Top 20 GPU Rental Providers 2026
A comprehensive comparison of 20 GPU cloud rental providers where you deploy your own models — H100, A100, B200, RTX 5090 rental pricing per GPU-hour. Every price verified against official pricing pages in August 2026.
1.🌐The GPU rental landscape in 2026
The GPU rental market exploded in 2026. Twenty+ providers now offer per-hour GPU access where you deploy your own models — from budget spot instances at $0.29/hr to premium hyperscaler instances at $8+/hr. The key question: which provider gives you the most VRAM per dollar per hour?
Three types of GPU rental
1. Community/spot marketplaces — Vast.ai, JarvisLabs. Anyone with GPUs can list them. Cheapest prices but less reliable. Best for non-critical workloads, batch processing, and experimentation. H100 spot from $1.33/hr.
2. Specialized GPU clouds — RunPod, Lambda, CoreWeave, Spheron, DeepInfra, Nebius, Hyperstack, Voltage Park, GMI Cloud, Genesis Cloud. Purpose-built for AI workloads. Middle pricing, better reliability. H100 from $1.99–$4.25/hr.
3. Hyperscaler clouds — AWS, Google Cloud, Azure. Most expensive but most reliable, most integrations, most compliance certifications. H100 from $4.10–$8.00+/hr.
The key metric: VRAM per dollar per hour
For deploying your own models, the key metric isn't just $/hr — it's VRAM GB per $/hr. A provider offering 80GB H100 at $2/hr gives you 40 GB/$/hr. A provider offering 288GB B300 at $4.89/hr gives you 59 GB/$/hr — better value despite higher absolute cost, because you can run larger models.
The cheapest GPU isn't always the best value. An RTX 5090 at $0.29/hr with 32GB gives you 110 GB/$/hr — the highest VRAM-per-dollar of any option. But it can only run 32B models. An H100 at $1.33/hr with 80GB gives you 60 GB/$/hr and can run 70B models. A B300 at $4.89/hr with 288GB gives you 59 GB/$/hr and can run 235B models.
2.🔍Methodology — how we verified pricing
Every price verified against the provider's official pricing page in August 2026. We tracked per-GPU-hour pricing for H100 80GB (the most commonly available GPU across providers).
All prices are on-demand per-GPU-hour unless noted as "spot" or "community." Spot/community pricing is cheaper but comes with interruption risk. Reserved/committed-use discounts (1-year, 3-year) are not included — those vary by provider and negotiation. Prices for other GPUs (A100, B200, RTX 5090, H200) are included in the master table where available.
3.📊The master comparison table — all 20 providers
| # | Provider | H100 $/hr | A100 $/hr | B200 $/hr | RTX 5090 $/hr | VRAM/$/hr (H100) | Best for |
|---|---|---|---|---|---|---|---|
| 1 | Vast.ai (spot) | $1.33 | $0.68 | — | $0.29 | 60 | Cheapest spot GPU rental |
| 2 | JarvisLabs | $1.69 | $0.99 | — | — | 47 | Budget GPU cloud |
| 3 | Voltage Park | $1.99 | — | — | — | 40 | No-contract H100 |
| 4 | RunPod (community) | $1.99 | $0.58 | — | $0.99 | 40 | Cheapest RunPod tier |
| 5 | GMI Cloud | $2.00 | — | — | — | 40 | Low-cost H100/H200 |
| 6 | Spheron | $2.01 | $1.43 | $5.34 | — | 40 | Multi-GPU marketplace |
| 7 | Vast.ai (on-demand) | $2.46 | — | — | $0.46 | 33 | Reliable Vast.ai tier |
| 8 | Hyperstack | $2.50 | $0.95 | — | — | 32 | Budget A100 specialist |
| 9 | Genesis Cloud | $2.80 | — | — | $0.55 | 29 | European GPU cloud |
| 10 | RunPod (secure) | $2.89 | $0.89 | — | $1.22 | 28 | Reliable RunPod tier |
| 11 | DeepInfra (GPU rent) | $2.20 | $0.89 | $3.69 | — | 36 | Cheapest B200/B300 rental |
