1 hour ago
15
For years, the AI trade has rested on a simple thesis: compute is scarce, demand is insatiable, and anyone who builds the infrastructure to deliver it will print money. Rich Privorotsky, head of Goldman Sachs’ Delta 1 trading desk, is now poking holes in that story.
Privorotsky flagged in mid-June 2026 that major AI companies are increasingly selling more compute hardware, directly challenging the idea that shortages will persist indefinitely.
The scarcity narrative is losing its grip
The core observation is straightforward. Major AI companies, the same hyperscalers that were previously hoarding every GPU they could get their hands on, are now making more compute hardware available to the market. That’s a meaningful shift from buyer to seller, and it suggests internal demand may be plateauing relative to the buildout pace.
Hyperscaler stocks have been underperforming recently, even as earnings commentary remains bullish on AI-driven revenue.
Rental prices in the compute market have also been softening. When the cost of renting GPU capacity declines, it tells you supply is catching up to demand, or demand is growing more slowly than expected.
Data center costs are climbing in the other direction
While compute hardware availability is improving, the cost of building the facilities to house it is moving the opposite way. Goldman Sachs estimates that next-generation AI data center costs have risen to $15 to $20 million per megawatt for advanced operational sites.
This creates an interesting tension. Hardware is becoming more available, which should push rental prices down. But the fixed costs of building and operating the facilities are going up. For hyperscalers, that means margins could get squeezed from both ends: lower revenue per unit of compute sold, higher costs per unit of capacity built.
What this means for investors
For semiconductor stocks, increased hardware availability means the supply-demand imbalance that supported premium pricing may be normalizing. Companies like Nvidia, which have benefited enormously from the perception of GPU scarcity, could face pressure on their forward earnings estimates if pricing power erodes.
If hyperscalers are selling more compute capacity into the market, it suggests they’ve overbuilt relative to their internal needs, or they’re monetizing excess capacity to offset the rising cost of new facilities.
The declining compute rental metrics that Privorotsky referenced deserve close attention. Rental prices are essentially a real-time signal of supply-demand balance in the compute market.
Investors should watch three signals closely: hyperscaler capex guidance in upcoming earnings calls, GPU rental pricing trends on platforms like CoreWeave and Lambda, and any signs that semiconductor order backlogs are shortening.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·