deploybase

RTX 4090 on AWS: Pricing, Availability & Setup

Deploybase · April 8, 2025 · GPU Pricing

Contents

AWS does not currently offer NVIDIA RTX 4090 GPUs within its infrastructure. The company focuses on professional-class GPUs including L4, A100, and H100 units specifically designed for production inference and training workloads. Teams seeking RTX 4090 capability should deploy on alternative providers, while AWS users should evaluate professional GPU alternatives for comparable inference performance.

AWS GPU Portfolio and RTX 4090 Absence

RTX 4090 on AWS simply doesn't exist. AWS won't carry it. The company focuses on professional-grade hardware: L4 for inference, A100 for general compute, H100 for heavy training work. That's it.

Why? AWS prioritizes production reliability and SLA commitments over cost-per-hour. Consumer GPUs like the 4090 don't fit that model. No professional support guarantees, no redundancy. Just a different category of hardware.

The L4 is as close as developers'll get on AWS. It's the inference workhorse. For production setups that already live in AWS, the L4 makes sense. For everyone else chasing raw price-to-performance, RunPod or Vast.ai will eat AWS's lunch.

RTX 4090 availability on AWS? Not happening. AWS's strategy is locked in: professional hardware, production guarantees, premium pricing.

AWS L4 as RTX 4090 Alternative

The L4 is what AWS offers if developers need that 24GB VRAM the 4090 has. Same amount, different card entirely.

Raw specs tell the story: L4 hits 30.3 TFLOPS FP32 (vs 4090's 82.6 TFLOPS FP32), though L4's TF32 tensor performance reaches 242 TFLOPS. Memory bandwidth is 300 GB/s on the L4, 1,008 GB/s on the 4090. On paper? The 4090 wins. But L4 has specialized inference optimizations that actually matter for transformer models. Developers won't notice the raw spec gap if developers're running LLM inference.

The catch: AWS charges $0.80/hour for g6.xlarge (one L4). RunPod RTX 4090? $0.34/hour. That's more than 2x more expensive for similar inference work. The premium buys developers AWS integration and SLA guarantees, not raw performance.

AWS g6 Instance Specifications

g6 instances scale from g6.xlarge (1 L4) to g6.12xlarge (8 L4s). Pick one L4 for experiments, stack them if developers're running multiple models concurrently.

The base: g6.xlarge has 4 vCPU, 16 GB RAM, one L4. Good for 10-100 concurrent inference requests. Step up to g6.2xlarge and developers get more CPU/RAM with 2 L4s. Scale linearly from there.

Storage comes from EBS (gp3 volume, ~1000 MB/s). Network: 25 Gbps on the xlarge, up to 100 Gbps on larger sizes. Nothing fancy, but standard cloud infrastructure.

Performance Characteristics of AWS L4

L4 pushes 30-80 tokens/sec for quantized language models. Depends on the model size. Bigger models actually perform better because batch sizes pack the GPU harder.

Time-to-first-token stays under 100ms for most models. Developers can batch 8-16 concurrent requests per GPU. That's enough for production inference services.

The 4090 has faster raw tensor ops, sure. But for transformer inference? L4's optimizations close that gap. Developers won't notice it in practice.

Need more throughput? Add more L4s. g6.2xlarge with 2 L4s handles double the workload or bigger models needing more VRAM.

Availability and Regional Distribution

L4s are available in most major regions: us-east-1, us-west-2, eu-west-1, and a few others. Availability fluctuates. Check AWS's page before committing.

Multi-AZ deployments give developers failover. Run across zones if uptime matters.

Reserve capacity if developers're planning to stick around. Reserved instances run 40-50% cheaper than on-demand. On-demand lets developers scale elastically but costs more per hour. Spot instances cost even less:good for fault-tolerant workloads.

Cost Analysis and Economic Justification

One L4 on g6.xlarge runs about $576/month (720 hours at on-demand $0.80/hr). Reserve it and drop to $280-320/month. That's the AWS cost reality.

