GPU Rental Prices Surge 94% in September 2026

TL;DR

RTX 3090 rental rates surged 94% in 30 days to a median of $0.40/hour, while the Tesla V100 declined 14% to $0.11/hour - the cheapest option across 126 GPU models tracked by GPUniq as of 2026-09-27. The RTX PRO 6000 commands the market ceiling at $2.31/hour, roughly 21x the V100 floor.

Which GPUs spiked most in 30 days?

The RTX 3090 is the standout mover. GPUniq's September 2026 index puts it at +94.4% in 30 days, median $0.402/hour. That's the steepest single-model climb across all 126 tracked SKUs.

Right behind it: RTX 3070 at +71.6% and RTX 4090 at +62.1%. Three cards from two generations, all training workhorses, all moving together. That pattern points to a capacity squeeze rather than isolated demand for any one model.

The Tesla V100 went the other direction. Down 14.2%, now $0.114/hour median.

RTX 5090 rose 30% - real, but modest compared to the older-gen cards. The newest silicon isn't the one under pressure.

What's the cheapest GPU to rent now?

Tesla V100 at $0.09/hour minimum, $0.114/hour median. That's the floor across the entire market.

RTX 3060 is close - $0.078/hour minimum, $0.122/hour median - but it's running at 80% use with only 152 available offers, so finding one when you need it is a different question from the sticker price. RTX 3070 starts at $0.097/hour, but its 71.6% price surge means that entry point may not hold long.

For inference workloads on a budget, the V100 has 1,225 active offers across 1,887 total units. Plenty of supply. The price drop reflects demand shifting toward newer cards, not V100s becoming unreliable.

RTX 5090 entry point is $0.408/hour. Newest generation, not cheapest.

Why is RTX 3090 up 94% this month?

Two things at once. The RTX 3090 runs at 70% use - matching the RTX 4090 - which means sustained booking, not spikes. The total pool is only 978 units with 267 available offers. Thin market. When demand holds and supply contracts as older-gen cards age out of rental fleets, prices move fast.

ML fine-tuning and small-batch inference are the likely culprits. The 3090's 24GB VRAM made it a workhorse for 7B-13B parameter models before the 4090 became widely available, and teams aren't migrating pipelines mid-project.

The simultaneous RTX 4090 surge (+62.1%) confirms this isn't 3090-specific. Both cards are hitting the same ceiling. Providers who might normally absorb demand by routing to 4090s can't, because those are also near-full.

Which GPUs have best availability right now?

Tesla V100 leads on raw offer count: 1,225 active offers, 40% use. Easiest path if you need to spin up today.

RTX 5090 has 968 offers and 50% use - the newest generation hasn't been absorbed yet. RTX 3090 is the tightest: 267 offers, 70% use. Book ahead for any multi-week training run.

RTX 3060 at 80% use with only 152 offers is worth flagging. Cheap, but genuinely constrained. RTX 5060 Ti is in the same position: 80% use, 181 offers.

A100 vs RTX 4090: how big is the gap?

A100 median is $1.413/hour. RTX 4090 median is $0.899/hour after its 62% surge. That's a 57% premium for the A100, or $0.51/hour more. At 700 hours of training, the gap is $357.

The A100 runs at 70% use with only 192 available offers - same use as the 4090 but far fewer units. Both are under pressure. The A100's edge is ECC memory, NVLink support, and enterprise SLAs that matter for production inference rather than experimental training.

RTX PRO 6000 at $2.31/hour is the real outlier. That's 2.57x the A100 median and 21x the V100 floor. It's a Blackwell-architecture professional card aimed at rendering and design.

Is RTX PRO 6000 worth $2.31/hour?

For most ML workloads, no. The RTX 5090 at $1.139/hour covers the same training scenarios at roughly half the cost. If you're optimizing cost-per-TFLOP on AI workloads, RTX 4090 at $0.899/hour or RTX 3090 at $0.402/hour are the rational choices.

The RTX PRO 6000 makes sense for professional rendering pipelines, CAD visualization, or workflows requiring certified drivers and professional-tier support contracts. The 50% use rate confirms it's a specialty market. At $1.658/hour minimum, half the fleet sits idle.

Why are RTX 5000-series modestly priced?

Supply is the answer. The RTX 50-series launched with enough inventory that the rental market hasn't absorbed it yet. RTX 5090 sits at 50% use with 968 active offers. RTX 5080 is at 60%. RTX 5070 Ti at 40%. None are under the same pressure as the 4090 and 3090.

There's also a substitution effect. The RTX 4090 is a known quantity for ML teams - stable CUDA support, existing benchmark data. Some teams will pay $0.899/hour for the 4090 rather than migrate to a 5090 at $1.139/hour mid-project. That keeps 5090 demand from running away.

The market is still digesting the RTX 50 supply wave. By Q1 2027 that picture may look different.

Market coverage

GPUniq's rental marketplace data as of 2026-09-27 covers 126 GPU models. Full price range: $0.09/hour (V100 minimum) to $2.311/hour (RTX PRO 6000 median). This is the spot rental and peer-to-peer market, not just headline cloud providers - which is where price movements show up first.

The RTX 3090's 94.4% move is the sharpest signal in the dataset. A card at 70% use with 267 available offers nearly doubling in price over 30 days isn't noise. That's a supply constraint playing out in real time.

Related on the GPUniq blog

• RTX 4090 Dominates GPU Rentals at 30K/Week
• GPU Rental Prices Surge 55% in 30 Days