56:05
Accounting Mismatch in AI Profits | Jim Chanos and Val Zlatev on Long/Short Alpha in AI & Semis
The Monetary Matters Network
Overview
This video features a discussion between investors Jim Chanos and Val Zlatev on the accounting and investment landscape surrounding the AI boom, particularly focusing on semiconductors and data centers. They explore the disconnect between companies selling AI infrastructure (picks and shovels) and those consuming it (hyperscalers), highlighting how differing accounting practices can inflate perceived profits. The conversation delves into the sustainability of current AI-driven capital expenditures, historical parallels with the dot-com bubble, and potential long and short investment opportunities within the AI ecosystem, including data centers, memory chips, and semiconductor equipment manufacturers. They also touch upon the speculative nature of space-based data centers and the valuation challenges in a rapidly evolving technological market.
How was this?
Save this permanently with flashcards, quizzes, and AI chat
Chapters
- The current AI boom is dominating financial markets, but its broader economic impact is uncertain, similar to the pre- and post-internet eras.
- While overall economic growth may not see a dramatic increase, there will be significant winners and losers within the AI economy.
- A key accounting disconnect exists: companies selling AI hardware (like chips) recognize immediate profits, while hyperscalers capitalizing these costs delay profit recognition.
- Historical parallels with the late 1990s tech boom suggest that rapid capital expenditure cycles can lead to sharp profit contractions.
Understanding this accounting difference is crucial for accurately assessing the true profitability of companies involved in the AI build-out and for anticipating potential market corrections.
During the dot-com bubble (1998-2001), S&P operating profits surged 30% in two years, only to drop 40% in the subsequent year, illustrating the volatility of tech-driven profit cycles.
- On a micro-level, AI is already having a significant positive impact on individual businesses, evidenced by increased operating profits with stable or declining headcount.
- The market sentiment around AI is healthy due to the presence of both bulls and bears, fostering critical evaluation rather than blind optimism.
- Despite the excitement, there's a cautionary note that exponential growth forecasts in technology often face real-world constraints.
- Experienced semiconductor industry professionals are generally more cautious than their Silicon Valley counterparts.
This chapter highlights that while AI is demonstrably boosting company profits now, historical patterns and industry caution suggest a need for realistic expectations regarding future growth.
Many tech companies have seen operating profits increase dramatically over the last few years while their headcount has barely budged or even decreased, indicating AI-driven efficiency gains.
- The discussion identifies inherently unprofitable business models within the AI ecosystem that are unlikely to generate significant returns on capital.
- These include entities like Bitcoin miners turned data center developers and 'neo-clouds' that lease out computing resources.
- Even with optimistic assumptions about chip longevity and profitability, these businesses struggle to achieve attractive returns.
- The core investment thesis favors owning the 'picks and shovels' (chip manufacturers) over the infrastructure providers or intermediaries.
Recognizing these speculative or fundamentally weak business models is key to avoiding investment pitfalls and focusing capital on more sustainable parts of the AI value chain.
Companies like Bitcoin miners or 'neo-clouds' that lease GPUs may only achieve 4-6% returns on capital even under best-case scenarios, making them less attractive investments.
- The capitalization and depreciation of data center assets, particularly GPUs, create a lag before costs impact reported earnings.
- The 'neo-cloud' or data center landlord model is viewed more as an equipment leasing or finance business than a high-tech operation.
- While some 'neo-clouds' incorporate software layers, the core technological value resides in the semiconductor chips themselves, not the data center infrastructure.
- The current tightness in GPU supply has temporarily improved economics for some data center providers, but this is a dynamic and potentially short-lived situation.
This distinction between the technology providers and the infrastructure 'landlords' helps investors differentiate between businesses with true technological moats and those acting as financial intermediaries.
Companies like Coreweave are likened to equipment leasing or finance companies, betting on chip lifespan and rental contracts, rather than core technology innovators.
- The high costs associated with launching data centers into space are a significant barrier.