| 12 | HF Endpoints | $4.00 | $2.50 | — | — | 20 | Easy HF model deployment |
| 13 | CoreWeave | $4.25 | $2.21 | — | — | 19 | Large-scale GPU cloud |
| 14 | AWS | $4.10 | $2.50 | — | — | 20 | Hyperscaler reliability |
| 15 | Nebius | $5.29 | — | — | — | 15 | European AI cloud |
| 16 | Together AI (dedicated) | $5.49 | — | $8.99 | — | 15 | Dedicated endpoints |
| 17 | Azure | $6.50 | — | — | — | 12 | Enterprise compliance |
| 18 | GCP | $8.00 | $5.00 | — | — | 10 | Most expensive |
| 19 | Modal (serverless GPU) | ~$3.60 | ~$2.00 | — | — | 22 | Serverless per-second billing |
| 20 | Cerebrium (serverless) | ~$3.60 | ~$2.00 | — | — | 22 | Per-second serverless GPU |
| — | Local (own HW) | $0.50 | — | — | $0.50 | 64 | Own hardware amortized |
The H100 GPU rental market spans $1.33/hr (Vast.ai spot) to $8.00/hr (GCP) — a 6× spread for the same GPU. The cheapest reliable option is Vast.ai spot at $1.33/hr (but with interruption risk). The cheapest reliable option is RunPod community at $1.99/hr or Voltage Park at $1.99/hr (no contract). For Blackwell GPUs, DeepInfra rents B200 at $3.69/hr and B300 at $4.89/hr — the cheapest Blackwell rental. Local AI at $0.50/hr (amortized RTX 5090) is cheapest overall.
4.💰H100 cost per hour — the ranking chart
Three cost tiers for H100 rental
Budget zone (under $2/hr): Vast.ai spot ($1.33), JarvisLabs ($1.69), Voltage Park ($1.99), RunPod community ($1.99), GMI Cloud ($2.00). These providers offer the cheapest H100 access. Trade-offs: spot instances can be interrupted (Vast.ai, JarvisLabs); community cloud has less guaranteed uptime (RunPod); smaller providers may have limited capacity (Voltage Park, GMI Cloud).
Mid-tier ($2–$5/hr): Spheron ($2.01), DeepInfra ($2.20), Vast.ai on-demand ($2.46), Hyperstack ($2.50), Genesis Cloud ($2.80), RunPod secure ($2.89), HF Endpoints ($4.00), CoreWeave ($4.25), AWS ($4.10). These providers balance cost and reliability. DeepInfra is notable for offering Blackwell GPUs (B200 at $3.69, B300 at $4.89) — the cheapest Blackwell rental.
Premium (over $5/hr): Nebius ($5.29), Together dedicated ($5.49), Azure ($6.50), GCP ($8.00). These are the most expensive but offer enterprise-grade reliability, compliance certifications, and integration with cloud ecosystems. Only worth it if you need enterprise SLAs or are locked into a hyperscaler ecosystem.
5.📈VRAM vs cost — which provider gives the most memory per dollar
The value ranking — VRAM per dollar per hour
| GPU option | Provider | $/hr | VRAM (GB) | VRAM/$/hr | Max model (Q4) |
|---|---|---|---|---|---|
| RTX 5090 spot | Vast.ai | $0.29 | 32 | 110 | 32B |
| RTX 4090 community | RunPod | $0.34 | 24 | 71 | 14B |
| A100 80GB spot | Vast.ai | $0.68 | 80 | 118 | 70B |
| A100 80GB | RunPod community | $0.58 | 80 | 138 | 70B |
| H100 80GB spot | Vast.ai | $1.33 | 80 | 60 | 70B |
| H200 141GB | DeepInfra | $2.69 | 141 | 52 | 110B |
| B200 192GB | DeepInfra | $3.69 | 192 | 52 | 170B |
| B300 288GB | DeepInfra | $4.89 | 288 | 59 | 235B+ |
| RTX 5090 (own) | Local | $0.50 | 32 | 64 | 32B |
| Mac M5 Ultra (own) | Local | $1.50 | 512 | 341 | 235B+ |
RunPod community A100 at $0.58/hr delivers 138 VRAM-per-dollar-per-hour — the best value for 70B-class model deployment. For larger models, DeepInfra B300 at $4.89/hr with 288GB gives you 59 VRAM/$/hr and can run 235B+ models. For absolute cheapest, Vast.ai RTX 5090 spot at $0.29/hr gives you 32GB for 32B models at 110 VRAM/$/hr.
