# OnlineGPU

> Cloud GPU provider (onlinegpu.io). 47 NVIDIA and AMD models in 3 regions, sold on three tiers: spot (interruptible, 30–53% below on-demand, 2-minute reclaim notice), on-demand (never reclaimed, 99.9% monthly uptime target) and reserved (fixed rate, 1–12 months). Billed per started minute at the hourly rate divided by 60, from $0.03 per GPU-hour (RTX A4000) to $2.49. Every tier is priced at least 20% below the lowest published price found for the same GPU across 102 providers. Prepaid credit in cryptocurrency from $30, no card, no identity document, no minimum commitment, no egress fees.

## Key facts
- Reference price: H100 SXM at $0.55 per GPU-hour on spot, $0.89 on-demand, $0.65 reserved.
- Price rule: every tier is at or below 80% of the lowest published price found across 102 providers (index of September 2026), checked by the catalogue builder before any price is published.
- Billing: per started minute (hourly price ÷ 60), no minimum duration, boot time and reclaim events not billed, no egress fees. Persistent storage $0.08 per GB-month.
- Payment: prepaid credit from $30 in USDT (TRON · TRC-20 or Ethereum · ERC-20), Bitcoin, Ethereum, Solana, Litecoin, Monero, TRON, confirmed on-chain. No card. Sign-up needs an e-mail address and a password only: no identity document, phone number or postal address.
- Interruption: spot instances are stopped after a 2-minute notice (metadata endpoint, webhook, console); the persistent disk, its data and the IP are kept; the 30-day reclaim rate is published per model.
- Regions: US East (Virginia, United States), US West (Oregon, United States), EU Central (Frankfurt, Germany). Prices identical in every region.
- Node sizes: 1 to 8 GPUs per instance, same price per GPU at every size. Instance reachable in under 60 seconds.
- Reliability: 99.98% average uptime across components over the last 90 days, 8 incidents in 90 days, every one listed with its timeline at https://onlinegpu.io/status.
- Support: tickets answered by the engineers who run the nodes, 24/7, first reply target 12 hours (2 hours when an instance is unreachable).
- Not certified: no SOC 2, no ISO 27001 as of November 2026 — stated plainly at https://onlinegpu.io/trust.
- Machine-readable: catalogue https://onlinegpu.io/catalog.json · status https://onlinegpu.io/status.json · full text export https://onlinegpu.io/llms-full.txt

## Quick answers
- How much does an H100 cost per hour? H100 SXM: $0.55 per GPU-hour on spot, $0.89 on-demand, $0.65 reserved. Billed per minute, identical in every region.
- How much does a B200 cost per hour? B200 SXM: $1.69 per GPU-hour on spot, $2.49 on-demand, $1.85 reserved.
- How much does an H200 cost per hour? H200 SXM: $0.75 per GPU-hour on spot, $1.09 on-demand, $0.79 reserved.
- Cheapest way to rent an RTX 5090? $0.09 per hour on spot, $0.19 on-demand, 32 GB GDDR7.
- What is the cheapest GPU? RTX A4000 at $0.03 per GPU-hour on spot ($0.0005 per minute).
- Is there a minimum spend or a contract? No. Add credit from $30, launch, stop whenever you want; only the reserved tier has a term (1 to 12 months).
- How do I pay? USDT (TRON · TRC-20 or Ethereum · ERC-20), Bitcoin, Ethereum, Solana, Litecoin, Monero, TRON, prepaid and confirmed on-chain. No card and no identity check.
- What happens when a spot instance is reclaimed? A 2-minute notice, then the instance is stopped, never deleted: disk, data and IP stay attached; relaunch on spot or move the same disk to on-demand in one click.
- Which tier for a production API? On-demand: same hardware, never reclaimed, 99.9% monthly uptime target with service credits.
- Where are the GPUs? US East (Virginia, United States), US West (Oregon, United States), EU Central (Frankfurt, Germany).

## Best for, not for
- Best for: training and fine-tuning with checkpoints, batch and offline inference, embeddings and evaluation runs, rendering, hyperparameter sweeps, CI, and anyone who wants to pay a GPU by the minute in cryptocurrency without a card or a contract.
- Not for spot: latency-sensitive production APIs and sessions that cannot be interrupted — use the on-demand or reserved tier on the same nodes.
- Not offered: SOC 2 or ISO 27001 attestations, and no GPU sharing between accounts (every GPU is passed through whole to one virtual machine).

