🌐

AI Agent Hosting Region / Latency Cost Comparator

Compare cloud GPU and inference API costs across AWS, GCP, Azure, CoreWeave, Lambda, Vast.ai, RunPod, and bare-metal regions. Factor in egress, latency, carbon, and spot/reserved pricing.

Scenario presets

Workload & region

Workload type User region Commitment / pricing model Monthly requests Input tokens / request Output tokens / request Cached input (%) LLM API model Egress / response bandwidth (GB / month) Carbon price ($ / tonne CO₂) Projection (months)

Estimates are directional. Regional prices, carbon intensities, and latency are representative. Last updated: 2026-07-10. See notes.

Cheapest region

—

Lowest monthly cost

—

Best latency region

—

Greenest region

—

Verdict

—

Region comparison

Provider / Region	Est. latency	Monthly compute	Monthly egress	Carbon cost	Total monthly	Score	Status
Run a calculation to compare regions.

Frequently asked questions

Why compare regions for AI agent hosting?

Latency, egress, carbon intensity, and spot/reserved discounts vary widely by provider and region. A region that looks cheap on compute can become expensive once egress and user latency are included.

How is end-to-end latency estimated?

The calculator adds the selected user-region base latency to the provider region's network bonus. It is a rough RTT proxy; actual latency depends on peering, CDN, and model time-to-first-token.

What does spot/reserved pricing mean?

Spot instances are interruptible and cheaper; reserved or committed-use instances trade flexibility for a discount. The calculator applies each provider's representative discount.

Is carbon cost included by default?

No. A carbon price per tonne is optional; when set, it multiplies estimated monthly kWh by the region's grid carbon intensity.

Should I pick the cheapest region?

Not always. Latency to users, compliance, support, and spot interruption risk matter. The tool ranks by a weighted score balancing cost, latency, and carbon.

Methodology notes

For LLM API workloads, compute cost = (uncached input tokens × input rate + cached input tokens × cached rate + output tokens × output rate) × requests / 1M. Rates come from the selected model.
For cloud GPU workloads, compute cost = base GPU hour × (1 + region premium) × (1 - spot or reserved discount) × GPU hours. Electricity is not included; add it separately if running bare metal.
For bare-metal colocation, compute cost = server count × base colocation GPU-hour equivalent × hours in month, plus bandwidth × egress rate and a carbon cost from power draw × grid intensity.
Egress cost = egress GB × provider/region egress rate. Some providers include a free tier; this calculator does not model free tiers.
Latency = user-region base RTT + provider region latency bonus. It is a coarse proxy; actual time-to-first-token depends on model, queue depth, and networking.
Score weights: 50% normalized cost, 35% normalized latency, 15% normalized carbon. Lower is better.