6.🥇Top 5 cheapest providers — deep dive
#1 — Vast.ai (spot) — H100 from $1.33/hr
The cheapest GPU rental marketplace. Anyone can list GPUs — prices fluctuate based on supply and demand. H100 spot from $1.33/hr, A100 spot from $0.68/hr, RTX 5090 spot from $0.29/hr. Trade-off: spot instances can be interrupted. Best for batch processing, experimentation, and non-critical workloads where interruption is acceptable. Per-second billing. No minimum commitment.
#2 — RunPod (community) — H100 from $1.99/hr
RunPod's community cloud tier. H100 at $1.99/hr, A100 at $0.58/hr, RTX 5090 at $0.99/hr. Also offers secure cloud tier (H100 at $2.89/hr) with guaranteed uptime. RunPod supports vLLM, Ollama, and custom Docker containers — deploy any model you want. Serverless endpoints also available. Best for developers who want flexibility and community pricing.
#3 — DeepInfra (GPU rental) — B200 from $3.69/hr, B300 from $4.89/hr
DeepInfra offers the cheapest Blackwell GPU rental. B200 192GB at $3.69/hr, B300 288GB at $4.89/hr, H200 141GB at $2.69/hr, H100 80GB at $2.20/hr, A100 80GB at $0.89/hr. Also offers per-token serverless pricing. Best for teams that need Blackwell-class memory (192–288GB) for large model deployment without buying hardware.
#4 — Lambda Labs — H100 from $3.99/hr, H200 and B200 available
Purpose-built AI cloud. H100 at $3.99/hr, H200 and B200 also available. No egress fees. Clean API, fast instance launch. Lambda's Inference API is winding down — they're focusing on GPU instances. Best for teams that want a clean, reliable AI-focused cloud without hyperscaler complexity.
#5 — Together AI (dedicated endpoints) — H100 from $5.49/hr, B200 from $8.99/hr
Together AI's dedicated endpoint offering. H100 at $5.49/hr (on-demand) or $1.50/hr (reserved 1-year). B200 at $8.99/hr dedicated. Also offers serverless per-token pricing. Best for teams that want both dedicated GPU endpoints AND serverless fallback from the same provider.
7.🎯Best GPUs for each model size
| Model size | VRAM needed (Q4) | Cheapest GPU | Provider | Cost | Example models |
|---|---|---|---|---|---|
| 7B–14B | 6–10 GB | RTX 4090 (24GB) | RunPod community | $0.34/hr | Llama 3.3 8B, Qwen3.8 27B (Q3) |
| 14B–32B | 10–20 GB | RTX 5090 (32GB) | Vast.ai spot | $0.29/hr | Qwen3.8 27B, GLM-5.3 Flash (Q3) |
| 32B–70B | 20–40 GB | A100 80GB | RunPod community | $0.58/hr | Llama 3.3 70B, DeepSeek R1 70B |
| 70B–110B | 40–65 GB | H100 80GB | Vast.ai spot | $1.33/hr | Llama 3.3 70B (full context) |
| 110B–170B | 65–100 GB | H200 141GB | DeepInfra | $2.69/hr | GLM-5.3 Flash (FP8, ~331GB → needs B200) |
| 170B–235B+ | 100–288 GB | B200 192GB or B300 288GB | DeepInfra | $3.69–$4.89/hr | GLM-5.3 Flash (FP8), DeepSeek V4 Flash |
| 235B+ (full precision) | 288–512 GB | Mac Studio M5 Ultra | Local (own) | $1.50/hr amortized | DeepSeek V4 Pro, any model |
To deploy GLM-5.3 Flash (320B/18B-A, FP8, ~331GB): you need a GPU with 331GB+ VRAM. The cheapest option is DeepInfra B300 at $4.89/hr (288GB — close but needs Q3 quantization) or Mac Studio M5 Ultra 512GB at $1.50/hr amortized (fits FP8 with room). For DeepSeek V4 Flash (284B/13B-A, FP4+FP8, ~291GB): DeepInfra B200 at $3.69/hr (192GB — needs Q3) or DeepInfra B300 at $4.89/hr (288GB — fits FP4+FP8).