## Prices and tiers
- [Pricing](https://onlinegpu.io/pricing): the three tiers, the full price table for all 47 models, a cost calculator and the method behind the price guarantee.
- [Spot instances](https://onlinegpu.io/spot): how reclaims work, the 2-minute notice on three channels, published 30-day reclaim rates per model.
- [On-demand instances](https://onlinegpu.io/on-demand): never reclaimed, 99.9% monthly uptime target, per-minute billing.
- [Reserved capacity](https://onlinegpu.io/reserved): fixed rate for 1 to 12 months, with the break-even utilisation for each model.
- [Persistent storage](https://onlinegpu.io/storage): $0.08 per GB-month, 10–4,000 GB volumes that survive a reclaim.
- [Pay with crypto](https://onlinegpu.io/pay-with-crypto): USDT (TRON · TRC-20 or Ethereum · ERC-20), Bitcoin, Ethereum, Solana, Litecoin, Monero, TRON — prepaid credit from $30, no card, no identity check.

## GPU catalogue
- [All GPUs](https://onlinegpu.io/gpus): filterable catalogue with memory, bandwidth, node sizes, availability and the three prices.
- [B200 SXM](https://onlinegpu.io/gpus/b200-sxm): 180 GB HBM3e, $1.69/h spot, $2.49/h on-demand, reclaim rate 5–10%.
- [H200 SXM](https://onlinegpu.io/gpus/h200-sxm): 141 GB HBM3e, $0.75/h spot, $1.09/h on-demand, reclaim rate 5–10%.
- [H200 NVL](https://onlinegpu.io/gpus/h200-nvl): 141 GB HBM3e, $0.69/h spot, $0.99/h on-demand, reclaim rate 5–10%.
- [H100 SXM](https://onlinegpu.io/gpus/h100-sxm): 80 GB HBM3, $0.55/h spot, $0.89/h on-demand, reclaim rate <5%.
- [H100 NVL](https://onlinegpu.io/gpus/h100-nvl): 94 GB HBM3, $0.49/h spot, $0.79/h on-demand, reclaim rate <5%.
- [H100 PCIe](https://onlinegpu.io/gpus/h100-pcie): 80 GB HBM2e, $0.45/h spot, $0.69/h on-demand, reclaim rate <5%.
- [GH200](https://onlinegpu.io/gpus/gh200): 96 GB HBM3, $0.49/h spot, $0.79/h on-demand, reclaim rate 5–10%.
- [A100 SXM 80GB](https://onlinegpu.io/gpus/a100-sxm): 80 GB HBM2e, $0.19/h spot, $0.29/h on-demand, reclaim rate <5%.
- [A100 PCIe 80GB](https://onlinegpu.io/gpus/a100-pcie-80): 80 GB HBM2e, $0.09/h spot, $0.19/h on-demand, reclaim rate <5%.
- [A100 40GB](https://onlinegpu.io/gpus/a100-40): 40 GB HBM2, $0.09/h spot, $0.17/h on-demand, reclaim rate <5%.
- [RTX PRO 6000 Blackwell](https://onlinegpu.io/gpus/rtx-pro-6000): 96 GB GDDR7, $0.19/h spot, $0.39/h on-demand, reclaim rate 5–10%.
- [RTX 5090](https://onlinegpu.io/gpus/rtx-5090): 32 GB GDDR7, $0.09/h spot, $0.19/h on-demand, reclaim rate 10–15%.
- [RTX 4090](https://onlinegpu.io/gpus/rtx-4090): 24 GB GDDR6X, $0.06/h spot, $0.09/h on-demand, reclaim rate 10–15%.
- [GPU comparisons](https://onlinegpu.io/compare): any two models side by side, specs and prices.
- [Price comparisons with other providers](https://onlinegpu.io/alternatives): published prices from the market index, dated and sourced.
- [OnlineGPU vs Lium](https://onlinegpu.io/alternatives/lium-io): 42 models compared, our price is 43% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs Vast.ai](https://onlinegpu.io/alternatives/vast-ai): 39 models compared, our price is 48% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs RunPod](https://onlinegpu.io/alternatives/runpod): 34 models compared, our price is 65% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs GPU.ai](https://onlinegpu.io/alternatives/gpu-ai): 26 models compared, our price is 62% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs Runcrate](https://onlinegpu.io/alternatives/runcrate): 21 models compared, our price is 83% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs Sesterce](https://onlinegpu.io/alternatives/sesterce): 21 models compared, our price is 79% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs EmpirioLabs AI](https://onlinegpu.io/alternatives/empiriolabs-ai): 19 models compared, our price is 84% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs Massed Compute](https://onlinegpu.io/alternatives/massed-compute): 16 models compared, our price is 74% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs GPU Outlet](https://onlinegpu.io/alternatives/gpu-outlet): 16 models compared, our price is 48% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs AtmosCompute](https://onlinegpu.io/alternatives/atmoscompute): 14 models compared, our price is 83% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs Omega Gradient](https://onlinegpu.io/alternatives/omega-gradient): 14 models compared, our price is 78% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs Salad](https://onlinegpu.io/alternatives/salad): 14 models compared, our price is 54% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs AWS](https://onlinegpu.io/alternatives/aws): 13 models compared, our price is 89% below theirs on average (published prices, index of September 2026).
- [OnlineGPU vs Shadeform](https://onlinegpu.io/alternatives/shadeform): 13 models compared, our price is 76% below theirs on average (published prices, index of September 2026).

