AI Computing Power: The Market Is Simultaneously Betting on Both "Shortage" and "Overcapacity" Bear Theses
nashnova research
How can two opposite bear cases coexist?
Bear case one: compute shortage — power and data-center buildout can't keep pace with 2027 capex plans. Morgan Stanley's Stephen Byrd estimates the U.S. faces a 38GW net power gap through 2028.
Bear case two: compute glut — massive overproduction destroys lease economics, and returns on existing chip inventory collapse.
These two narratives are factually incompatible: supply cannot be both insufficient and excessive. In plain terms = the market doesn't need to pick a bear story — it just needs whichever excuse is trending that day to cut exposure.
Privorotsky's read: "When positioning is crowded, every new worry becomes another reason to de-risk."
Is demand actually strong?
Hyperscaler earnings say yes, overwhelmingly: Azure grew 48%, AWS 39%, Google Cloud 103% quarter-on-quarter.
Combined cloud revenue for Amazon, Alphabet, and Microsoft hit 54% year-on-year growth in Q3 — the sixth consecutive quarter of acceleration. Privorotsky calls it "extraordinary demand."
This means → demand-side data does not support the bearish narrative. The sell case isn't "nobody's buying" — it's "the bill is too large."
How large is the bill?
Goldman's Lee Coppersmith puts a number on it: hyperscaler capex is projected to reach $1.2 trillion by 2027.
Combined free cash flow already turned negative in Q3, burning roughly $31 billion in a single quarter. This means → the next wave of mega-investment will be funded primarily through the bond market, not operating cash flow.
AI-related off-balance-sheet debt has ballooned to $4.1 trillion, and financing demand is getting increasingly crowded. In plain terms = the demand is real, but so is the debt taken on to meet it — whether returns can cover the interest cost is the unresolved core question.
What does Intel's crash actually tell us?
Intel plunged 29% in a single day, becoming a snapshot of market sentiment.
Privorotsky's view: Intel's warning "is more about consumer PCs, margins, and the foundry business — not AI demand itself" — yet the market traded it as an AI-thesis casualty.
This reflects a deeper signal: the market is not distinguishing between "what specifically went wrong." Any bad news gets pinned on the AI trade.
How much positioning has actually been cleared?
Goldman's systematic long-short strategies posted a 2.17% weekly drawdown — the worst week since December 2023. Single stocks have seen net selling for seven consecutive weeks.
Morgan Stanley found that hedge-fund AI exposure has dropped to its first underweight since early 2023. The market just went through one of the largest sustained de-leveraging episodes in over fifteen years.
But Coppersmith is cautious: despite low net exposure, gross leverage still sits at 314% — the 97th percentile over the past five years. In plain terms = on the surface, positions have been trimmed significantly, but the underlying leverage hasn't been fully unwound — fuel for further volatility remains.
Will there be a year-end rally?
Privorotsky forecasts a "painful rally" into year-end — under-positioned, low participation, hard to chase.
But he simultaneously raises the core question: with free cash flow already negative and the 2027 capex bill about to land squarely on the bond market, any year-end rally could be a window to sell.
This means → even if a short-term bounce materializes, the medium-term logic remains the same: "who pays for this bill?"
市场有风险，内容仅供研究参考，不构成投资建议。