Goldman Sachs: AI Returns Hard to Quantify, Market Volatility Risks Rising
nashnova research
Goldman's hedge-fund-coverage chief Tony Pasquariello warns that AI capex returns may be impossible to capture in traditional financial models, while the risk of market 'air pockets and dislocations' is rising — the broad spending trend holds, but the road ahead will be rougher than the early phase.
From $154 billion to $1.3 trillion — where is the money going?
Hyperscale cloud operators — Meta, Google, Amazon, Microsoft — spent roughly $154 billion in capex in 2023. The consensus forecast for 2028 already exceeds $1.3 trillion.
This means → spending is set to grow nearly tenfold in five years. The compounding effect of financial commitments makes the market's demand for proof of returns increasingly urgent.
Each of these companies has flagged positive AI-revenue signals in 2026 earnings calls, but Pasquariello notes that the signals "haven't lasted long" each time.
Why might traditional return-on-capital math fail here?
Pasquariello argues that forcing AI spending into a binary ROIC framework — return on invested capital, the classic "how much does each dollar earn back" metric — may be the wrong analytical path.
He draws a parallel to cloud computing and cybersecurity spending: companies invest not to boost revenue directly but because competitors are investing — falling behind is the real cost. In plain terms = much of AI capex is essentially defensive, not offensive.
A senior CIO he quotes puts it bluntly: "Everyone wants a spreadsheet — but there is no spreadsheet for this." Goldman partner George Lee adds that part of the return will show up as competitive pressure and ultimately flow into consumer surplus. This means → the money goes out, but the winner may not be the company spending it — it may be the users across the ecosystem.
Will competitive pressure pull every industry in?
Pasquariello expects this competitive dynamic to cut across every sector.
Some AI spending will persist even at very low or zero returns, driven purely by necessity. In plain terms = companies won't spend because it pays — they'll spend because not spending risks being knocked out. That provides a durable floor under enterprise AI budgets.
What risks are building up?
Political risk: increasingly complex political dynamics are brewing; midterm elections will test the question of "who bears the cost of AI."
Misuse risk: the probability of AI being used for harm is rising, yet the market has not fully priced it in.
Power bottleneck: the next phase of expansion is constrained by insufficient electricity supply — a problem Pasquariello calls "both widely known and still underestimated."
What does this mean for investors?
Pasquariello sees the risk of "air pockets and dislocations" in the AI theme rising, citing the DeepSeek episode as a reference point for "lightning-strike" shocks.
His core conclusion: the long-term trend of broad AI spending remains intact, but "the path ahead will be bumpier than the early phase."
This means → realized volatility between AI leaders and laggards is likely to keep climbing — picking the right stocks will matter more than simply betting on the sector.
市场有风险，内容仅供研究参考，不构成投资建议。