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

Jim Chanos describes the AI investment boom as a near-universal consensus trade dominating both equity and credit markets, while acknowledging that nobody yet knows whether there will be a meaningful return on investment from AI spending. He contextualizes the uncertainty by noting that US economic growth and corporate profitability, running at roughly 6 percent per year, were virtually unchanged in the decade before and after Netscape's introduction around 1995 to 1996, suggesting the internet did not meaningfully accelerate overall growth. He also notes that S&P 500 operating profits rose approximately 30 percent from mid-1998 to mid-2000 before dropping roughly 40 percent by 2001, a decline matching the Global Financial Crisis, and says current S&P earnings growth is running at a similar or slightly higher clip than that period.

Val Zlatev takes a more constructive view, noting that AI effects are already visible across more than 500 hard tech companies he tracks, where headcounts have barely moved or declined while operating profits have risen dramatically. He also observes that there are roughly as many AI bears as bulls, which he views as healthier than the one-sided enthusiasm of the 1990s, and he describes himself as net long AI relative to his shorts.

Chanos identifies a structural accounting disconnect in the current capex boom: companies selling chips and data center equipment recognize revenues and profits immediately, while hyperscalers capitalize those same costs and depreciate them over time, with a lag of 12 to 18 months between spending and revenue generation. He argues that neoclouds such as CoreWeave are effectively equipment leasing or finance companies rather than high-technology companies, because they buy chips from Nvidia, rent data center space, and then rent chips to hyperscalers. Even under heroic profitability assumptions and a conservative 10-year GPU life, he calculates neoclouds generate only approximately 4 to 6 percent returns on capital in the out years, with deals where granular economics are available penciling out at single-digit pre-tax ROICs of 5 to 8 percent even during peak capacity shortages. His stated investment preference is to be long where chips are produced rather than where chips reside.

Zlatev partially pushes back, noting that GPU rental prices for older chips were down 20 to 30 percent year on year until December but have since risen 40 to 50 percent or more since January, which he attributes to token usage growing faster than available GPU supply. He distinguishes neoclouds like CoreWeave and Nimbius, which have software and optimization layers above commodity infrastructure, from legacy colocation providers like Equinix or Digital Realty. He acknowledges he has only invested in neoclouds on the long side and agrees with Chanos that the real technology value in the AI infrastructure chain resides in chips and chip wrappers, not in land, power access, or transformer installation. Zlatev predicts power and labor bottlenecks will likely persist for the next 18 months but are unlikely to remain constraints three years from now, and he argues equities should be priced on core business over a full cycle rather than on current spot prices during a shortage.

Chanos dismisses the space data center thesis by noting that power costs are only approximately 5 to 7 percent of data center revenue, that heat radiation is a major engineering problem in a vacuum, and that redundancy and maintenance costs are prohibitive because replacing broken components requires additional launches. Zlatev frames Musk's argument differently, saying it is not about cost savings but about the sheer volume of compute Musk believes will be needed, specifically one terawatt compared to the roughly 15 gigawatts of capacity being built by current hyperscaler capex of approximately 750 billion dollars. Chanos notes SpaceX is approaching a valuation of nearly two trillion dollars, that Starlink earns approximately four billion dollars annually in operating profit on an estimated 25 to 30 billion dollars of invested capital with growth slowing and prices cut to drive unit growth, and that SpaceX's launch business still loses money despite billions spent.

Zlatev argues that semiconductor equipment makers like ASML and Applied Materials face a hard physical ceiling on capacity growth of 30 to 35 percent per year due to supply chain complexity, and that memory makers compounded this constraint by never building additional clean room space during the downturn through early 2025. DRAM and flash prices have since risen four to five times and are expected to continue rising at roughly 30 percent per quarter. The shift from chatbots to reasoning models and AI agents has pushed memory from roughly 20 percent of the bill of materials for PCs and smartphones to approximately 50 percent, with unit volumes in both markets currently down mid-teens year over year. Zlatev sees shorting opportunities in component makers that supply PC and smartphone manufacturers but lack pricing power and are absorbing those unit declines.

Zlatev places the most exaggerated valuations on the networking side, where companies trade at roughly 50 to 60 times forward multiples, while Nvidia trades at approximately 15 times 2027 earnings and Broadcom at approximately 12 times 2028 earnings. Chanos states he has never made a single dollar being short a DRAM company in 40 years of investing because timing is nearly impossible, and he is not currently short any semiconductor companies.