RTX 4090 rental rates compared across networks

The RTX 4090 is the commodity tier of the GPU market: 24 GB of consumer VRAM renting from $0.36/hr at the floor - an order of magnitude under the datacenter tiers. The cheap rate is real and the discount is honest: preemption is priced in, and the job design either absorbs it or pays for bonded capacity. 14 providers list 4090-class capacity here; this page prices the tier and its discipline.

The mining-history dividend, honestly

The 4090's mining-era fleet is why this tier has supply at all [1] - and it cuts both ways. This is the only tier on the site where the mining history is part of the product, so it gets its own section:

• The overlap is the supply. The same marketplaces that rented GPU mining hashrate now rent AI compute; the population on this page and the compute hub share hosts. That is why mining-era cards carry wear-risk questions and why platform reputation systems price them [2].
• The legacy hashrate question - ex-mining hosts list 4090s for AI work at commodity rates; the rent-gpu guide covers the host-reputation checks that make that supply usable.
• Wear risk is priced, not ignored - reputable platforms track host records; ask for the uptime history before long jobs, the same standard the mining hub applies to operator track records.

Why commodity hours are cheap

The cheapest legitimate compute in this market is also the tier most dependent on your own job design. The sections below cover the discount mechanics, the fits, and the checkpoint discipline - in that order, because the order is the decision.

Three structural reasons [1][2]:

1. The supply is enormous. Consumer 4090s exist in far greater numbers than datacenter cards, and their owners rent the same hardware the mining era used - supply overlaps between the compute population and this page.
2. Preemption is the price. A higher bid takes the card; the rate discounts the interruption. Your checkpoint discipline is the other half of that trade [2].
3. No fabric premium. Single-card jobs do not need interconnect - which is the commodity tier's whole advantage for fits that stay on one card.

The flat feed line is price discovery, the same as every tier at this stage - the floor is documented, the spread is developing, and the gpu-prices page is the record.

Live from our database - caveats and sources in each linked review.

What fits on 24 GB

The fit arithmetic from the llm-inference hub, applied at the commodity tier - fit is the first filter, and on this tier it is also the graduation question:

• 7B-class at FP16 - fits with KV-cache headroom; the tier's sweet spot for inference.
• 13B-class quantized - fits with quality tradeoffs; test on one card before fleet-quantizing.
• Long contexts - KV cache is the constraint; budget by your real context length, not a demo prompt [5].
• Multi-GPU jobs do not belong here - PCIe-only fabric; the a100 hub is where model-parallel work prices.
• Rendering and burst fits - the tier's original workload; the rent-gpu guide covers the practical setup for burst rendering on marketplace hosts.

The checkpoint discipline, commodity edition

The compute hub's interruption rules, applied at the tier where they matter most:

• Checkpoint every few minutes - treat any job as if the machine disappears at a random point [2].
• Keep state external - weights and data on object storage; the node's disk vanishes with the lease.
• Diversify across hosts - one host's preemption costs minutes, not days.
• Bid the finished job, not the hour - rate times expected runtime plus re-run overhead; the price-compare tool does the arithmetic.
• Test at the bid you will use - preemption frequency is bid-dependent; a test run at your production bid is the only sample that predicts your production experience [2].

Where the tier graduates

The honest question at this tier is not "is it cheap" but "when does the discount stop paying" [2]:

• Long unattended runs graduate to the a100 hub - fewer interruptions beat a cheaper rate once re-run overhead crosses the price gap.
• Compliance-sensitive inference graduates to attestation - the h100 hub and llm-inference hub price that requirement.
• Steady utilization graduates when the crossover replay says dedicated capacity is cheaper than interrupted bursts.

The graduation is the point: the commodity tier is where you learn the checkpoint discipline cheaply, and the discipline is what makes every tier above it affordable.

14 providers make the live table; 86% of all rows carry facts we verified with dates. The commodity tier's honest role: the cheapest legitimate compute in this market for workloads designed to absorb interruption - and the graduation path to the a100 and h100 tiers when production reliability outranks the discount. The free-trial hub lists where to test the loop without commitment, and the change log records how the tier's floor moves as supply grows.