Self-Hosted AI Agent Costs: Real Monthly Numbers

Hardware Costs and Amortization

Hardware is the largest upfront investment but amortizes over 3 to 5 years. These are representative 2026 prices for complete systems (GPU, CPU, RAM, storage, PSU, case).

Entry tier: RTX 4060 Ti 16 GB system, roughly $1,500 total. Runs 7B to 13B models. Monthly amortization over 4 years: $31.

Mid tier: RTX 4090 24 GB system, roughly $4,000 total. Runs models up to 34B parameters. Monthly amortization over 4 years: $83.

Professional tier: Dual RTX 4090 system, roughly $6,500 total. Runs 70B models with model splitting. Monthly amortization over 4 years: $135.

Enterprise tier: A100 80 GB or H100 server, $25,000 to $45,000. Runs 70B+ models at high throughput. Monthly amortization over 4 years: $520 to $937.

These hardware costs are one-time capital expenditures. Unlike cloud API costs, they do not increase with usage. Whether you generate 1 million tokens or 1 billion tokens per month, the hardware cost stays the same.

Electricity Costs

GPU-equipped systems consume meaningful electricity, especially under sustained inference load. Actual power draw depends on GPU utilization and your local electricity rate.

RTX 4060 Ti system: Approximately 200 watts at typical inference load. At the US average rate of $0.16 per kWh, running 24/7 costs about $23 per month. At European rates (around $0.30 per kWh), approximately $43 per month. In practice, most systems are not under continuous load, so real costs are often 30 to 50 percent lower.

RTX 4090 system: Approximately 350 watts at inference load. US: $40 per month continuous. EU: $76 per month continuous. With realistic utilization patterns: $20 to $50 per month.

H100 server: Approximately 700 watts per GPU at load. US: $81 per month continuous per GPU. In a 4-GPU configuration: $324 per month for electricity alone.

For organizations with solar panels, off-peak electricity rates, or data center contracts with bulk pricing, these numbers can be substantially lower. Conversely, peak-rate electricity in expensive markets (parts of California, Germany, or Japan) can push costs 50 percent higher than these averages.

VPS and Hosting Costs (if not on-premise)

If you do not host hardware on-premise, GPU-equipped VPS and cloud instances provide an alternative with monthly rental costs.

Budget GPU VPS: $40 to $80 per month for a Hetzner or OVH server with an older GPU (GTX 1080 Ti, RTX 3060). Suitable for running 7B models.

Mid-range GPU VPS: $100 to $250 per month for an RTX 4090 or A10 instance from providers like Lambda, Vast.ai, or RunPod. Suitable for models up to 34B parameters.

Enterprise GPU cloud: $2 to $4 per GPU-hour for A100 or H100 instances from Lambda, CoreWeave, or major cloud providers. At continuous usage, $1,500 to $3,000 per month per GPU. These make sense for burst workloads or organizations that want to avoid capital expenditure.

VPS hosting converts capital expense to operational expense. The per-month cost is higher than owning hardware, but there is no upfront investment, and you can scale up or down. For organizations testing self-hosting before committing to hardware purchases, this is a practical starting point.

Software and Licensing Costs

Most self-hosted AI software is open source and free.

Inference engines: Ollama, vLLM, llama.cpp, and TGI are all free and open source.

Orchestration platforms: Dify community edition, Flowise, n8n community edition, LangGraph, and CrewAI are all free. Some offer paid tiers (n8n cloud, Dify cloud) but the self-hosted versions are free.

Models: Llama, Mistral, Qwen, DeepSeek, and most other open-weight models are free to download and use, including for commercial purposes. Check individual model licenses, but the dominant open-weight models use permissive licenses.

Vector databases: pgvector (PostgreSQL extension), Qdrant, and Weaviate all have free open-source self-hosted editions.

Monitoring: Langfuse community edition is free. Grafana, Prometheus, and similar observability tools are free.

In practice, the software licensing cost of a self-hosted AI agent stack is $0 per month for the vast majority of deployments.

Engineering and Maintenance Time

This is the cost category most people either ignore or overestimate. Real-world engineering time breaks down into two phases.

Initial setup: 8 to 40 hours depending on complexity. A simple Ollama + Dify deployment takes a day for someone familiar with Docker. A production Kubernetes deployment with monitoring, backups, and high availability takes a week or more.

Ongoing maintenance: 2 to 10 hours per month for a stable deployment. This includes monitoring system health, applying security updates, evaluating new model releases, debugging occasional issues, and expanding storage or capabilities as needed. The lower end applies to simple single-server setups; the higher end applies to multi-component production deployments.

If you value engineering time at $75 per hour, ongoing maintenance costs $150 to $750 per month in labor. For organizations with existing DevOps or system administration staff, this is marginal additional work rather than a new hire.

Cloud API Cost Comparison

To determine your breakeven point, compare your projected cloud API spend to the total self-hosted cost.

Low usage (10M tokens/month): Cloud APIs cost approximately $30 to $100. Self-hosted costs (mid-tier hardware amortization + electricity) approximately $103 to $133. Cloud wins at this volume.

Medium usage (100M tokens/month): Cloud APIs cost approximately $300 to $1,000. Self-hosted costs remain approximately $103 to $133 (hardware does not change with volume). Self-hosting wins for all but the cheapest cloud models.

High usage (500M tokens/month): Cloud APIs cost approximately $1,500 to $5,000. Self-hosted costs remain approximately $103 to $133. Self-hosting saves $1,400 to $4,867 per month.

Very high usage (1B+ tokens/month): Cloud APIs cost $3,000 to $10,000+. Self-hosted costs remain approximately $103 to $133 (possibly $135 with dual GPUs for larger models). Self-hosting saves $2,900 to $9,900 per month.

The breakeven point for most configurations falls between 30 and 100 million tokens per month, typically reached within 6 to 18 months of purchasing hardware. Beyond that point, every additional token is essentially free, costing only the marginal electricity to generate it.

Per-user subscription models change the math further. If your team uses cloud AI through per-seat subscriptions ($20 to $200 per user per month), the breakeven calculation is simpler: multiply the per-seat cost by the number of users and compare to your self-hosted monthly cost. A team of ten users at $30 per seat spends $300 per month, meaning a $1,200 hardware investment pays for itself in four months even before accounting for the unlimited usage you gain.

Hidden Costs to Watch

Cooling: GPU systems generate significant heat. In a home office, this is noticeable in summer. In a server room, you need adequate cooling infrastructure. Data center colocation includes cooling in the rack fee, but on-premise setups may need additional HVAC capacity.

Noise: GPU fans under load are loud, typically 40 to 55 dBA for consumer GPUs and louder for server-grade hardware. If the system is in a shared workspace, acoustic isolation or a separate room may be necessary.

Internet bandwidth for model downloads: Downloading large models (40 to 130 GB for full-precision 70B+ models) requires adequate bandwidth. This is a one-time cost per model but can be slow on limited connections.

Backup infrastructure: Your conversation logs, vector database indices, and configuration files need backups. This might be existing backup infrastructure or a new backup target, typically $5 to $30 per month for cloud backup storage.