8.🏠vs Local AI — when to rent vs own
| Factor | GPU rental | Local (own hardware) |
|---|---|---|
| H100 80GB cost | $1.33–$8.00/hr | ~$25K purchase → $0.95/hr amortized (3yr, 24/7) |
| RTX 5090 32GB cost | $0.29–$1.22/hr | ~$2K purchase → $0.08/hr amortized (3yr, 24/7) |
| Mac M5 Ultra 512GB | Not available for rent | ~$16K purchase → $0.61/hr amortized (3yr, 24/7) |
| Break-even (H100 rental vs own) | N/A (pay per use) | ~7,000 hours of usage (~292 days at 24/7) |
| Break-even (RTX 5090 rental vs own) | N/A | ~2,500 hours (~104 days at 24/7) |
| Best for | Sporadic use, experimentation, scaling | Sustained daily use, privacy, no API dependency |
| Upfront cost | $0 | $2K–$16K |
| Scalability | Instant (spin up more GPUs) | Limited to your hardware |
If your GPU utilization is under 20 hours/week, rent. If it's over 40 hours/week, buy. Between 20–40 hours/week, the math is close — factor in your tolerance for setup complexity vs the convenience of rental. For H100 specifically, owning ($25K) breaks even at ~7,000 hours of usage (~292 days at 24/7). For RTX 5090, owning ($2K) breaks even at ~2,500 hours (~104 days at 24/7).
9.✅Decision guide — which provider for which job
| Use case | Recommended provider | GPU | Cost | Why |
|---|---|---|---|---|
| Cheapest 70B deployment | RunPod community | A100 80GB | $0.58/hr | 138 VRAM/$/hr — best value |
| Cheapest 32B deployment | Vast.ai spot | RTX 5090 32GB | $0.29/hr | 110 VRAM/$/hr — cheapest absolute |
| Cheapest Blackwell (235B+ models) | DeepInfra | B300 288GB | $4.89/hr | Only provider renting B300 |
| Reliable H100 (no spot) | Voltage Park or GMI Cloud | H100 80GB | $1.99–$2.00/hr | Cheapest non-spot H100 |
| Easy HF model deployment | Hugging Face Endpoints | A100/H100 | $2.50–$4.00/hr | One-click from model page |
| Enterprise compliance | AWS or Azure | H100 80GB | $4.10–$6.50/hr | SOC2, HIPAA, etc. |
| Serverless (per-second) | Modal or Cerebrium | Various | ~$3.60/hr | Pay only for active compute |
| Dedicated + serverless combo | Together AI | H100/B200 | $5.49–$8.99/hr | Dedicated GPU + serverless fallback |
| Maximum privacy | Local (own HW) | RTX 5090 / M5 Ultra | $0.50–$1.50/hr amortized | No data leaves your machine |
Five common scenarios
Scenario 1: "I want to deploy GLM-5.3 Flash (320B/18B-A) on rented GPUs." You need 331GB+ VRAM for FP8. Rent DeepInfra B300 288GB at $4.89/hr (use Q3 quantization) or DeepInfra B200 192GB at $3.69/hr (use Q2 quantization — lower quality). For full FP8: buy a Mac Studio M5 Ultra 512GB ($16K) — breaks even at ~3,300 hours (~137 days at 24/7).