## Guides
- [How much does an H100 cost per hour?](https://onlinegpu.io/guides/h100-price-per-hour): Published on-demand and spot prices for the H100 from every provider in the index, sourced and dated, with the median, the cheapest, and the monthly cost of each.
- [Cheapest cloud GPU providers](https://onlinegpu.io/guides/cheapest-cloud-gpu-providers): The five lowest published prices for eight popular GPUs, from the H100 to the RTX 4090, and the providers that keep showing up at the bottom of the table.
- [How much does a B300 cost per hour?](https://onlinegpu.io/guides/b300-price-per-hour): Published B300 prices from every provider in the index, sourced and dated, and how many B300s or B200s six large models need — with the node price for each.
- [H100 vs H200 vs B200](https://onlinegpu.io/guides/h100-vs-h200-vs-b200): Memory, bandwidth, tensor throughput and price per hour side by side, what fits on one GPU of each, and which one to rent for inference, training and fine-tuning.

## Model GPU requirements and calculators
- [LLM GPU requirements](https://onlinegpu.io/models): VRAM at BF16, FP8 and 4-bit (or the published format), KV cache at 32K context and the cheapest node for 64 open models; architecture read from each model's config.json.
- [DeepSeek V4 Flash GPU requirements](https://onlinegpu.io/models/deepseek-v4-flash): 608 GB at BF16, 167 GB at FP4+FP8; cheapest node 2× RTX PRO 6000 Blackwell at $0.38/h spot.
- [DeepSeek V4 Pro GPU requirements](https://onlinegpu.io/models/deepseek-v4-pro): 3.2 TB at BF16, 865 GB at FP4+FP8; cheapest node 8× MI325X at $8.72/h spot.
- [DeepSeek V4.1 Flash GPU requirements](https://onlinegpu.io/models/deepseek-v4-1-flash): 1.53 TB at BF16, 510 GB at FP4+FP8; cheapest node 8× RTX PRO 6000 Blackwell at $1.52/h spot.
- [DeepSeek V3.2 GPU requirements](https://onlinegpu.io/models/deepseek-v3-2): 1.37 TB at BF16, 343 GB at 4-bit; cheapest node 8× A40 at $0.48/h spot.
- [Gemma 4 31B GPU requirements](https://onlinegpu.io/models/gemma-4-31b): 62.5 GB at BF16, 15.6 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Gemma 4 26B A4B GPU requirements](https://onlinegpu.io/models/gemma-4-26b-a4b): 51.6 GB at BF16, 12.9 GB at 4-bit; cheapest node 1× RTX A4000 at $0.03/h spot.
- [Gemma 4 12B GPU requirements](https://onlinegpu.io/models/gemma-4-12b): 23.9 GB at BF16, 5.98 GB at 4-bit; cheapest node 1× RTX 3070 at $0.03/h spot.
- [Gemma 4 E4B GPU requirements](https://onlinegpu.io/models/gemma-4-e4b): 16.0 GB at BF16, 4 GB at 4-bit; cheapest node 1× RTX 3070 at $0.03/h spot.
- [MiniMax M2.5 GPU requirements](https://onlinegpu.io/models/minimax-m2-5): 457 GB at BF16, 114 GB at 4-bit; cheapest node 2× A100 PCIe 80GB at $0.18/h spot.
- [MiniMax M3 GPU requirements](https://onlinegpu.io/models/minimax-m3): 854 GB at BF16, 214 GB at 4-bit; cheapest node 4× A100 PCIe 80GB at $0.36/h spot.
- [Mistral Large 3 GPU requirements](https://onlinegpu.io/models/mistral-large-3): 1.35 TB at BF16, 338 GB at 4-bit; cheapest node 8× A40 at $0.48/h spot.
- [Devstral 2 123B GPU requirements](https://onlinegpu.io/models/devstral-2): 250 GB at BF16, 62.5 GB at 4-bit; cheapest node 2× A40 at $0.12/h spot.
- [Kimi K2.6 GPU requirements](https://onlinegpu.io/models/kimi-k2-6): 2.05 TB at BF16, 595 GB at INT4; cheapest node 8× RTX PRO 6000 Blackwell at $1.52/h spot.
- [Kimi K3 GPU requirements](https://onlinegpu.io/models/kimi-k3): 5.56 TB at BF16, 1.56 TB at INT4; cheapest node 8× MI325X at $8.72/h spot.
- [Nemotron 3 Nano GPU requirements](https://onlinegpu.io/models/nemotron-3-nano): 63.2 GB at BF16, 15.8 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Nemotron 3 Super GPU requirements](https://onlinegpu.io/models/nemotron-3-super): 247 GB at BF16, 61.8 GB at 4-bit; cheapest node 1× A100 PCIe 80GB at $0.09/h spot.
- [Nemotron 3 Ultra GPU requirements](https://onlinegpu.io/models/nemotron-3-ultra): 1.12 TB at BF16, 280 GB at 4-bit; cheapest node 4× A100 PCIe 80GB at $0.36/h spot.
