Quick Answer

io.net offers access to 200,000+ GPUs across 130+ countries, including NVIDIA's latest H100 SXM and H100 PCIe, A100 80GB/40GB, L40S, RTX 4090, RTX 4080, A40, and RTX 3090 Ti. Unlike traditional cloud providers with limited inventory and waitlists, io.net's decentralized network aggregates GPUs from independent providers worldwide, ensuring instant availability for both consumer and enterprise workloads. Prices range from $0.18/hr for RTX 4090 (ideal for inference) to $2.20/hr for H100 SXM (optimal for training), with most GPU types available on-demand in under 2 minutes.

Complete GPU Inventory Breakdown

Here's every GPU type available on io.net with pricing, availability, and ideal use cases:

GPU ModelMemoryCompute (TFLOPS)Price/hrAvailabilityBest For
NVIDIA H100 SXM80GB HBM360 FP16$2.20HighLLM training (70B+), distributed training
NVIDIA H100 PCIe80GB HBM351 FP16$1.49HighLLM inference, fine-tuning (7-13B)
NVIDIA A100 80GB80GB HBM2e39 FP16$1.49Very HighLLM training (7-30B), multi-task workloads
NVIDIA A100 40GB40GB HBM2e39 FP16$1.20Very HighMid-size training, batch inference
NVIDIA L40S48GB GDDR691 FP16$0.75HighInference, rendering, video processing
NVIDIA A4048GB GDDR637 FP16$0.85HighProfessional visualization, AI inference
NVIDIA RTX 409024GB GDDR6X82 FP16$0.18Very HighCost-efficient inference, small models
NVIDIA RTX 408016GB GDDR6X55 FP16$0.35HighGaming, development, light AI workloads
NVIDIA RTX 3090 Ti24GB GDDR6X40 FP16$0.28HighBudget training, inference, image gen
NVIDIA RTX 309024GB GDDR6X35 FP16$0.28Very HighLegacy workloads, cost optimization

Availability status: Very High (95%+ uptime), High (90%+ uptime). Pricing as of April 2026.

GPU Architecture Comparison

Understanding GPU architectures helps you choose the right hardware:

Hopper Architecture (H100):
Released: 2022
Manufacturing: TSMC 4N process
Key features: Transformer Engine (2x faster LLM training), FP8 precision, 80GB HBM3 at 3 TB/s bandwidth
Best for: Llama 3 70B training, GPT-4 scale inference, distributed training clusters
Premium over Ampere: 1.5x performance at 1.4x cost (better value)

Ada Lovelace Architecture (RTX 4090/4080):
Released: 2022
Manufacturing: TSMC 4N process
Key features: DLSS 3, AV1 encoding, 4th gen Tensor Cores
Best for: Llama 3 8B inference, Stable Diffusion generation, ComfyUI workflows
Value proposition: 90% of A100 inference performance at 15% of the cost

Ampere Architecture (A100, A40, RTX 3090):
Released: 2020-2021
Manufacturing: TSMC 7nm
Key features: Multi-Instance GPU (MIG), sparse tensor cores, PCIe Gen 4
Best for: Proven training workloads, established production inference
Mature ecosystem: Widest framework support, most optimization guides

GPU Availability by Region

io.net's decentralized network provides global GPU access:

RegionH100A100RTX 4090L40SAvg. Latency to US
North AmericaHighVery HighVery HighHigh<20ms
Europe (West)HighHighVery HighHigh80-120ms
Europe (East)MediumHighHighMedium100-140ms
Asia-PacificMediumHighVery HighHigh150-200ms
South AmericaLowMediumHighLow120-180ms

Regional Selection Tips:
Lowest cost: Eastern Europe, Southeast Asia (15-25% cheaper electricity)
Lowest latency: Select region closest to your end users
Compliance: EU regions for GDPR, US regions for HIPAA/SOC 2
No regional lock-in: Move workloads between regions instantly

