NVIDIA GPU Rental Rates Rise in 2026, Led by a 143% Jump for the H200
Renting NVIDIA GPUs has become more expensive since the start of 2026, including older models such as the A100. Ornn Data figures reported by a16z show increases ranging from 9% for the A100 to 143% for the H200, alongside a sharp rise in AI token consumption.
Layla Haddad
Higher rates across GPU generations
The H200, launched in 2024, recorded the largest rental-rate increase in the figures reported by a16z. Behind its 143% gain were the B200 at 111%, the RTX PRO 6000 at 73% and the H100 at 54%. The A100, a 2020 model, rose 9%.
Ornn's October 23, 2026, Compute Price Index settlement put the B200 at $9.58 per GPU-hour and the H200 at $5.74. The H100 SXM settled at $3.17, followed by the RTX PRO 6000 at $1.54 and the A100 SXM4 at $1.09.
Those figures reflect completed rentals across the neocloud market, not advertised prices or quotes from individual providers. Ornn calculates the index once per trading day using volume-weighted, winsorized transaction averages.
An earlier snapshot shows how B200 rates had already been climbing. In an Ornn-based analysis published September 25, Data Gravity reported a rate of $8.01 per hour as of September 24, 2026, up 79% over three months.
Demand extends to older hardware
The rental charts show an upward trend alongside the spring surge in AI token consumption, but they do not identify how much demand comes from agentic workloads, enterprise adoption or other uses. Nor do those charts break out the effects of GPU supply, data center capacity or power constraints.
Data Gravity's September analysis offers a separate view of A100 availability and use. From March through September 2026, occupancy increased from 74% to 90% even as supply expanded 13%. The analysis put the resulting increase in the number of rented A100s at about 37%.
Older GPUs also remain useful for running medium-sized large language models. Wccftech reported that four Tesla V100s connected through an NVLink board can outperform systems such as the M5 Ultra and DGX Spark by up to 230%. Higher-VRAM V100 variants have begun selling at a $200 premium, with the hardware's value proposition attracting consumers more than enterprises.
Rental income and provider price increases
Rental revenue has also changed the ownership economics in Data Gravity's analysis. It estimated that an H100 purchased for about $19,000 in September 2025 would have a resale value of about $20,500 by the time of its September 25, 2026, report. With roughly $13,500 earned from rentals after costs, the estimated return was about 76%, including rent.
The same report described price increases across several providers. AWS increased reserved GPU block rates by about 20% on July 1, 2026, while CoreWeave reportedly raised prices by about 25%. Nebius listed a B200 rate of $8.50 effective October 1, up from $7.15.
Oracle, meanwhile, renewed expiring GPU contracts at a 20% premium, according to the analysis, which reported that 97.9% of its AI capacity was in use.
Memory pricing provides further context
Data Gravity also found a close relationship between memory capacity and rental rates. Since July 2026, the H200 had rented for about 1.76 times as much as the H100, matching their memory-capacity ratio. Measured per gigabyte of memory, hourly rental prices for the B200, H200 and H100 were nearly equal, at 3.6 to 4.2 cents.
The report cited TrendForce figures showing a 93% to 98% increase in DRAM contract prices during the first quarter of 2026. It also put Micron's gross margin at 84.9%, compared with NVIDIA's 75.0%, and cited TrendForce's expectation that high-bandwidth memory contracts would reset sharply higher in 2027.