- [Qwen3.8 27B GPU requirements](https://onlinegpu.io/models/qwen3-8-27b): 55.6 GB at BF16, 13.9 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Qwen3.8 Flash-Next GPU requirements](https://onlinegpu.io/models/qwen3-8-flash-next): 360 GB at BF16, 90.0 GB at 4-bit; cheapest node 2× A100 PCIe 80GB at $0.18/h spot.
- [Qwen3.8 2.4T A95B GPU requirements](https://onlinegpu.io/models/qwen3-8-2-4t): 4.89 TB at BF16, 1.22 TB at 4-bit; cheapest node 8× MI300X at $7.92/h spot.
- [Qwen3.6 35B A3B GPU requirements](https://onlinegpu.io/models/qwen3-6-35b-a3b): 71.9 GB at BF16, 18.0 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Qwen3.6 27B GPU requirements](https://onlinegpu.io/models/qwen3-6-27b): 55.6 GB at BF16, 13.9 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Qwen3.5 397B A17B GPU requirements](https://onlinegpu.io/models/qwen3-5-397b-a17b): 807 GB at BF16, 202 GB at 4-bit; cheapest node 4× A100 PCIe 80GB at $0.36/h spot.
- [Qwen3.5 122B A10B GPU requirements](https://onlinegpu.io/models/qwen3-5-122b-a10b): 250 GB at BF16, 62.5 GB at 4-bit; cheapest node 1× A100 PCIe 80GB at $0.09/h spot.
- [Qwen3.5 35B A3B GPU requirements](https://onlinegpu.io/models/qwen3-5-35b-a3b): 71.9 GB at BF16, 18.0 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Qwen3.5 27B GPU requirements](https://onlinegpu.io/models/qwen3-5-27b): 55.6 GB at BF16, 13.9 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Qwen3 Coder Next GPU requirements](https://onlinegpu.io/models/qwen3-coder-next): 159 GB at BF16, 39.8 GB at 4-bit; cheapest node 1× A40 at $0.06/h spot.
- [GLM-5.3 GPU requirements](https://onlinegpu.io/models/glm-5-3): 1.51 TB at BF16, 377 GB at 4-bit; cheapest node 8× A100 PCIe 80GB at $0.72/h spot.
- [GLM-5.3 Flash GPU requirements](https://onlinegpu.io/models/glm-5-3-flash): 643 GB at BF16, 161 GB at 4-bit; cheapest node 4× A40 at $0.24/h spot.
- [GLM-5 GPU requirements](https://onlinegpu.io/models/glm-5): 1.51 TB at BF16, 377 GB at 4-bit; cheapest node 8× A100 PCIe 80GB at $0.72/h spot.
- [GLM-4.7 Flash GPU requirements](https://onlinegpu.io/models/glm-4-7-flash): 62.4 GB at BF16, 15.6 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Hunyuan Hy3 GPU requirements](https://onlinegpu.io/models/hy3): 598 GB at BF16, 149 GB at 4-bit; cheapest node 4× A40 at $0.24/h spot.
- [gpt-oss-120b GPU requirements](https://onlinegpu.io/models/gpt-oss-120b): 234 GB at BF16, 65.2 GB at MXFP4; cheapest node 1× A100 PCIe 80GB at $0.09/h spot.
- [gpt-oss-20b GPU requirements](https://onlinegpu.io/models/gpt-oss-20b): 41.8 GB at BF16, 13.8 GB at MXFP4; cheapest node 1× A30 at $0.05/h spot.
- [DeepSeek R1 GPU requirements](https://onlinegpu.io/models/deepseek-r1): 1.37 TB at BF16, 342 GB at 4-bit; cheapest node 8× A40 at $0.48/h spot.
- [DeepSeek R1 Distill Llama 70B GPU requirements](https://onlinegpu.io/models/deepseek-r1-distill-llama-70b): 141 GB at BF16, 35.3 GB at 4-bit; cheapest node 1× A100 PCIe 80GB at $0.09/h spot.
- [DeepSeek R1 Distill Qwen 32B GPU requirements](https://onlinegpu.io/models/deepseek-r1-distill-qwen-32b): 65.5 GB at BF16, 16.4 GB at 4-bit; cheapest node 2× RTX A4000 at $0.06/h spot.
- [DeepSeek R1 Distill Qwen 14B GPU requirements](https://onlinegpu.io/models/deepseek-r1-distill-qwen-14b): 29.5 GB at BF16, 7.39 GB at 4-bit; cheapest node 1× RTX A4000 at $0.03/h spot.
- [DeepSeek R1 Distill Llama 8B GPU requirements](https://onlinegpu.io/models/deepseek-r1-distill-llama-8b): 16.1 GB at BF16, 4.02 GB at 4-bit; cheapest node 1× RTX 3080 at $0.03/h spot.
- [Kimi K2 GPU requirements](https://onlinegpu.io/models/kimi-k2): 2.05 TB at BF16, 513 GB at 4-bit; cheapest node 8× A100 PCIe 80GB at $0.72/h spot.
- [GLM-4.6 GPU requirements](https://onlinegpu.io/models/glm-4-6): 714 GB at BF16, 178 GB at 4-bit; cheapest node 4× A100 PCIe 80GB at $0.36/h spot.