Real-Time GPU Availability Dashboard

Unlike AWS/Azure with instance waitlists, io.net shows live availability:

# Check real-time GPU availability
io availability --gpu all

# Output:
GPU Model          Available  Price/hr  Regions
H100 SXM          1,247      $2.20     US-West, EU-West, APAC
A100 80GB         3,891      $1.49     All regions
A100 40GB         5,203      $1.20     All regions
RTX 4090          28,432     $0.18     All regions
L40S              2,156      $0.75     US-West, EU-West

Availability Guarantees:
RTX 4090/3090: 99.5% availability (28,000+ GPUs)
A100 40GB/80GB: 99.2% availability (9,000+ GPUs)
H100 SXM/PCIe: 98% availability (1,500+ GPUs)
L40S: 95% availability (2,000+ GPUs)

During peak demand (US business hours), availability may temporarily drop to 90-95% for H100s, but you can usually provision within 5-10 minutes.

GPU Selection Guide by Use Case

Large Language Model Training:
| Model Size | GPU Recommendation | Quantity | Rationale |
|------------|-------------------|----------|-----------|
| <7B (Llama 3 8B) | A100 80GB | 1-2 | Fits in single GPU, full fine-tune possible |
| 7-13B | A100 80GB or H100 | 2-4 | LoRA on 1-2 GPUs, full fine-tune on 4 |
| 30-40B | H100 SXM | 4-8 | Requires NVLink, FSDP or DeepSpeed |
| 70B+ | H100 SXM | 8-16 | NVSwitch interconnect critical |

LLM Inference:
| Throughput | GPU Recommendation | Quantity | Cost/Month (24/7) |
|------------|-------------------|----------|-------------------|
| <100 req/day | RTX 4090 | 1 | $130 |
| 1K-10K req/day | RTX 4090 or L40S | 2-4 | $260-2,160 |
| 10K-100K req/day | L40S or H100 PCIe | 4-8 | $2,160-8,582 |
| 100K+ req/day | H100 SXM cluster | 8-32 | $12,672-50,688 |

Image Generation (Stable Diffusion, Midjourney-scale):
| Images/Day | GPU Recommendation | Quantity | Cost/Month |
|------------|-------------------|----------|------------|
| <1,000 | RTX 4090 | 1 | $65 (12hr/day) |
| 1K-10K | RTX 4090 | 2-4 | $130-260 (12hr/day) |
| 10K-100K | RTX 4090 or L40S | 8-16 | $1,036-2,160 (24/7) |
| 100K+ | RTX 4090 cluster | 32+ | $4,147+ (24/7) |

Video Processing and Rendering:
| Workload | GPU Recommendation | Key Features |
|----------|-------------------|--------------|
| Real-time rendering | RTX 4090 | DLSS 3, ray tracing |
| Video encoding | L40S or RTX 4090 | AV1 encoding, NVENC |
| 3D visualization | A40 or L40S | Professional drivers, ECC memory |

GPU Performance Benchmarks

Real-world performance comparison for common AI tasks:

Llama 3 8B Inference (tokens/second):
| GPU | FP16 | INT8 | FP8 (H100) |
|-----|------|------|------------|
| H100 SXM | 142 | 287 | 385 |
| H100 PCIe | 118 | 245 | 320 |
| A100 80GB | 95 | 178 | N/A |
| L40S | 87 | 165 | N/A |
| RTX 4090 | 82 | 156 | N/A |

Stable Diffusion XL (512x512 image generation time):
| GPU | SDXL (default) | SDXL (optimized) |
|-----|----------------|------------------|
| H100 | 0.8s | 0.4s |
| RTX 4090 | 1.2s | 0.6s |
| A100 | 1.4s | 0.7s |
| RTX 3090 | 2.1s | 1.0s |

Fine-tuning Llama 3 8B LoRA (samples/second):
| GPU | Batch 1 | Batch 4 | Batch 16 |
|-----|---------|---------|----------|
| H100 SXM | 3.2 | 11.8 | 42.5 |
| A100 80GB | 2.1 | 7.8 | 28.3 |
| RTX 4090 | 1.8 | 6.5 | 24.1 |

