Goldman Sachs 1-Delta Head: The AI Capex Trade Rubber Band Has Been Stretched to Its Limit
0xBroomberg
Goldman Sachs 1-Delta trading desk head Rich Privorotsky warns that markets have ignored every negative signal on AI capital spending for weeks — the rubber band is near its breaking point, and even a modest spending trim by any hyperscaler could reprice the entire trade.
What exactly has the market been ignoring?
GLM-5.2 (a frontier model from Zhipu AI, trained entirely on domestic chips), fusion architectures, small-model advances, and ongoing cost compression — none of it has been priced into hyperscaler spending expectations.
This means → the market has piled all AI-beneficiary bets into semis, memory, power, networking, and infrastructure, while assigning zero probability to the scenario that frontier intelligence can be developed in the East at far lower cost.
In plain terms = GLM-5.2 was trained on 100,000 Ascend 910B processors with zero Nvidia chips. If that path scales, the biggest capital allocators are also the most exposed to over-investment risk.
How is the "snap point" defined?
Privorotsky defines it as: the moment any major spender concludes that spending slightly less would better serve shareholders.
The problem is that "slightly less" is not in anyone's model at all. The entire AI complex is priced on one premise: inference demand grows → capex keeps rising, with no deceleration option in between.
This reflects an extreme one-sided consensus — one hyperscaler flinching would trigger a chain of valuation challenges across the stack.
What are technicals and market structure saying?
The Nasdaq failed to make a decisive new high. The world's largest companies by market cap are showing what Privorotsky calls an "unstable equilibrium."
Last week's options-expiry tailwind has faded. Markets are entering month-end and quarter-end rebalancing — a window that theoretically favors selling equities, buying bonds.
Market-maker gamma (how sensitive dealers' hedging is to price moves — lower gamma means bigger swings) is low near current spot and drops further on a move down. CTA trend strategies (systematic trend-following funds) remain net long but are highly convex to the downside — once a trigger level breaks, selling accelerates fast.
All roads lead to the same few stocks — where is the risk?
Prime-brokerage data and price action confirm the same picture: the global market has become one extraordinarily concentrated trade — AI drives equities, equities drive economic expectations, and every path ends at the same handful of names.
Privorotsky also flags a risk the market has long ignored: the highly deflationary forces embedded in token economics (the per-token pricing structure of AI inference).
This means → the key signal for when this trade actually snaps is singular: the stock-price action of the hyperscalers themselves. The day their shares waver is the day the rubber band breaks.
Content is for reference only, not financial advice.