Across 52 providers tracked on October 11, 2026, NVIDIA H100 80GB on-demand rentals spanned $1.71 per GPU-hour at Gcore to $10.98 at the top of the hyperscaler band, with a market median of $3.23 per GPU-hour and a separate H100 SXM listing from $2.27/hr [S1][S10].
Last-released GPUs (B200, B300, MI300X, RTX 5090) saw their on-demand median rise from $2.12 per GPU-hour in October 2024 to $4.72 in September 2026, while the Modern cohort (H100/H200/A100/L40S/RTX 4090/A10G/T4/L4) moved only $1.35 to $1.45 over the same window [S3].
H100 Rate Range and Short-Term Volatility
GetDeploying's index of 5,225 listings across 85 providers put the cheapest H100 80GB listing at $1.71 per GPU-hour (Gcore on-demand), against a 12-month trend of +8% [S1]. Separately, the Silicon Data GPU Rental Index recorded an isolated H100 spike of +10% in four weeks, from $2.00/hr on December 9, 2025 to $2.20/hr on January 6, 2026, with A100 and B200 flat over the same window [S5].
That spike ran counter to a multi-year normalization: H100 hourly rates fell from roughly $8/hr in early 2024 to a $1.50-$3.00/hr range by late 2025 across specialist providers, and H100 lead times compressed from over 50 weeks in 2023 to shorter delivery windows in 2024 [S5]. The October 2026 cross-provider range of $1.71 to $10.98 per GPU-hour now straddles both the discount end (boutique clouds, neoclouds, marketplace hosts) and the premium hyperscaler tier [S1][S4].
B200, B300, GB200 and GB300 Rate Cards
The Blackwell generation carries a steep premium: B200 180GB lists from $3.75 per GPU-hour at Packet.ai with a $6.69 median, B300 288GB starts at $4.99 with a $7.87 median, and GB200 186GB starts at $10.50 with a $16.00 median across five providers [S1][S3]. The Grace Blackwell Ultra GB300 288GB tops the table at $18.00 per GPU-hour on Oracle Cloud, and Vast.ai's October 2026 live snapshot shows B200 at a $6.00 floor and $7.50 median, with H200 NVL at $3.71/$4.25 and H200 at $4.37/$5.00 [S1][S6].
Reserved versus on-demand spreads are GPU-specific: B300's reserved median is $5.71 against $7.87 on-demand, H100 has a tight $0.13 gap ($3.10 reserved vs $3.23 on-demand), and MI300X actually inverts the pattern, with reserved at $3.39 above its $2.91 on-demand median [S3]. For engineering procurement, the on-demand column is the safe reference; reserved medians combine mixed commitment terms and cannot be read as a same-provider discount [S3].
Cheaper Tiers: A100, L40S, RTX 4090, RTX 5090
The Ampere and Ada Lovelace tiers carry the budget rungs: A100 80GB lists from $0.43 per GPU-hour at Vast.ai with a $1.80 median, RTX 4090 24GB from $0.31 with a $0.47 median, and RTX 5090 32GB from $0.35 with a $0.69 median [S1][S3]. L40S sits at a $1.29 per GPU-hour median, and A100 80GB provider rates run $1.09-$5.03/hr per Thunder Compute's October 2026 market scan [S3][S4].
The cheapest listing across the entire 108-model table is consumer-grade: the Nvidia GTX 1050 Ti at Salad at $0.020 per GPU per hour on-demand, useful for testing pipelines rather than training frontier models [S1]. For H100 class workloads, the GPU server rental price comparison pages indexed in October 2026 use the A100 to H100 ratio of roughly 1:3 at median (A100 80GB at $1.76 vs H100 at $3.25) as a quick planning benchmark [S3].
Cost Drivers: Chip Generation, VRAM Tier, and Provider Class
Four variables move the hourly rate more than anything else. (1) Chip generation: B-series Blackwell cards cost 2-3x the H100 median, and the Grace Blackwell modules (GB200/GB300) add another 1.5-2x on top [S1]. (2) VRAM and HBM tier: 80GB HBM3 H100 lists cheaper than 141GB HBM3e H200 in many markets, but the H200's $4.40 median is justified by bandwidth for LLM inference [S3]. (3) Provider class: hyperscalers (AWS, Azure, GCP) cluster at the high end, with Azure H200 at $13.78/hr, while neoclouds and marketplaces (Vast.ai, Gcore, Beam) anchor the floor [S1][S3]. (4) Billing tier: spot rates sit below on-demand, but spot capacity is unreliable for long jobs [S3].
Total cost of ownership extends well past the hourly line item. A dedicated 8x H100 private cloud configuration carries minimum commitments far above the $2.27/hr SXM sticker, and reserved-instance quotes bundle storage, egress, and SLA terms that vary by provider [S9][S10].
Market Size, Supply, and the 2026 Trajectory
The global GPU rental market is valued near $52.04 billion in 2026, up from $34.62 billion in 2025, with a projected $198.74 billion by 2031 at a 30.73% CAGR per Mordor Intelligence figures republished in October 2026 [S4]. The wider data center GPU market tracks a parallel curve, from $138.88 billion in 2026 to $624.17 billion by 2034 [S4].
Supply is uneven across SKUs. Confirmed availability is 7% for MI300X, 11% for B300, 16% for B200, 33% for H100, 68% for RTX 4090, and 79% for RTX 5090, with 88% of MI300X listings carrying an unknown stock status that prevents a clean sold-out read [S3]. For shop-floor compute planning tied to industrial AI workloads, this maps to the same logic used in SCARA robot pricing curves and EV charger power module supply: newest-generation silicon carries the highest premium and the tightest confirmed-availability, so a dual-source plan across H100 and A100 is the lowest-risk posture for Q4 2026 budgets.
Track three signals through Q1 2027: the spread between H100 cheapest and median (currently about $1.52 per GPU-hour, a useful stress gauge) [S1]; B200 and B300 reserved-tier medians as Blackwell inventory rotates out of the on-demand pool [S3]; and the cross-provider median for Last-released GPUs, which broke $4.50 for the first time in September 2026 and will test $5.00 if the October trajectory holds [S3].
Spec-level background on the components involved: pressure transmitter, and flow meter.