PodcastThe Monetary Matters NetworkJim Chanos

Jim Chanos and Val Zlatev on Long/Short Alpha in AI, Semis, & Data Centers | MacroMinds Symposium

Speakers see persistent supply constraints and conservative memory producers limiting near‑term capacity, even as empirical AI scaling laws continue to drive compute demand. That mismatch has driven GPU rental spikes, a roughly 40–50% price reversal, and a 4–5x memory surge that markets appear to be mispricing as a near‑term rollover. The panel warns of concentration risk—markets and the U.S. economy are increasingly a bet on AI—while neo‑clouds and leasing layers face low single‑digit ROICs, and semiconductors look constructive but carry bubble risks. They expect a market rotation within 18–24 months that will separate high‑return winners from low‑return, capital‑intensive bets.

Artificial IntelligenceSemiconductorsAI InfrastructureSupply ChainsInvestment StrategyMarket Breadth

Preview · 20 more points in the app

Read the Full Summary

Key Takeaways

Physical supply constraints—equipment vendors can only scale shipments ~30–35% annually and fabs take years and huge capital—help explain the 4–5x surge in DRAM/NAND driven by datacenter AI; speakers expect elevated memory demand to persist for a shallow multi‑year window (roughly 2–4 years), making the market’s pricing that expects a sharp rollover in 6–9 months (forward multiples ~6–7) likely optimistic and shorting DRAM timing‑risky.

Empirical AI scaling laws (more compute → better models) are driving explosive token usage that is straining GPU capacity and pushing rental prices up roughly 40–50%; Nvidia is deliberately allocating chips across hyperscalers and new cloud/neocloud providers (e.g., CoreWeave) to avoid excessive concentration, effectively routing chips through intermediaries rather than selling only to the largest buyers.

Neo‑cloud operators function economically like equipment‑leasing/finance firms—buying chips, renting powered rack space and leasing compute—and while tight supply recently improved near‑term economics, current powered‑capacity deals are producing single‑digit pre‑tax ROICs (~5–8%), implying the leasing/middleman layer may offer inferior long‑term returns relative to other stack segments.

Accounting differences amplify apparent supplier strength: picks‑and‑shovels vendors recognize revenue and profit immediately while hyperscalers capitalize chips, labor and interest for ~12–18 months, mechanically inflating supplier margins even as at least one large model company reported 'absolutely staggering' operating losses in 2025—supporting the view that supplier profitability can diverge from model‑company economics.

The U.S. economy and equity market are an increasingly concentrated bet on AI—speakers warn of large asymmetric outcomes (huge gains if AI succeeds, large losses if it doesn’t)—and they note historical tech waves (e.g., Netscape/internet era) changed sectoral dynamics without reliably boosting long‑run macro growth, with a market rotation expected to reveal winners and losers within 18–24 months.

TECHNOLOGY

Supply constraints — Equipment vendors and memory suppliers are cautious and capacity‑constrained. Equipment makers cannot scale shipments much faster than ~30–35% a year, and fabs take years and large capital to build. Memory producers went through a downturn that made management conservative, limiting near‑term capacity growth despite strong demand.

“it might happen but history tells us that that these kind of insane exponential growth rates tend to get constrained by the real world…”—

AI scaling laws — Speakers say empirical scaling relationships (larger clusters and more compute → better model performance) continue to drive compute demand. The January 2025 'Deep Sea' episode reduced token costs by recombining known algorithms but did not overturn those scaling laws, though a new architecture could change the empirical relationships.

“What that really means the scaling loss in AI are the bigger the cluster the more compute you use to train a code the better the output the higher the…”—

Nvidia allocation — Nvidia deliberately balances supply between hyperscalers and new cloud/neocloud providers (e.g., CoreWeave) to avoid concentration with a few large customers. The company effectively routes chips through intermediaries rather than selling only to hyperscalers.

“And Nvidia wants to create a balance between the hyperscalers and the new clouds because Nvidia doesn't want to get beholden on four customers in perpetuity.”—

ECONOMY

Tech ≠ macro growth — Looking at the decade before and after the Netscape/internet inflection, U.S. economic growth and corporate profitability did not show a sustained increase. Major tech waves can cause sectoral disruption without materially changing long‑run macro growth or aggregate corporate profit growth rates.

