Skip to content
Advertisement
GPU rental prices climb 24% as AI trade cools, Silicon Data shows

GPU rental prices climb 24% as AI trade cools, Silicon Data shows

Photo: Pixabay / Pexels

H100 rental rates keep rising even as lenders tighten financing on chip-backed loans and AI output prices fall

The AI trade is supposedly coming apart at the seams. Nobody told the price of renting a GPU.

Median rental prices for Nvidia’s H100 reached approximately $3.70 per GPU-hour in the week of October 19, 2026, according to Silicon Data. That is up 24% from $2.98 a year earlier, a strange look for a sector that has spent the autumn absorbing bad headlines.

The numbers behind the rally

Silicon Data is a New York-based firm that tracks what it costs to rent graphics processors, the chips that do the heavy lifting for AI models. Its indices cover 95%+ of GPU providers. That breadth makes it something close to a price oracle for an industry that until recently ran on opaque, handshake-style deals.

The H100 is not the only chip getting pricier. Independent tracker GetDeploying shows similar increases, with gains also showing up across other generations, including Hopper and Nvidia’s newer Blackwell line.

The momentum shows up in shorter windows too. Silicon Data’s H100 Neo index rose 11.3% between September 30 and November 10, 2026, a move the firm tied to tight supply and demand conditions.

Longer commitments have moved even more. H100 contract rates climbed roughly 38-40% from late 2025 to early 2026, and that rise has carried into the fourth quarter.

Then there is CoreWeave, the specialized GPU cloud provider that serves as a bellwether for the space. On its Q3 2026 earnings call on November 10, the company reported price increases of 30-50% quarter over quarter for its older Ampere and Hopper fleets.

Advertisement

Silicon Data’s indices agree on the direction but not the magnitude. They show average Q3 rental gains of approximately 2.4% to 13.1%, though certain indices have climbed as much as roughly 30% year over year.

That gap matters. A single provider repricing its own fleet is one data point, while an index spanning most of the market is a broader read. Both point up, but one is pointing much more enthusiastically.

Several forces appear to be driving the climb. Demand has risen significantly, memory costs are going up, and fewer GPUs are rolling off existing contracts as the market approaches 2027-2028.

A bull chart in a bear market

ZeroHedge described rental pricing as the one chart AI bulls kept pointing to all autumn, even as the broader “AI supercycle” narrative unraveled around them. By its account, the AI trade peaked in June, and CoreWeave imploded in October.

The credit side tells a less cheerful story. Lenders have cut advance rates on GPU-backed loans from 80% to 50%.

An advance rate is how much a lender will hand over against collateral. Borrowers who once got 80 cents for every dollar of chips pledged now get 50 cents.

The crunch spread problem

Silicon Data has also flagged a pressure point it calls the “crunch spread.” It measures how much work a rented GPU must do to break even when rental costs rise while token prices fall.

Tokens, in this context, are the units of text that AI models process and generate. AI providers charge customers per token, so token prices are effectively the revenue side of running a model.

For Nvidia’s B300 chips, the break-even throughput requirement has risen 75% since May 31, 2026, according to the firm’s analysis. Silicon Data’s analysis suggests these price pressures are expected to persist into Q4 2026.

What this means for the AI infrastructure trade

For GPU owners, rising rents look like good news on paper. Pricing power on older Ampere and Hopper fleets suggests aging hardware is holding its value better than many skeptics assumed.

The catch sits with the lenders. A cut in advance rates from 80% to 50% means owners can borrow far less against the same chips, which could slow the pace at which new capacity gets financed.

There is also a market-structure angle. Silicon Data’s indices serve as benchmarks for financial products, including planned CME GPU futures. Divergence between a major provider reporting 30-50% increases and broad indices showing far more modest averages highlights why a trusted benchmark matters.

The watch list is fairly clear. Token pricing trends, lender appetite for GPU collateral, and the volume of chips rolling off contracts heading into 2027-2028 will likely determine whether this chart keeps rising or finally catches up with the mood of the rest of the AI trade.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.