GPU Rental Prices Rebound, but Risk of AI Compute Demand Ceiling Still Looms
Nashnova编辑部
Ahead of Nvidia's earnings, GPU rental transaction prices are climbing while listing prices hold flat — neither metric signals weakening demand. But Bloomberg strategist Simon White warns that Jevons' Paradox won't hold forever, and compute demand faces an existential ceiling risk.
Why are two GPU rental indexes telling different stories?
The Ornn Compute index, which tracks actual transaction prices, rose noticeably over the past two weeks. The Silicon Data index, which tracks seller listing prices, barely moved.
This means → real buyers are bidding up GPU rentals, but sellers haven't raised their ask prices to match — supply is tightening, yet sellers haven't confirmed the trend.
Bloomberg macro strategist Simon White notes the two indexes cover different market segments, which may explain the gap. Either way, neither shows compute demand weakening.
Token prices are falling — doesn't that mean demand is slowing?
The per-token price for AI tasks — what users pay each time they call a model — keeps dropping. On the surface, it looks like shrinking demand.
In plain terms = tokens got cheaper not because nobody wants GPUs, but because open-source models grabbed market share from closed-source rivals and drove prices down.
Open-source models aren't necessarily more efficient — they may actually consume more compute per task. This means → profit margins at OpenAI and Anthropic are under pressure, but Nvidia's GPU demand isn't necessarily hurt.
Does "cheaper means more usage" always hold?
White invokes Jevons' Paradox: when the cost of using compute falls → usage rises → total demand keeps expanding.
In plain terms = it's like widening a highway — more lanes attract more cars. Cheapness itself creates new demand.
But White adds a caveat: this logic doesn't hold at every time horizon and every price level. He is skeptical of the claim that compute demand is "effectively limitless."
Why is Nvidia rushing to turn compute into an "asset class"?
Nvidia partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to package AI-factory compute as an investable asset class.
This means → Nvidia wants GPUs to stop being a one-time hardware sale and start behaving like commercial real estate — generating ongoing rental income as a long-duration asset.
The premise: rental demand for older chips like the H100 stays strong — proving chips remain scarce and supporting longer depreciation cycles.
What is the real risk here?
White warns: if rental prices collapse at some point — meaning Jevons' Paradox fails and compute demand actually hits a ceiling — Nvidia's asset-class strategy could unravel before it fully launches.
This reflects a deeper market anxiety about AI — not this quarter's earnings, but the longer question: is capital spending sustainable over time?
Put simply = the short-term news is fine, but no one has answered the question of whether demand has a ceiling.
市场有风险，内容仅供研究参考，不构成投资建议。