’s semiconductor effort is serious and improving, especially in memory, but scale and leading-edge manufacturing still lag the best global players. In hard tech investing, staying power matters more than trading around every earnings print; a sector-specific knowledge base and long dataset create an edge. Team continuity and operational trust are essential because research focus and portfolio construction matter more than back-office micromanagement in generating returns.

Data Points: Hard tech market cap: Over $14–$15 trillion - Estimated market capitalization of the hard tech space within TMT. Large-cap hard tech universe: 500–600 companies above $1 billion market cap - Size of the investable public-company universe in hard tech. Fab cost increase: From about $1 billion to about $50 billion - A leading-edge fab that once cost ~$1B now costs roughly $50B. Fab build time: 5–6 years - Approximate time to build a state-of-the-art semiconductor fab. Transistor density: 2–3–5 billion transistors on a fingernail - Illustrative scale of modern chip complexity. Transistor spacing: About five atoms apart - Why quantum effects now matter in chip manufacturing. US/World equipment spend: $100–$110 billion per year - Current global spend on semiconductor equipment for capacity. TSMC annual capex: $40–$50 billion+ - Cited as the biggest spender among semiconductor manufacturers. Samsung annual capex: $25–$30 billion - Major fab investment cited in the discussion. Intel annual capex: $15–$20 billion - Intel’s investment level described as an attempt to stay competitive. Texas Instruments annual capex: $6–$8 billion - Example of a large but smaller long-tail spender. Cloud capex trend: Up every year since 2012 - Used as a comparison for AI capex growth. Inference cost decline: 5–10x per year - Zlatov’s estimate for how fast AI inference costs are falling. AI investment stage: About year 2.5 - His view that AI is still very early in a multi-decade cycle. Data center cost mix: 85% equipment / 15% power and shell - Breakdown of a data center’s cost structure. NVIDIA networking share: About 15% of data center revenues - Portion of NVIDIA data center revenue tied to networking-related products. NVIDIA accelerator shipments: 5–6 million units - Estimated annual shipments cited for NVIDIA accelerators. Accelerator price: $30,000–$40,000 each - Rough price range for NVIDIA accelerators. Analog revenue decline: Cut down in half - Analog semiconductor company revenues after inventory digestion. Automotive semiconductor content: From $100 to $700–$800 to $3,000 - Growth in semiconductors per vehicle from traditional ICE to EV. Market neutral exposure: At least 70% EDO exposed - Target for factor exposure in the market-neutral strategy. Portfolio size: About 60 companies - Typical combined long/short portfolio size. Long book composition: 25 longs - Typical long book size in the strategy. Short book composition: 35 shorts - Typical short book size in the strategy. Position risk cap: 7% at market - Maximum market exposure per position in the long-short strategy.

Pivotal Quotes: "Hard tech is think of tech technology, but take out software and internet." — Val Zlatov: Definition of hard tech at the start of the interview. "We're in a very, very early stage, like we're a year two and a half in what I think it will be a multi-decade investment cycle." — Val Zlatov: His core view on AI as a durable capex supercycle. "The only way to now the trick here is to make like 70, 80, 100, 1,000 of these chips work as one." — Val Zlatov: Explaining how networking, not just raw transistor scaling, extends performance.

Implications: Listeners should see semis and AI infrastructure as long-duration, research-intensive opportunities, not just trading vehicles. The biggest gains may come from hidden supply-chain enablers, while fake AI exposure and obsolete cyclicals may underperform.