Jim Chanos & Val Zlatev: Long and Short Alpha in AI, Semiconductors, Data Centers, Neoclouds, and Data Centers | MacroMinds Symposium 2026

Jim Chanos argues that AI has become a near-universal investment theme in both equity and credit markets, but draws a cautionary parallel to the internet era by noting that US economic growth and corporate profitability grew at roughly 6 percent per year in the decade before and the decade after the internet was introduced, leaving the macro picture essentially unchanged. He acknowledges the capex boom is agreed upon by bulls and bears alike but says nobody yet knows whether there is a real return on investment at the end of the AI buildout.

Val Zlatev takes a more constructive view, noting that across a universe of over 500 hard tech companies headcount has barely budged or declined while operating profits have risen dramatically, suggesting AI's effects on individual businesses are already visible. He describes himself as net long AI relative to his shorts and considers the roughly equal number of bulls and bears a healthy contrast to the one-sided enthusiasm of the 1990s. He also disputes the fiber glut analogy directly, noting that fiber companies and CLECs spent a combined total of only about 100 billion dollars over five years from 1998 to 2002, and that roughly 70 percent of fiber installation cost was fixed construction labor that incentivized massive overbuilding, a dynamic he does not see repeating today.

Chanos identifies a structural accounting disconnect in the AI supply chain. Companies like Nvidia and GE Vernova recognize revenues and profits immediately when they sell chips and equipment, while hyperscalers such as Alphabet, Microsoft, Amazon, and Oracle capitalize those same costs, with GPUs treated as construction in progress during the 12 to 18 month lag before a data center comes online, meaning no depreciation is recorded during that period. His investment thesis follows directly: be long where chips are produced, not where chips reside. He is not short any semiconductor companies and is instead skeptical of neoclouds and legacy data centers.

Chanos is sharply critical of neoclouds, which he describes as equipment leasing or finance companies rather than high-tech businesses. Even under heroic profitability assumptions and a 10-year chip life, he calculates neoclouds generate only 4 to 6 percent returns on capital in out years, with current deals penciling out at single-digit pre-tax ROICs of 5 to 8 percent even during the best supply environment. He cites Blackstone entering the neocloud business via a new REIT as confirmation that these are finance vehicles and predicts neoclouds and low-return data center developers will be winnowed away within 18 to 24 months as capital stops flowing to mundane business models. Zlatev pushes back partially, arguing that neoclouds like CoreWeave have software and optimization layers above commodity infrastructure and that Nvidia deliberately allocates supply to neoclouds to avoid dependence on only four hyperscaler customers.

On GPU rental pricing, Zlatev notes that prices for chips aged six to eight years were down 20 to 30 percent year on year through December, which he considers normal given that new GPU architectures offer lower cost per token. Since January, however, spot GPU rental prices have risen 40 to 50 percent or more due to supply tightness driven by inference adoption growing faster than GPU supply, though he cautions he does not know whether that increase will continue. He traces the explosion in AI memory demand over the last 12 months to three compounding forces: the shift from chatbot to reasoning models, expanding context windows, and the emergence of AI agents. DRAM and NAND prices have risen four to five times driven entirely by data center demand, and memory as a share of bill of materials for PC and smartphone makers has climbed from roughly 20 percent to roughly 50 percent, pushing PC and smartphone unit volumes down mid-teens year over year and creating shorting opportunities in component makers supplying those manufacturers.

On valuations, Zlatev pushes back against the idea that semiconductors are uniformly frothy. Networking companies trade at around 50 to 60 times forward multiples, which he identifies as the most extreme valuation in the space. Nvidia by contrast trades at approximately 15 times 2027 EPS, or roughly 12 times 2028 EPS following a recent decline, and Zlatev considers Nvidia meaningfully cheaper than Intel, which has not made money in a couple of years. Semiconductor equipment companies such as Lam Research and ASML trade at approximately 35 to 40 times earnings, which he views as defensible given their monopoly or semi-monopoly positions, though less attractive than Nvidia. Chanos adds a personal caveat that in 40 years of investing he has never successfully made money being short DRAM companies because the business is highly cyclical and timing is nearly impossible.

Zlatev identifies the empirical nature of AI scaling laws as the key risk to the entire AI infrastructure investment thesis. These laws, which hold that larger clusters and more compute produce higher-quality outputs, have existed for approximately 12 years but are not grounded in physics or mathematics and could in principle be broken by a new architecture. He notes that DeepSeek did not break scaling laws but rather combined already-known algorithms to reduce token costs, representing an incremental step rather than a paradigm shift. The host adds that the current situation in which semiconductors earn large profits while model companies report large losses is unsustainable but could persist for a year or more, and cautions that technological booms frequently last longer than many people expect.