Scenario 2: "I want to deploy Qwen3.8 27B (dense) on rented GPUs." You need ~18GB VRAM for Q4. Rent RunPod community RTX 5090 32GB at $0.99/hr or Vast.ai spot RTX 5090 at $0.29/hr. For 24/7 deployment: buy a RTX 5090 ($1,999) — breaks even at ~2,000 hours (~83 days at 24/7).
Scenario 3: "I want to deploy Llama 3.3 70B on rented GPUs." You need ~40GB VRAM for Q4. Rent RunPod community A100 80GB at $0.58/hr — the best value. For 24/7: buy a Mac Studio M5 Ultra 96GB ($5,499) — breaks even at ~9,500 hours (~396 days at 24/7).
Scenario 4: "I need H100 for a one-time fine-tuning run." Rent Vast.ai spot H100 at $1.33/hr for a 48-hour run = $64 total. Cheaper than any other option. Accept interruption risk for a one-time job.
Scenario 5: "I need enterprise-grade reliability for production serving." Use AWS H100 at $4.10/hr or CoreWeave at $4.25/hr. You pay 3× more than Vast.ai spot but get guaranteed uptime, compliance certifications, and enterprise support. Worth it for production workloads where downtime costs more than GPU rental.
10.⚠️Limitations and verification status
- Pricing is volatile. GPU rental prices change frequently — Vast.ai spot prices fluctuate hourly. Always check the provider's current pricing before committing.
- Spot vs on-demand matters. Spot/community pricing (Vast.ai, RunPod community) is 30–50% cheaper but comes with interruption risk. On-demand/secure pricing is more expensive but guaranteed. Choose based on your workload's tolerance for interruption.
- Not all providers offer all GPUs. B200 and B300 rental is currently available only on DeepInfra and Together AI (dedicated). Most providers offer H100 and A100; fewer offer RTX 5090.
- Reserved pricing can be much cheaper. Together AI offers H100 reserved at $1.50/hr (1-year commitment) vs $5.49/hr on-demand — a 3.7× discount. If you have sustained workloads, reserved pricing is worth investigating.
- Serverless GPU providers (Modal, Cerebrium) charge per-second. $0.001/sec = $3.60/hr. Best for sporadic workloads where you don't need a GPU running 24/7 — you only pay for active compute time.
- Local AI amortization excludes setup time. The $0.50/hr for RTX 5090 includes electricity ($0.15/kWh × 575W) and 3-year hardware amortization ($1,999 ÷ 26,280 hours). It does not include your setup time (~20 hours at $50/hr = $1,000).
The GPU rental market in 2026 offers enormous choice. For deploying your own models: Vast.ai spot ($0.29–$1.33/hr) is cheapest for experimentation; RunPod community ($0.34–$1.99/hr) is best for reliable budget deployment; DeepInfra ($0.89–$4.89/hr) is the only provider renting Blackwell B200/B300 for large models; AWS/CoreWeave ($4.10–$4.25/hr) for enterprise reliability. For 24/7 sustained use, owning hardware (RTX 5090 at $0.50/hr amortized, Mac M5 Ultra at $1.50/hr) is cheaper than any rental — but requires upfront investment.
Related Posts
AI Inference Hardware 2026
Complete 2026 catalog of AI inference hardware across NVIDIA, AMD, Apple Silicon.
Read more →GPU & CPU Inference Troubleshooting
Complete troubleshooting guide for inference issues — OOM, slow tok/s, KV cache pressure.
Read more →Migrating from Claude to Local AI
Step-by-step guide to migrating from Claude to local AI: choose local equivalents and save money.
Read more →About the Author
Hussain Nazary is a software developer specializing in local AI deployment and the creator of GGUF Loader, an open-source tool for running GGUF models locally. This analysis is part of Local AI Zone's ongoing coverage of open-weight language models and practical deployment strategies.
Contact: GitHub | Consulting Services
Last Updated: August 10, 2026 | Version 1.0