What Investors Are Overlooking in AI & Semis | Val Zlatev

Episode Summary

Executive Summary: Val Zlatov argues hard tech—especially semiconductors—has shifted from a cyclical trading space back toward a secular investment theme, driven by Moore’s Law constraints, huge fab costs, and AI demand. He says investors still underappreciate the sector, misread AI cycles, and often mistake marketing for real product impact. The conversation also covers U.S./China industrial policy, DeepSeek, AI supply-chain winners, and Analog Century’s long-short and market-neutral process.

Main Topics: What hard tech is and why semiconductors dominate (Priority: 5/5): Zlatov defines hard tech as hardware outside software/internet, including semiconductors, networking gear, and semiconductor equipment. He emphasizes that semiconductors are the largest, most innovative, and most economically important part of the universe. Why Wall Street underweights hard tech (Priority: 5/5): He explains that talent and attention shifted to internet/software after the 1990s because those were the secular growth drivers, while semis became more cyclical and harder to model due to fragmented supply chains and many different business models. Moore’s Law, capex inflation, and reduced cyclicality (Priority: 5/5): Zlatov argues the end of Moore’s Law has made semiconductor manufacturing vastly more expensive and slower to expand, reducing boom-bust oversupply behavior and making the sector more investable as a long-term theme. AI as a multi-decade demand cycle (Priority: 5/5): He compares AI more to cloud computing than the dot-com boom, saying the biggest hyperscalers are funding AI from free cash flow because customers are pulling demand through real ROI, not speculation. DeepSeek and misconceptions about efficiency (Priority: 4/5): He says the market initially misread DeepSeek as a revolutionary cost break, but it was more of an incremental step in inference-cost reduction that ultimately increased usage and demand for chips. Government policy in chips and the China race (Priority: 4/5): He discusses U.S. export controls, the CHIPS Act, and Chinese state support. He is skeptical of government picking winners but believes China is serious, well funded, and capable of catching up in select areas. Analog Century’s investment framework and portfolio construction (Priority: 5/5): The fund uses a framework based on inflecting growth, winner identification, balance sheet strength, and reasonable valuation, supported by a 20-year proprietary dataset to model cycles and construct long-short and market-neutral portfolios.

Key Arguments: Hard tech is broader than semiconductors, but semis are the core of the space because they power nearly all electronic devices and represent the biggest share of innovation and market value. Investors missed semis because the 2000s rewarded software/internet, causing knowledge depth in hard tech to atrophy while semiconductor supply chains remained fragmented and difficult to analyze. Moore’s Law’s physical limits have made fabs and equipment much more expensive, slowing capacity additions and reducing the severity of historical semiconductor cyclicality. AI spending resembles the cloud buildout: large hyperscalers are investing from free cash flow because enterprise customers want the capability and ROI is visible. DeepSeek did not change the economics as radically as the market initially thought; instead, it accelerated a broader downward trend in inference costs, which increased usage. The main AI winners are not only chip designers; networking, server assembly, power management, and other data-center infrastructure names can also benefit. Many companies are “faking” AI exposure through marketing labels like AI PCs or AI smartphones, but end-demand has not yet justified the claims. Government industrial policy can provide carrots and sticks, but bureaucrats are poor at selecting winners versus the market; some recipients of CHIPS Act support are struggling businesses. China is behind in semis overall, but the country has made meaningful progress in memory, digital chips, and manufacturing capacity, and should be taken seriously. Analog Century’s process relies on a 20-year cycle database and a framework emphasizing inflecting growth, market share, balance sheet quality, and valuation. The firm’s market-neutral strategy grew from investor demand for portable alpha, using much of the same research as the long-short book but with different sizing and risk constraints. Long-term conviction matters more than trading quarterly earnings; the firm aims to compound alpha through durable positions rather than constantly flip names around earnings.

Data Points: Hard tech market cap: over $14-$15 trillion - Zlatov estimates the total market capitalization of the hard tech universe Number of hard tech companies over $1B market cap: 500-600 companies - Approximate count in the hard tech universe Fab cost in 2000: about $1 billion - Historical cost of a state-of-the-art semiconductor fab Current fab cost: about $50 billion - Current cost for a leading-edge semiconductor fab Semiconductor equipment spend (current year): $100-$110 billion - Worldwide annual equipment spend for capacity buildout Total CapEx project spend: about $150 billion - Includes equipment plus bricks-and-mortar and related costs TSMC annual spend: $40-$50 billion - One of the largest ongoing capex spenders in semiconductors Samsung annual spend: $25-$30 billion - Estimated yearly semiconductor investment Intel annual spend: $15-$20 billion - Intel’s attempt to re-invest and catch up Texas Instruments annual spend: $6-$8 billion - Smaller but meaningful capital spender AI use-case rollout: about 25% or less - Zlatov says current hyperscaler capacity supports only a minority of AI use cases Inference cost decline: 5x to 10x per year - He says AI inference costs are falling very rapidly as models improve DeepSeek inference step-down cadence: every 3-4 months - He characterizes DeepSeek as one step in a recurring cost-reduction pattern NVIDIA accelerator sales: 5-6 million units this year - Approximate annual volume mentioned for AI accelerators NVIDIA accelerator price: $30,000-$40,000 each - Illustrative pricing for high-end accelerators Data center cost mix: 85% hard-tech equipment / 15% shells and utilities - He emphasizes where most of the money is spent in AI data centers Analog Century portfolio size: about 60 companies - Typical combined portfolio size across strategies Long-short portfolio mix: 25 longs / 35 shorts - Approximate composition of the main long-short book Factor exposure target in market neutral: at least 70% idiosyncratic exposure, no more than 30% style factors - Risk construction goal for market-neutral strategy Balance sheet/position risk limit: no single long above 7% at market - Example of portfolio risk management constraint Chinese memory share: about 20% of world memory capacity - Zlatov says Chinese firms have become increasingly relevant in memory

Pivotal Quotes: "Hard tech is think of tech technology, but take out software and internet." — Val Zlatov: Defining the sector at the start of the interview "The reality is actually very different. So the reality, I'd like to make a different parallel for AI. I think AI right now is much closer to the cloud development." — Val Zlatov: Explaining why AI should be compared to cloud adoption rather than the dot-com bubble "The only way to really now extend Moore's Law is not from making the chips bigger to be more powerful. This doesn't work anymore because they have reached what's called the radical limit." — Val Zlatov: Describing why semiconductor manufacturing has become capital-intensive and why cyclicality is changing

Implications: Listeners should see hard tech as a long-duration thematic space, not just a trading sector. AI demand, supply-chain constraints, and rising fab complexity may support durable winners, while many “AI” labels will prove superficial.