- [GLM-4.5-Air GPU requirements](https://onlinegpu.io/models/glm-4-5-air): 221 GB at BF16, 55.2 GB at 4-bit; cheapest node 1× A100 PCIe 80GB at $0.09/h spot.
- [Qwen3 235B A22B GPU requirements](https://onlinegpu.io/models/qwen3-235b-a22b): 470 GB at BF16, 118 GB at 4-bit; cheapest node 2× A100 PCIe 80GB at $0.18/h spot.
- [Qwen3 Coder 480B A35B GPU requirements](https://onlinegpu.io/models/qwen3-coder-480b-a35b): 960 GB at BF16, 240 GB at 4-bit; cheapest node 4× A100 PCIe 80GB at $0.36/h spot.
- [Qwen3 Next 80B A3B GPU requirements](https://onlinegpu.io/models/qwen3-next-80b-a3b): 163 GB at BF16, 40.7 GB at 4-bit; cheapest node 1× A40 at $0.06/h spot.
- [Qwen3 30B A3B GPU requirements](https://onlinegpu.io/models/qwen3-30b-a3b): 61.1 GB at BF16, 15.3 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Qwen3 Coder 30B A3B GPU requirements](https://onlinegpu.io/models/qwen3-coder-30b-a3b): 61.1 GB at BF16, 15.3 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Qwen3 32B GPU requirements](https://onlinegpu.io/models/qwen3-32b): 65.5 GB at BF16, 16.4 GB at 4-bit; cheapest node 2× RTX A4000 at $0.06/h spot.
- [Qwen3 14B GPU requirements](https://onlinegpu.io/models/qwen3-14b): 29.5 GB at BF16, 7.39 GB at 4-bit; cheapest node 1× RTX A4000 at $0.03/h spot.
- [Qwen3 8B GPU requirements](https://onlinegpu.io/models/qwen3-8b): 16.4 GB at BF16, 4.1 GB at 4-bit; cheapest node 1× RTX 3080 at $0.03/h spot.
- [QwQ 32B GPU requirements](https://onlinegpu.io/models/qwq-32b): 65.5 GB at BF16, 16.4 GB at 4-bit; cheapest node 2× RTX A4000 at $0.06/h spot.
- [Qwen2.5 72B GPU requirements](https://onlinegpu.io/models/qwen2-5-72b): 145 GB at BF16, 36.4 GB at 4-bit; cheapest node 1× A100 PCIe 80GB at $0.09/h spot.
- [Qwen2.5 Coder 32B GPU requirements](https://onlinegpu.io/models/qwen2-5-coder-32b): 65.5 GB at BF16, 16.4 GB at 4-bit; cheapest node 2× RTX A4000 at $0.06/h spot.
- [Llama 4 Maverick GPU requirements](https://onlinegpu.io/models/llama-4-maverick): 803 GB at BF16, 201 GB at 4-bit; cheapest node 4× A100 PCIe 80GB at $0.36/h spot.
- [Llama 4 Scout GPU requirements](https://onlinegpu.io/models/llama-4-scout): 217 GB at BF16, 54.3 GB at 4-bit; cheapest node 1× A100 PCIe 80GB at $0.09/h spot.
- [Llama 3.1 405B GPU requirements](https://onlinegpu.io/models/llama-3-1-405b): 812 GB at BF16, 203 GB at 4-bit; cheapest node 4× A100 PCIe 80GB at $0.36/h spot.
- [Llama 3.3 70B GPU requirements](https://onlinegpu.io/models/llama-3-3-70b): 141 GB at BF16, 35.3 GB at 4-bit; cheapest node 1× A100 PCIe 80GB at $0.09/h spot.
- [Llama 3.1 8B GPU requirements](https://onlinegpu.io/models/llama-3-1-8b): 16.1 GB at BF16, 4.02 GB at 4-bit; cheapest node 1× RTX 3080 at $0.03/h spot.
- [Gemma 3 27B GPU requirements](https://onlinegpu.io/models/gemma-3-27b): 54.9 GB at BF16, 13.7 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Gemma 3 12B GPU requirements](https://onlinegpu.io/models/gemma-3-12b): 24.4 GB at BF16, 6.1 GB at 4-bit; cheapest node 1× RTX 3080 at $0.03/h spot.
- [Mistral Small 3.2 24B GPU requirements](https://onlinegpu.io/models/mistral-small-3-2): 48.0 GB at BF16, 12.0 GB at 4-bit; cheapest node 1× A30 at $0.05/h spot.
- [Mixtral 8x7B GPU requirements](https://onlinegpu.io/models/mixtral-8x7b): 93.4 GB at BF16, 23.4 GB at 4-bit; cheapest node 2× RTX A4000 at $0.06/h spot.
- [Mistral 7B GPU requirements](https://onlinegpu.io/models/mistral-7b): 14.5 GB at BF16, 3.63 GB at 4-bit; cheapest node 1× RTX 3080 at $0.03/h spot.
- [Phi-4 GPU requirements](https://onlinegpu.io/models/phi-4): 29.3 GB at BF16, 7.33 GB at 4-bit; cheapest node 1× RTX A4000 at $0.03/h spot.
- [LLM VRAM calculator](https://onlinegpu.io/tools/vram-calculator): weights + KV cache for a context length and number of concurrent requests, and the GPU nodes that fit.
- [GPU rent vs buy calculator](https://onlinegpu.io/tools/gpu-rent-vs-buy): the utilisation where owning a GPU beats renting it.