RunPod's 4090 at $0.34/hour is less than half the cost. That gap matters over time. AWS charges developers for infrastructure, SLA, managed services. Developers get what developers pay for.

Running multiple L4s? Reserved instances + spot purchasing brings the cost down. Spot instances are 30% of on-demand. Good for workloads that can handle interruptions.

Don't forget data transfer costs. Large model downloads and external traffic cost extra. That bandwidth multiplier adds up fast.

Deployment Workflows on AWS

AWS ships pre-built Deep Learning containers with vLLM, Text Generation WebUI, and similar tools. Use them. Saves time getting g6 instances running.

SageMaker Endpoints let developers deploy models on g6 instances with auto-scaling. Load balancing is automatic. Good if developers already live in the SageMaker ecosystem.

ECS or EKS for orchestration. Both work fine. They distribute workloads across the L4s as capacity fills up.

Auto Scaling kicks in when request load spikes. New instances launch automatically, so developers don't spike latency during traffic spikes.

Networking and Integration

VPC keeps instances private. Security groups and ACLs control who hits the endpoints.

ELB spreads traffic across multiple g6 instances. Health checks pull failed ones out of rotation automatically.

VPN or Direct Connect if developers need on-premises connectivity. Secure tunneling for compliance workloads.

S3 for model storage. Stream directly to GPU memory rather than burning through EBS. Cleaner, cheaper approach.

Comparison to Alternative GPU Providers

RunPod RTX 4090 runs $0.34/hour. That's 55% cheaper than AWS's L4. Simple math.

Vast.AI's marketplace is even cheaper: $0.20-0.40/hour for 4090s. 60-80% savings if developers can handle marketplace variability (instances disappear without warning).

Need more power? B200 on AWS at $80-100/hour crushes L4 for training and large-scale inference. But developers're paying for performance, not inference efficiency.

AWS's value is infrastructure, SLA, support. Developers pay the premium for that. If developers need those guarantees, it's worth it. If not, developers're leaving money on the table.

When to Choose AWS L4 vs RTX 4090 Alternatives

Already in AWS? L4 makes sense for avoiding migration friction. Switching providers means operational work. Sometimes that cost is bigger than the hourly rate savings.

Running production inference that needs 99%+ uptime? AWS's SLA and managed infrastructure justify the premium. That's the actual value proposition.

HIPAA, SOC2, or other compliance needs? AWS's certification infrastructure handles that. Regulatory workloads often require it anyway.

Single-cloud policy at the company? Developers'll eat the AWS premium. That's a business decision, not a technical one.

When Not to Choose AWS

No AWS lock-in? RunPod or Vast.AI save developers 40-80%. That's real money.

Testing concepts? Don't waste money on AWS SLA. Go cheap, validate, then upgrade if it sticks.

Comfortable with multi-cloud? Do it. Cost optimization beats vendor lock-in most days.

Got GPUs on-prem already? Don't buy AWS. Use what developers have first.

Migration from AWS to Cheaper Alternatives

Docker containers move between providers without changes. Standard compatibility across platforms.

PyTorch and TensorFlow models don't need retraining. Framework code runs the same on RunPod, Vast.AI, AWS. Infrastructure is infrastructure.

CloudWatch monitoring doesn't port. Developers'll reconfigure observability. That's the real migration cost, not the model transfer.

If inference is the main use case? The cost savings from migration usually beat the effort. Move the workload.

AWS's Future GPU Strategy

AWS is testing new GPUs. The 4090 won't show up:that's not their lane.

Watch their announcements for L4 competitors or better inference hardware. Custom silicon (Trainium, Inferentia) is their long game. If developers're betting on AWS long-term, that's worth tracking.

Final Thoughts

No RTX 4090 on AWS. Period. Developers get L4 instead at $0.526/hour. It's a solid card for inference, but not cheap.

If developers're already in AWS and need production reliability, L4 makes sense. The SLA and integration are real benefits.

Chasing price? RunPod at $0.34/hour or Vast.AI even cheaper. The 55%+ savings matter if developers're running this long-term and inference is the main game.