“we took a look at the decade that preceded the introduction of Netscape in late '95 early 96 to the decade the 10 years that had happened after Netscape the…”—

MARKETS

GPU scarcity — Exploding token usage is straining available GPU capacity and pushing rental prices for even older GPUs up. Near‑real‑time token metrics correlate with the rent spike, indicating infrastructure scarcity is driving compute rental inflation.

“These chips are so tight as we speak that the prices rental prices for GPUs which are really old like six, seven, eight years old are going up in price…”—

Price reversal — GPU rental prices, after a normal year‑over‑year decline into December, tightened sharply since January and jumped roughly 40–50% or more. Newer GPU architectures are more efficient (lower $/token), which should favor replacements, but tight supply changed near‑term economics for neo‑clouds.

“They started to increase by the way into December these prices were down 20 30% year on year which is very normal.”—

Neo‑cloud leasing — Neo‑cloud operators act like equipment‑leasing/finance companies: they buy chips, rent powered data‑center capacity and lease compute. Their economics resemble finance/leasing bets on chip life rather than classic high‑margin tech firms; some (e.g., CoreWeave, Nimbus) add software/optimization layers and derive spot revenue from inference.

“And remember in this business model, these are equipment leasing companies…”—

FORECASTS

18–24 month rotation — Speakers expect capital inflows into AI and semiconductors will eventually stop flowing to companies with mundane, low‑return business models. They expect a market clarification (a rotation) that reveals winners and losers within the next 18–24 months.

“One of the things we do know is capital is flowing immensely into this space and that tends to reduce returns…”—

Trailer

Trailer highlights acute GPU tightness and rising rental prices, flagging skepticism about unchecked exponential compute demand and industry caution.

Jack's Introduction

Host Jack introduces the panel, the MacroMinds Symposium context, and guests Jim Chanos and Val Zlatev, framing the long/short AI debate.

"Finding Value On Both the Long and Short Side in AI"

Panelists debate where to find alpha in AI: both long and short opportunities exist amid meaningful micro impacts but uncertain macro growth effects.

"There's A Disconnect In The Profitability" Between Semis and Hyperscalers

Speakers emphasize an accounting disconnect: semiconductors and vendors recognize profits while hyperscalers capitalize spending, risking sharp earnings reversals when depreciation catches up.

NeoClouds and Coreweave

NeoClouds (CoreWeave, Nimbus) are framed as equipment-leasing finance plays with low single-digit pre-tax ROICs and concentration risk versus owning the chips themselves.

SpaceX & Data Centers in Space

Speakers are skeptical about data centers in space — launch, radiation, redundancy and unproven Starship economics make space-based data centers a long-term, speculative bet.

DotCom Bubble Analogy on Assumptions of Secular Hypergrowth

They compare today's narrative to the late‑90s dotcom cycle: rapid scaling claims can mislead capex plans and create volatile order-book-driven profit cycles.

Potential "Nightmare" Scenario for Longs?

A true nightmare for longs would be a new AI architecture that breaks current empirical scaling laws, collapsing demand for massive compute and picks‑and‑shovels revenues.

Huge Bull Market in Memory

Memory prices have surged 4–5x driven by AI storage and context demands; supply constraints and conservative capacity expansion make an imminent sharp rollover unlikely.

Semi Cap Equipment Names

Semiconductor equipment makers are high-quality, oligopolistic businesses but their growth is supply‑constrained (~30% p.a. max), making some current valuations relatively rich versus peers.

Jack's Conclusion

Jack closes urging balanced exposure: semiconductors have long-side value (Nvidia highlighted) but beware concentration risk, and expect a 18–24 month rotation separating winners from capital‑intensive losers.

Related Episodes

Important Disclaimers

Not Investment Advice: For education only — not financial, tax or legal advice.

AI-Generated: Made by AI from source transcripts and may contain errors.

No Guarantees: Past performance does not predict future results; investing carries risk.

Our Commentary · Request Removal