Podrobný průvodce již brzy
Pracujeme na komplexním vzdělávacím průvodci pro Model Hosting Cost Calculator. Brzy se vraťte pro podrobné vysvětlení, vzorce, příklady z praxe a odborné tipy.
The Self-Hosted Model Cost Calculator estimates the total expense of running open-source language models like Llama 3, Mistral, Mixtral, and Phi-3 on your own cloud GPU infrastructure, and compares this against equivalent API costs to determine the break-even point. Self-hosting gives you full control over data privacy, customization, and per-request costs, but requires GPU infrastructure management, model serving frameworks, and operational expertise. As of 2025, running Llama 3 8B on a single A10G GPU ($1.10/hr) can serve approximately 20 to 50 requests per second, costing roughly $0.0006 per request. The equivalent API call on GPT-4o-mini costs $0.0003 to $0.001 per request depending on token counts. Running Llama 3 70B on 2 A100 GPUs ($5/hr total) serves 5 to 15 requests per second at approximately $0.001 per request, comparable to GPT-4o pricing but with full data control and no per-token charges. This calculator is essential for ML engineering teams evaluating build-versus-buy decisions, companies with data residency requirements that preclude sending data to third-party APIs, and organizations with high enough volume to justify the fixed infrastructure investment. The break-even typically occurs at $2,000 to $5,000 per month in equivalent API spend, below which API pricing is more economical, and above which self-hosting delivers significant savings.
Monthly Self-Hosting Cost = GPU Instances x Hourly Rate x Hours per Month + Storage + Bandwidth + Ops Overhead. Break-even API Spend = Monthly Hosting Cost / (1 - Utilization Overhead). For example: 2 A100 GPUs at $2.50/hr running 24/7: Monthly = 2 x $2.50 x 730 = $3,650. If equivalent API usage would cost $8,000/month, self-hosting saves $4,350/month (54%).
- 1Select the open-source model you want to host based on your performance requirements. Llama 3 8B is suitable for most general tasks and runs on a single 24GB GPU. Llama 3 70B requires 2 to 4 A100 80GB GPUs for full-precision inference or a single A100 with quantization. Mistral 7B offers excellent performance-to-size ratio. Mixtral 8x7B provides expert-mixture capabilities on 2 A100s. Choose the smallest model that meets your quality requirements to minimize GPU costs.
- 2Determine the GPU type and count needed. Model size in billions of parameters roughly equals the GB of GPU memory needed for FP16 inference: 7B needs approximately 14GB (fits on A10G 24GB), 13B needs approximately 26GB (needs A100 40GB), 70B needs approximately 140GB (needs 2 A100 80GB). Quantization (4-bit or 8-bit) reduces memory by 50 to 75 percent, allowing larger models on fewer GPUs at a 5 to 15 percent quality trade-off.
- 3Estimate your throughput requirements in requests per second. A single A10G serving Llama 3 8B with vLLM achieves 20 to 50 requests per second depending on input/output lengths. An A100 serving the same model achieves 40 to 100 requests per second. H100 achieves 80 to 200 requests per second. Your peak traffic determines the minimum GPU count, while average traffic determines cost efficiency. Auto-scaling between minimum and maximum GPU counts optimizes cost.
- 4Calculate the monthly GPU compute cost. On-demand pricing: A10G at $1.10/hr, A100 80GB at $2.50 to $3.00/hr, H100 at $3.50 to $8.00/hr. Reserved instances save 30 to 60 percent for 1 to 3 year commitments. For 24/7 operation, multiply hourly rate by 730 hours per month. For auto-scaling workloads, estimate average hours per month based on traffic patterns.
- 5Add infrastructure overhead costs. Storage for model weights (Llama 3 70B is approximately 140GB), inference logs, and cached data costs $10 to $50 per month. Bandwidth for serving responses adds $5 to $20 per month for moderate traffic. A model serving framework (vLLM, TGI, or TensorRT-LLM) is free but requires setup and maintenance. Load balancers and monitoring add $20 to $50 per month.
- 6Factor in operational labor costs. Setting up and maintaining self-hosted model infrastructure requires ML engineering expertise. Initial setup takes 20 to 40 hours. Ongoing maintenance, monitoring, and troubleshooting requires 5 to 15 hours per month. At $100 to $200 per hour for ML engineering time, this adds $500 to $3,000 per month in labor costs that are often excluded from cost comparisons but are real and significant.