## Documentation
- [Overview](https://onlinegpu.io/docs): What the platform does and where everything lives.
- [Quickstart](https://onlinegpu.io/docs/quickstart): From an empty account to a running GPU in five steps.
- [Interruption handling](https://onlinegpu.io/docs/interruptions): Catch the reclaim notice, checkpoint, resume.
- [Disks and data](https://onlinegpu.io/docs/storage): NVMe scratch, persistent volumes, moving data in and out.
- [Service limits](https://onlinegpu.io/docs/limits): Quotas, node sizes, rate limits and how to raise them.
- [CLI](https://onlinegpu.io/docs/cli): Install the client, every command and flag.
- [API reference](https://onlinegpu.io/docs/api): REST endpoints, authentication, webhooks, errors.
- [Changelog](https://onlinegpu.io/changelog): What changed on the platform, newest first.
- [Public pricing feed (JSON)](https://onlinegpu.io/pricing.json): every model with spot, on-demand and reserved price per GPU-hour, availability, node sizes and regions; no authentication.

## Workloads
- [Serving a language model](https://onlinegpu.io/use-cases/llm-inference): Hold the weights, read them fast, answer requests. Memory capacity decides what runs; memory bandwidth decides how quickly.
- [Fine-tuning and training](https://onlinegpu.io/use-cases/fine-tuning): Optimiser states need several times the memory of the weights. Checkpoints make the job interruptible, and interruptible is cheap.
- [Image and video generation](https://onlinegpu.io/use-cases/image-generation): Diffusion models are small enough to fit almost anywhere; throughput comes from bandwidth and from keeping the pipeline resident.
- [Batch and offline jobs](https://onlinegpu.io/use-cases/batch-jobs): Embeddings, evaluation runs, transcoding, rendering, data processing: work that can stop and resume — the natural home of spot capacity.