Multi-GPU Configurations

For distributed training and high-throughput inference:

Pre-Configured Clusters:

# 2-GPU NVLink cluster for 13B model training
io launch --gpu A100 --count 2 --network nvlink --disk 200GB

# 8-GPU NVSwitch cluster for 70B model training
io launch --gpu H100 --count 8 --network nvswitch --disk 1TB

# 4-GPU inference cluster with load balancing
io launch --gpu RTX4090 --count 4 --mode inference --autoscale

Cluster Networking Options:

ConfigurationBandwidthLatencyUse CasePremium Cost
PCIe Gen 464 GB/s~5μsIndependent tasksIncluded
NVLink (2-way)600 GB/s<1μsSmall model parallel+$0.10/hr per GPU
NVSwitch (8-way)900 GB/s<1μsLarge model training+$0.25/hr per GPU
InfiniBand (multi-node)200-400 Gb/s~2μs16+ GPU clusters+$0.50/hr per GPU

GPU Availability Alerts and Reservations

Never wait for GPU availability:

Real-Time Alerts:

# Get notified when H100s become available in US-West
io notify --gpu H100 --region us-west --email [email protected]

# Auto-provision when GPU becomes available
io autoprovision --gpu H100 --count 8 --max-price 2.50

Enterprise Reserved Capacity:
For mission-critical workloads requiring guaranteed availability:
Reserved pools: Dedicated GPU allocation for your account
Volume discounts: 10-20% off on top of base pricing
SLA guarantees: 99.9% availability with credits for downtime
Minimum commitment: 10 GPUs for 3 months

Contact [email protected] for reserved capacity pricing.

How quickly can I get a GPU on io.net?

Most GPUs provision in under 2 minutes. RTX 4090 and A100 (highest availability) typically spin up in 30-60 seconds. H100s may take 2-5 minutes during peak demand. There are no reservation queues or waiting lists - if a GPU shows as available, you can provision it immediately. For guaranteed instant access, enterprise plans include reserved capacity.

Can I switch GPU types during my project?

Yes. io.net allows you to stop one instance and start another with a different GPU type instantly. Your data persists in attached storage volumes. For example, you might use RTX 4090 for development ($0.18/hr), then switch to 8x H100 for production training ($17.60/hr), then back to L40S for inference ($0.75/hr). No migration fees or data transfer costs.

What if my preferred GPU type is unavailable?

io.net shows live availability across all regions. If H100s are unavailable in US-West, check EU-West or US-East. The platform also suggests alternative GPUs - for example, 2x A100 can often replace 1x H100 for training at similar total cost. Set up availability alerts to get notified when your preferred GPU becomes available.

Are GPUs shared or dedicated?

All io.net GPUs are dedicated bare-metal instances. You get 100% of the GPU's compute, memory, and bandwidth - no virtualization, no sharing, no "noisy neighbor" issues. This is different from some cloud providers that use MIG (Multi-Instance GPU) to split GPUs. You have root access and full control over the GPU.

How does GPU quality control work on a decentralized network?

io.net runs automated health checks on every GPU every 6 hours: memory tests, compute benchmarks, and thermal monitoring. GPUs that underperform (>10% below expected benchmarks) are automatically removed from the marketplace. Provider reputation scores factor in uptime, performance, and user ratings. You can see provider ratings before provisioning and request GPU replacement if performance issues occur.

Browse io.net's GPU Inventory

Access GPUs instantly:
H100 SXM at $2.20/hr - 68% cheaper than AWS ($6.98/hr)
RTX 4090 at $0.18/hr - Best price-performance for inference
A100 80GB at $1.49/hr - Proven training performance
No waitlists - 99%+ availability across all GPU types

View real-time availability → or launch a GPU now →


Last updated: April 2026 | GPU availability and pricing subject to real-time market conditions