- Power is a relatively small component of data center costs on Earth, so leveraging free solar power in space offers limited financial advantage.
- Challenges like radiation, heat dissipation (radiators), and the difficulty of maintenance and redundancy in space make it impractical.
- The ambitious projections for space-based data centers are often tied to unproven launch technologies, like SpaceX's Starship, which has a history of failures.
This section debunks the hype around space-based data centers by focusing on practical engineering and economic realities, steering investors away from highly speculative ventures.
The need for massive radiators to dissipate heat in a vacuum and the logistical nightmare of launching replacement parts via rockets make space data centers economically unviable compared to terrestrial ones.
- The semiconductor equipment manufacturing sector faces physical constraints, limiting annual growth to around 30-35% due to supply chain complexities and long build times for facilities.
- Memory (DRAM and NAND) prices have surged due to increased demand from AI data centers, significantly impacting consumer electronics costs.
- While memory companies' stock valuations appear low, this reflects market expectations of an imminent downturn, which may be delayed by supply constraints.
- Valuations in the semiconductor space are not universally frothy; some areas like networking are expensive, while others like memory appear cheap, and dominant players like Nvidia are relatively more reasonably valued than competitors in crowded markets.
Understanding the supply constraints in semiconductor manufacturing and the cyclical nature of memory markets is crucial for assessing current valuations and future profitability.
The cost of memory as a component in PCs and smartphones has risen from 20% to 50% of the bill of materials, leading to price increases for consumers and potential unit sales declines.
- History suggests that exponential growth rates in technology, like internet traffic in the late '90s, can be overestimated, leading to mispriced investments.
- The current AI boom's trajectory is being driven by empirical scaling laws, which are not immutable physical laws and could change with new architectures.
- Investors should be wary of applying 'magical' valuations to mundane businesses, as intense capital flow into the AI space tends to reduce overall returns.
- Opportunities exist on both the long and short sides of the AI market, but careful analysis is needed to distinguish sustainable businesses from speculative ones.
Drawing lessons from past technological booms and busts is essential for navigating the current AI investment landscape with a critical and realistic perspective.
The belief that internet traffic was doubling every quarter in the late 1990s led to massive overinvestment, a lesson applicable to current AI compute demand forecasts.
Key takeaways
- The AI boom presents a complex investment landscape where accounting methods can obscure true profitability, and historical parallels suggest caution.
- Focus on the 'picks and shovels' of the AI revolution (semiconductor manufacturers) rather than the 'landlords' or intermediaries (data center operators) for potentially more sustainable returns.
- Be skeptical of extremely high growth forecasts, especially when they are based on empirical scaling laws that could be disrupted by technological innovation.
- The semiconductor industry faces physical supply constraints that limit rapid capacity expansion, influencing pricing and profitability.
- While memory prices have surged, the long-term outlook is cyclical, and current low valuations may reflect anticipated downturns.
- Valuations across the tech sector are uneven; avoid overpaying for companies in highly competitive markets, even if they are growing rapidly.
- The sustainability of the current AI profit dynamic, where hardware providers profit while AI model companies potentially incur losses, is a key area to monitor.
Key terms
HyperscalersCapitalization (Accounting)DepreciationPicks and Shovels (Investment Metaphor)Neo-cloudsGPU Rental PricesScaling Laws (AI)Bill of MaterialsDRAMNAND FlashCapEx BoomOperating Profit
Test your understanding
- How does the accounting treatment of costs differ between companies selling AI hardware and those consuming it, and why does this difference matter for investors?
- What historical parallels does Jim Chanos draw to the current AI boom, and what lessons can be learned from those periods?
- Why are 'neo-clouds' and data center 'landlords' considered less attractive investments compared to semiconductor manufacturers in the AI ecosystem?
- What are the primary practical and economic challenges that make space-based data centers currently unfeasible?
- How do supply chain constraints and the cyclical nature of the memory market influence the valuations of semiconductor companies?