## Company, status and policies
- [About](https://onlinegpu.io/about): how the company works and what it deliberately does not do.
- [Trust and security](https://onlinegpu.io/trust): isolation, encryption, account security, vulnerability disclosure, and the honest list of gaps.
- [Status](https://onlinegpu.io/status): live state of every component, 90-day uptime, response times, every incident and maintenance window since launch (Atom feed at https://onlinegpu.io/status/feed).
- [Incident history](https://onlinegpu.io/status/history): month by month, with the updates posted at the time.
- [Support](https://cloud.onlinegpu.io/support): tickets in the console, answered 24/7.
- [Terms of service](https://onlinegpu.io/legal/terms): The contract between you and OnlineGPU. Revised 2026-09-15.
- [Acceptable use policy](https://onlinegpu.io/legal/aup): What you may and may not run on our GPUs. Revised 2026-09-15.
- [Service level agreement](https://onlinegpu.io/legal/sla): Availability targets for on-demand instances, and credits. Revised 2026-09-15.
- [Privacy policy](https://onlinegpu.io/legal/privacy): What we collect, why, how long we keep it. Revised 2026-09-15.
- [Data processing addendum](https://onlinegpu.io/legal/dpa): Terms that apply when we process personal data for you. Revised 2026-09-15.

## Optional
- [Changelog](https://onlinegpu.io/changelog): platform changes, newest first.
- [Templates](https://onlinegpu.io/templates): 9 ready-made environments (PyTorch, vLLM, ComfyUI, ROCm…).
- [Console](https://cloud.onlinegpu.io): customer area — account, deployment, billing, support.
- [Full text export](https://onlinegpu.io/llms-full.txt): every model, every price, node sizes, FAQ, comparisons and status in one plain-text file.

Generated 2026-11-01 19:30 UTC from the live catalogue. Prices in USD per GPU-hour.