A100 and H100 Rental Prices Rip Higher as the "Old GPU" Bear Case Starts to Crack

Rental rates are repricing higher, fast

This chart shows a clear February squeeze in cloud GPU rentals. As of Feb 24, the Silicon Data A100 Rental Index prints 1.40, while the H100 Rental Index prints 2.43. That puts H100 at about a 74% premium versus A100 (2.43 vs 1.40), and A100 still clears roughly 58% of H100's level, which is a surprisingly "non-obsolete" price for an older generation.

The simplest read is brutal and boring: buyers are paying up for immediate capacity, which usually means utilization is rising and lead times are not easing fast enough.

"Old GPUs are worthless" runs into a timeline problem

$NVIDIA (NVDA.US)$ A100 was announced in May 2020 and NVIDIA said it was "in full production and shipping" at launch. H100 was introduced at GTC in March 2022, and NVIDIA later said H100 was in full production with the first wave of partner products rolling out in October 2022.

So here is the awkward fact for AI bears: nearly six years after A100's launch, its rental price is rising, not collapsing. If A100 were truly a fast-depreciating science project with a short usable life, you would expect the clearing price for scarce "instant GPU supply" to drift toward zero. Instead, it is behaving like a workhorse asset with a long tail of demand, especially for inference, fine-tuning, and capacity bridging when the newest silicon is tight.

There is also a real-world maintenance economy around these GPUs, with repair and refurbishment demand rising where supply is constrained, which extends usable life rather than abruptly ending it.

Why this matters for the 2026 to 2027 capex anxiety?

Markets are already nervous that hyperscalers keep stepping up 2026 and 2027 capex. A stronger "old GPU" rental tape helps the math: it supports high utilization and monetization of the installed base while new buildouts ramp. That makes the transition less dependent on a single generation's ROI and more like a rolling fleet model where multiple vintages earn.

Who benefits？

NVIDIA and other AI chip suppliers benefit first, because tight rentals usually point to strong end-demand and a willingness to pay for performance tiers.

Cloud platforms benefit next, because high rental clearing prices expand the opportunity to monetize both new and existing fleets. That includes the major hyperscalers and also the neocloud cohort that is explicitly positioned around GPU capacity, such as $CoreWeave (CRWV.US)$ and $NEBIUS (NBIS.US)$ .

80K Views