- 7Calculate the break-even point against API pricing. Compare your total monthly self-hosting cost (GPU + storage + bandwidth + labor) against the equivalent API cost for the same workload. The break-even API spend is typically $2,000 to $5,000 per month. Below this, API pricing is more economical. Above this, self-hosting savings grow linearly with volume because GPU costs are largely fixed while API costs scale with every request.
At low volume, the fixed GPU cost of $803/mo plus $500 in ops labor makes self-hosting more expensive than GPT-4o-mini API pricing. The break-even point is approximately 12 million requests per month, where the API would cost $1,320/mo while self-hosting remains at $1,318.
At enterprise scale with 100 million monthly requests, self-hosted Llama 3 70B on reserved H100s saves 70 percent versus GPT-4o API pricing. The quality difference is measurable but acceptable for many use cases. The reserved pricing commitment reduces GPU costs from $11,680 to $5,840 per month.
Auto-scaling with spot instances reduces GPU cost by 50 percent compared to always-on at full capacity. Average utilization of 1.8 GPUs versus max 4 GPUs saves $1,601/mo. Spot pricing at $0.45/hr versus on-demand $1.10/hr saves an additional 60 percent on the GPU component.
Healthcare companies self-host Llama 3 to comply with HIPAA regulations that restrict sending patient data to third-party APIs. A hospital system runs Llama 3 70B on 4 A100 GPUs in their private cloud at approximately $8,000 per month. Processing 2 million clinical queries per month would cost $15,000 to $25,000 on commercial APIs. The 50 to 70 percent cost savings, combined with guaranteed data privacy, makes self-hosting the clear choice for healthcare AI applications.
Financial services firms self-host models for regulatory compliance, ensuring that proprietary trading strategies and client data never leave their infrastructure. A hedge fund running Mistral 7B on 2 A10G GPUs at $1,606 per month processes 500,000 daily queries for market analysis. The equivalent API cost would be $3,000 to $5,000 per month, but the real value is in data control: no risk of training data leakage through third-party API providers.
AI SaaS companies self-host models to control their COGS and maintain margins at scale. A content generation platform serving 100,000 users self-hosts Llama 3 70B on reserved H100s at $6,000 per month, handling 50 million monthly requests. The equivalent GPT-4o-mini API cost would be approximately $12,000 per month. Self-hosting saves $6,000 per month while providing unlimited request capacity within their GPU allocation.
Government agencies self-host models in air-gapped environments for classified or sensitive workloads. A defense contractor runs Llama 3 on on-premises H100 servers in a SCIF (Sensitive Compartmented Information Facility). The amortized hardware cost is $4,000 per month. No commercial API can serve classified environments, making self-hosting the only option for government AI applications with security clearance requirements.
For multi-model serving where you need to run several different models
For multi-model serving where you need to run several different models simultaneously (for example, a small model for classification and a large model for generation), GPU memory must be partitioned across models. A single A100 80GB can serve both a 7B model (14GB) and a 13B model (26GB) simultaneously with room for KV cache. This multi-model serving on shared infrastructure is impossible with API providers and can reduce total GPU costs by 30 to 50 percent compared to dedicated GPUs per model.
For burst-traffic applications where peak load is 10 to 50 times average load,
For burst-traffic applications where peak load is 10 to 50 times average load, self-hosting faces a capacity planning dilemma. Provisioning for peak means 90 percent idle capacity during normal times. Provisioning for average means dropped requests during peaks. A hybrid approach using self-hosted GPUs for baseline traffic and API calls for overflow (known as cloud bursting) provides cost-effective coverage. This requires application logic to route requests to self-hosted or API based on current load.
