24GB
VRAM
GDDR6X memory
AIVory · GPU Marketplace
Live spot pricingNVIDIA RTX 3090 — ten cents an hour.
The budget king of cloud GPU rental. Same 24 GB VRAM as the RTX 4090 at a fraction of the hourly cost — from $0.10/hr on the spot market. When you need raw compute at the lowest possible price and can tolerate spot interruptions, nothing in the market comes close.
At a glance
RTX 3090 specifications.
Key hardware specs that determine what workloads this GPU handles.
936 GB/s
Memory Bandwidth
peak throughput
350W
TDP
thermal design power
Ampere
Architecture
NVIDIA GPU architecture
Spot pricing
RTX 3090: live hourly rates.
Every provider offering this GPU on the spot market, sorted cheapest first.
Prices in USD per GPU-hour · spot instances · sorted cheapest first
Recommended models
AI models that run well on RTX 3090.
Tested model-GPU pairings with notes on why each is a good fit.
Mistral Nemo 12B
Fits in FP16 with 12 GB to spare. The ideal pairing for high-volume, low-cost inference where absolute latency is secondary to price.
View model pricing →
Gemma 2 9B
At 9B parameters in FP16, this model uses under half the 3090's VRAM. Run large batches concurrently without hitting memory limits.
View model pricing →
Voxtral Mini 3B
Tiny model, huge headroom. Run dozens of concurrent inference streams on a single RTX 3090 for voice or text applications.
View model pricing →
Llama 3.3 70B (4-bit GPTQ)
Aggressive quantization squeezes 70B into 24 GB. Quality drops slightly, but at $0.10/hr it is unbeatable for non-critical workloads.
View model pricing →
Use cases
What the RTX 3090 is built for.
-
Ultra-cheap batch processing at $0.10/hr
Process millions of tokens overnight for pennies. Document classification, sentiment analysis, entity extraction, embedding generation — any workload where you measure success in tokens-per-dollar rather than tokens-per-second. A single RTX 3090 at $0.10/hr processing 50 tokens/second runs through 18 million tokens in 100 hours for $10 total.
-
Horizontal scaling — ten GPUs for the price of one A100
At $0.10/hr per card, renting 10 RTX 3090s costs $1/hr — less than a single A100 on most providers. For embarrassingly parallel workloads like dataset labeling, eval harness runs, or embedding databases, this horizontal approach delivers more aggregate throughput with built-in redundancy. If one spot instance gets reclaimed, nine keep running.
-
Development environments and CI/CD testing
Spin up a 3090 instance for your CI pipeline's model evaluation suite. Run unit tests against a real GPU for $0.10/hr instead of paying for expensive reserved capacity that sits idle between builds. The RTX 3090 is fast enough to catch performance regressions and verify model outputs without slowing down your development cycle.
FAQ
Common questions.
Why is the RTX 3090 the cheapest GPU on the spot market?
Oversupply from the cryptocurrency mining boom. Millions of RTX 3090s were purchased for Ethereum mining before the merge to proof-of-stake in 2022. Those GPUs flooded the secondary market and rental platforms. The hardware is still excellent for AI inference — 24 GB of fast GDDR6X — but market dynamics keep the price at a fraction of newer cards.
RTX 3090 vs RTX 4090 — when does the extra cost of a 4090 justify itself?
Only when latency matters. The 4090 delivers roughly 40-60% faster inference than the 3090 on the same model thanks to Ada Lovelace's improved tensor cores. For real-time chat serving where sub-second response time affects user experience, pay the premium. For batch processing, overnight jobs, async pipelines, or development, the 3090's lower hourly rate makes it the better value.
What AI models fit on the RTX 3090's 24 GB VRAM?
In FP16: all models up to ~12B parameters (Mistral Nemo 12B, Gemma 2 9B, Phi-3 14B with tight fit). In INT8: models up to ~22B (Mistral Small 24B with some KV cache pressure). In INT4/GPTQ: up to ~65B (Llama 3.3 70B is very tight). Practically, the sweet spot is 3B-12B models in FP16 where you get full quality and throughput.
RTX 3090 power efficiency — how does 350W compare?
The 3090 draws 350W under load, the same as the L40S and less than the RTX 4090's 450W. At $0.10/hr, the electricity cost is essentially included in the rental price and negligible. In a colocation scenario where you own the card, the 3090 offers the best inference-per-watt at the budget tier because the older Ampere architecture requires less cooling infrastructure than the 4090.
Rent a RTX 3090. Right now.
Spot pricing, per-second billing, no commitment.
Browse the live marketplace, pick your GPU, deploy in one click. Credits from $10.