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

Episode Summary

Executive Summary: Jim Chanos and Val Zlotev debated where value may lie in the AI boom, agreeing that AI is real and already affecting company profits, but disagreeing on where returns will accrue. Chanos argued the best long exposure is in chips/picks-and-shovels, while short ideas are the “middlemen” and speculative data-center/neocloud models. Zlotev emphasized supply constraints, rising GPU/memory prices, and continued upside for semis, especially NVIDIA.

Main Topics: AI as a market-defining capex boom (Priority: 5/5): Both speakers framed AI as the dominant force in equities and increasingly credit markets, but warned that its broader macro benefits are still uncertain. Where profits accrue in the AI stack (Priority: 5/5): Chanos argued that returns concentrate in chip makers and certain infrastructure providers, while neoclouds and data-center intermediaries may earn poor long-term returns. Depreciation, accounting, and inflated near-term earnings (Priority: 5/5): The panel discussed how capex is capitalized and depreciated later, which can make current profits for hyperscalers and suppliers look stronger than the eventual economics justify. Memory, DRAM, NAND, and supply constraints (Priority: 4/5): Zlotev argued memory is tight, prices are rising sharply, and physical/equipment bottlenecks limit rapid supply response, supporting the bull case for semis. Valuation dispersion within semiconductors (Priority: 4/5): They contrasted expensive but high-growth names with cheaper dominant franchises, arguing the sector is not uniformly frothy the way it was in 1999-2000. Skepticism toward space data centers and SpaceX-style narratives (Priority: 3/5): Chanos dismissed the idea that data centers in space solve the economics of AI, highlighting launch, radiation, maintenance, and redundancy problems. Long-term uncertainty of AI scaling laws (Priority: 4/5): The conversation closed on the risk that AI demand or scaling laws may change, which would dramatically alter token demand, GPU needs, and the whole investment case.

Key Arguments: AI is already lifting profits for hardware vendors, but the impact on long-run GDP and aggregate corporate profitability is still unproven. In capex booms, sellers recognize revenue immediately while buyers capitalize costs and delay depreciation, overstating near-term earnings power. Chanos believes the best way to express AI enthusiasm is to own the chips, not the companies that merely rent or house them. Neoclouds resemble finance/leasing businesses more than true tech businesses, and their returns on capital may remain only single digits. Zlotev argued GPU rental prices have risen sharply because supply is tight and token usage is expanding rapidly. Memory pricing strength is driven by AI needs such as reasoning models, longer context windows, and agents that require more storage. The semiconductor supply chain cannot expand instantly because fab equipment output is constrained and new facilities take years to build. The AI boom is not uniformly overvalued; within semis there is significant valuation dispersion between dominant and competitive businesses. Space data centers are not economically compelling on power alone because power is a small share of data-center cost, and launch/radiation/reliability issues dominate. The biggest risk to current bullish semiconductor assumptions is a new AI architecture or scaling-law break that sharply reduces compute and memory demand.

Data Points: U.S. economic growth comparison: Virtually the same before and after the internet era - Chanos cited a comparison of the decade before Netscape and the decade after, arguing AI may not boost aggregate growth as much as bulls expect. Corporate profitability growth: 6% per year in both decades - Chanos said profitability growth did not meaningfully accelerate after the internet introduction. GPU useful-life assumption in modeling: 10 years - Chanos said his models assume a 10-year life for GPUs as a conservative depreciation assumption. Estimated spot delay for data-center revenue: 12 to 18 months - Chanos noted construction-in-progress lags before data centers come online and begin depreciating/revenue generation. Neocloud ROIC under heroic assumptions: 4% to 6% - Chanos argued even optimistic modeling for data-center/neocloud businesses produces only low single-digit returns on capital. Hyperscaler/AI spend this year: About $750 billion - Zlotev compared current industry spending to Musk’s compute ambitions. Compute capacity implied by current spend: About 15 gigawatts - Zlotev translated the current spend into power terms. Musk’s implied compute need: 1 terawatt (1,000 gigawatts) - Used to illustrate how far Musk’s long-term AI compute ambitions exceed current buildout. U.S. electric grid size: About 1.5 terawatts - Zlotev noted Musk’s vision would require near full-grid-scale capacity. Power cost share of data center costs: 5% to 7% of revenue - Chanos argued power is a relatively small part of total data-center economics. Memory price increase: 4x to 5x - Zlotev said DRAM/flash prices had risen dramatically due to AI-driven demand. GPU rental price move: Up 40% to 50%+ recently - Zlotev said rental prices for older GPUs had surged since January because of supply tightness. Pre-Dec AI rental trend: Down 20% to 30% year over year - Zlotev said GPU rental prices were falling before the recent squeeze, which is normal over time. Equipment makers' annual growth ceiling: 30% to 35% - Zlotev argued semiconductor equipment suppliers cannot grow shipments faster than this due to physical/supply-chain constraints. PC/smartphone memory bill of materials: About 20% historically; about 50% now - Zlotev said rising memory costs are forcing higher end-user prices. Expected unit demand impact: Mid-teens decline - Zlotev estimated PC and smartphone unit volumes are down mid-teens because higher prices suppress demand. Current revenue multiple examples: NVIDIA about 15x; Broadcom about 12x; memory at 6x-7x - The panel used these figures to show valuation dispersion within semis. 1999-2000 Cisco valuation: Around 160x P/E - Zlotev contrasted today’s more selective valuations with the dot-com bubble. S&P 500 operating profit decline after tech capex peak: Down about 40% - Chanos cited the 2000-2001 collapse following the telecom/fiber capex boom. Network traffic growth myth: Doubling every quarter vs. actually doubling every year - Chanos recalled a mistaken late-1990s narrative that helped fuel overinvestment.

Pivotal Quotes: ""You want to be long with the chips produced, not where the chips reside."" — Jim Chanos: Chanos summarizing his preferred AI exposure: semiconductors over neocloud/data-center intermediaries. ""I think people should be a little bit careful about extrapolating much broader impact to global economic growth and earnings growth."" — Jim Chanos: Chanos warning against assuming AI will automatically raise macro growth and profitability. ""The reality, though, is that these chips are so tight, as we speak..."" — Val Zlotev: Zlotev explaining why GPU rental prices and AI infrastructure economics remain strong near term.

Implications: The AI trade still offers opportunity, but selection matters: semis and scarce infrastructure may benefit most, while leveraged intermediaries and hype-driven models could disappoint. Investors should watch depreciation, supply constraints, and any break in AI scaling laws.