When deploying models in edge locations (factory floors, retail stores, medical
When deploying models in edge locations (factory floors, retail stores, medical facilities) with limited internet connectivity, self-hosting on local hardware is the only option. NVIDIA Jetson devices ($500 to $2,000) can run 7B models with quantization for on-device inference with no ongoing API costs. The amortized hardware cost over 3 years is $14 to $56 per month, far cheaper than any API option. Edge deployment also eliminates network latency and provides offline capability.
| Model | GPU Setup | Monthly GPU Cost | Throughput | Equivalent API Cost at Same Throughput |
|---|---|---|---|---|
| Llama 3 8B | 1x A10G | $803/mo | 30-50 req/s | GPT-4o-mini: ~$400-1,200 |
| Llama 3 8B | 1x A100 | $1,825/mo | 50-100 req/s | GPT-4o-mini: ~$800-2,400 |
| Mistral 7B | 1x A10G | $803/mo | 25-45 req/s | GPT-4o-mini: ~$350-1,000 |
| Llama 3 70B | 2x A100 | $3,650/mo | 10-25 req/s | GPT-4o: ~$5,000-15,000 |
| Llama 3 70B | 4x H100 | $7,300/mo | 25-60 req/s | GPT-4o: ~$12,000-35,000 |
| Mixtral 8x7B | 2x A100 | $3,650/mo | 15-35 req/s | GPT-4o-mini: ~$2,000-6,000 |
When does self-hosting become cheaper than API calls?
The break-even point is typically $2,000 to $5,000 per month in equivalent API spend when including operational labor costs, or $800 to $1,500 when considering only infrastructure costs. The exact point depends on GPU utilization rate, model size, and the commercial API you are comparing against. At $10,000+ per month in API costs, self-hosting almost always saves 40 to 70 percent.
Which serving framework should I use?
vLLM is the most popular open-source serving framework, offering 2 to 4 times higher throughput than naive HuggingFace inference through PagedAttention and continuous batching. Text Generation Inference (TGI) by HuggingFace is simpler to set up but slightly slower. TensorRT-LLM by NVIDIA offers the highest throughput on NVIDIA GPUs but requires more setup complexity. For most teams, vLLM provides the best balance of performance and ease of use.
Can I match GPT-4o quality with open-source models?
Llama 3 70B and Mixtral 8x22B approach GPT-4o quality on many benchmarks but do not fully match it on complex reasoning, creative writing, and nuanced instruction following. For tasks like classification, extraction, summarization, and straightforward Q&A, open-source models perform within 5 to 10 percent of GPT-4o. For the most demanding tasks, the quality gap may be 10 to 20 percent. Fine-tuning on domain-specific data can close or eliminate the gap for specialized use cases.
How much VRAM do I need for different models?
In FP16 precision: 7B model needs 14GB (A10G 24GB), 13B needs 26GB (A100 40GB), 34B needs 68GB (A100 80GB), 70B needs 140GB (2x A100 80GB). With 4-bit quantization (GPTQ/AWQ): 7B needs 4GB (any modern GPU), 13B needs 8GB, 34B needs 20GB (A10G), 70B needs 40GB (A100 40GB). Quantization reduces memory by 75 percent with only 5 to 15 percent quality loss, enabling larger models on fewer GPUs.
Should I use spot instances for model serving?
Spot instances save 50 to 80 percent but can be interrupted. For development, testing, and batch processing, spot instances are excellent. For production serving, use spot instances as additional capacity on top of a baseline of on-demand or reserved instances. Auto-scaling that adds spot GPU instances during peak traffic and scales down during off-peak provides a good balance of cost savings and reliability.
What are the hidden costs of self-hosting?
Beyond GPU compute: ML engineering setup time (20 to 40 hours at $150/hr = $3,000 to $6,000 one-time), monthly operations (5 to 15 hours at $150/hr = $750 to $2,250), monitoring and alerting infrastructure ($20 to $100/mo), model weight storage ($10 to $50/mo), load balancing ($20 to $50/mo), and the opportunity cost of engineering time spent on infrastructure instead of product features. Total hidden costs typically add $1,000 to $3,000 per month on top of raw GPU costs.
Pro Tip
Start with API-based models for development and early production, then migrate to self-hosting once you have stable traffic patterns and a clear cost incentive. Premature self-hosting often leads to over-provisioned infrastructure sitting idle while you figure out product-market fit. Once your monthly API bill consistently exceeds $3,000 to $5,000 and your traffic patterns are predictable, begin a parallel self-hosted deployment and gradually shift traffic from API to self-hosted based on quality validation.
Did you know?
Meta released Llama 3 70B as an open-source model that matches or exceeds the performance of GPT-3.5-turbo on most benchmarks. Running it on two A100 GPUs costs approximately $5 per day, which means any individual or organization can now operate a model as capable as what was state-of-the-art commercial AI just 18 months ago for less than the price of a coffee per day.