Q: How often is GPU Tracker pricing data updated?

GPU Tracker refreshes pricing every 6 hours through automated scrapers and provider APIs across 54+ clouds. Each price entry carries a timestamp so users can verify freshness. The full dataset (5,000+ instances) is free to browse and is also available via a $199/mo JSON API.

Q: Which GPU cloud providers does GPU Tracker cover?

GPU Tracker covers 54+ providers including AWS, Azure, GCP, Oracle Cloud, CoreWeave, Lambda Labs, RunPod, Vast.ai, TensorDock, Modal, Lightning.ai, FluidStack, Cudo Compute, Nebius, Vultr, Scaleway, OVHcloud, Hyperstack, Crusoe, Nscale, Verda, E2E Networks, Shadeform, and many more hyperscalers and specialty GPU clouds.

Question 1

What is the cheapest GPU cloud provider in 2026?

Accepted Answer

For budget and spot instances, Vast.ai and TensorDock are cheapest — A100 from $0.08/hr (spot), RTX 4090 from $0.20/hr. For reliable on-demand, RunPod leads (A100 from $0.34/hr). For enterprise SLAs, CoreWeave, AWS, Azure, and GCP cost 3–10x more. GPU Tracker compares live prices across 54+ providers, refreshed every 6 hours.

Question 2

How much does an H100 GPU cost per hour?

Accepted Answer

NVIDIA H100 80GB pricing ranges from $0.80/hr (Verda, specialty providers) to $8.46/hr (AWS p5.48xlarge per-GPU) to $97.44/hr on premium enterprise tiers. The median cheap-tier on-demand H100 is roughly $1.90–$2.50/hr. H100 SXM5 sits slightly higher than PCIe variants.

Question 3

How much does an A100 GPU cost per hour?

Accepted Answer

NVIDIA A100 80GB prices start at $0.08/hr on Vast.ai spot and $0.34/hr on RunPod on-demand. Enterprise providers like AWS and Azure charge $3–$5/hr. A100 40GB is typically 20–30% cheaper. Over 350 live A100 listings are tracked on GPU Tracker.

Question 4

What is the best GPU for LLM inference?

Accepted Answer

For 7B–13B models, RTX 4090 (24GB) or L40S (48GB) offer the best price-performance at $0.20–$0.80/hr. For 70B models, A100 80GB or H100 80GB are required. For 180B+ models, H200 (141GB) or multi-GPU H100 configurations. Quantization (Q4/Q8) halves the VRAM needed.

Question 5

What is the best GPU for training large language models?

Accepted Answer

H100 SXM5 and H200 are the standard for training frontier LLMs (70B+). H100 delivers ~3x the training throughput of A100 thanks to Transformer Engine and FP8. B200 and GB200 are emerging for the largest jobs. For smaller research models (7B–13B), A100 80GB remains cost-effective.

Question 6

How much VRAM do I need for an LLM?

Accepted Answer

Rule of thumb: parameters (B) × 2 bytes (fp16) = GB VRAM needed, plus 20% for KV cache. 7B model: ~14GB fp16 or 4GB Q4. 13B: ~26GB fp16, 7GB Q4. 70B: ~140GB fp16, 35GB Q4. Add context length: 128K context adds ~8–16GB.

Question 7

What is the difference between spot and on-demand GPU instances?

Accepted Answer

Spot instances are interruptible and cost 40–70% less than on-demand. RunPod spot H100: ~$1.20/hr vs $1.87/hr on-demand. Vast.ai A100 spot: $0.08/hr vs $0.34/hr on-demand. Use spot for fault-tolerant training with checkpointing; use on-demand for production inference or time-sensitive jobs.

Question 8

Is renting a GPU cheaper than buying?

Accepted Answer

Buying an H100 ($25K–$40K) breaks even vs renting at ~$2/hr in roughly 14–22 months of 24/7 use. Rent if you need <12 months, bursty workloads, or latest hardware. Buy if you need 24/7 utilization for 2+ years and can absorb depreciation. GPU Tracker's buy-vs-rent calculator runs this math live.

Question 9

How often is GPU Tracker pricing data updated?

Accepted Answer

GPU Tracker refreshes pricing every 6 hours through automated scrapers and provider APIs across 54+ clouds. Each price entry carries a timestamp so users can verify freshness. The full dataset (5,000+ instances) is free to browse and is also available via a $199/mo JSON API.

Question 10

Which GPU cloud providers does GPU Tracker cover?

Accepted Answer

GPU Tracker covers 54+ providers including AWS, Azure, GCP, Oracle Cloud, CoreWeave, Lambda Labs, RunPod, Vast.ai, TensorDock, Modal, Lightning.ai, FluidStack, Cudo Compute, Nebius, Vultr, Scaleway, OVHcloud, Hyperstack, Crusoe, Nscale, Verda, E2E Networks, Shadeform, and many more hyperscalers and specialty GPU clouds.

Provider	Instance	GPUs	VRAM	Region	Type	Price/hr
CHEAPESTCloudRift	mi350x-15-250-1000-gv.1	1×	288GB	EU-Central	On-Demand	$3.9000	Rent
CloudRift	mi350x-15-250-1000-gv.2	2×	576GB	EU-Central	On-Demand	$7.8000	Rent
CloudRift	mi350x-15-250-1000-gv.3	3×	864GB	EU-Central	On-Demand	$11.7000	Rent
CloudRift	mi350x-15-250-1000-gv.4	4×	1152GB	EU-Central	On-Demand	$15.6000	Rent
CloudRift	mi350x-15-250-1000-gv.5	5×	1440GB	EU-Central	On-Demand	$19.5000	Rent
CloudRift	mi350x-15-250-1000-gv.6	6×	1728GB	EU-Central	On-Demand	$23.4000	Rent
CloudRift	mi350x-15-250-1000-gv.7	7×	2016GB	EU-Central	On-Demand	$27.3000	Rent
CloudRift	mi350x-15-250-1000-gv.8	8×	2304GB	EU-Central	On-Demand	$31.2000	Rent

MI350X GPU Cloud Prices

All MI350X Instances — Cheapest First

Compare MI350X vs

Browse Other GPU Models

Best For (by Use Case)

Browse by VRAM Tier

MI350X GPU Cloud Rental — Frequently Asked Questions