Cloud vs Local: Calculadora ROI Precios reales de API (abril 2026). Elija modelo cloud y hardware local para comparar.
Escenario PYME 5K consultas/mes, Mac Mini M4, 10h ahorradas/semana Enterprise 25K consultas/mes, M4 Pro, 30h ahorradas/semana Gobierno 10K consultas/mes, Sovereign Node, privacidad máxima
Modelo Cloud (API) GPT-5 — $1.25/$10 per 1M tok (OpenAI) GPT-5 Mini — $0.25/$2 per 1M tok (OpenAI) Claude Sonnet 5 — $2/$10 per 1M tok (Anthropic) Claude Haiku 4.5 — $1/$5 per 1M tok (Anthropic) Gemini 1.5 Pro — $1.25/$5 per 1M tok (Google) Gemini 1.5 Flash — $0.075/$0.3 per 1M tok (Google) Gemini 2.0 Flash — $0.1/$0.4 per 1M tok (Google) Mistral Large — $2/$6 per 1M tok (Mistral) Mistral Small — $0.1/$0.3 per 1M tok (Mistral) Llama 3.1 70B — $0.88/$0.88 per 1M tok (Meta/Together) Llama 3.1 8B — $0.18/$0.18 per 1M tok (Meta/Together) DeepSeek V3 — $0.27/$1.1 per 1M tok (DeepSeek)
Hardware Local Jetson Orin Nano Super — EUR 230 (8GB, 8B max, ~15 tok/s)
RPi 5 + AI HAT+ 2 — EUR 200 (16GB, 4B max, ~8 tok/s)
Beelink SER8 — EUR 550 (32GB, 14B max, ~11 tok/s)
ASUS NUC 14 Pro — EUR 650 (64GB, 8B max, ~10 tok/s)
Minisforum UM890 Pro — EUR 750 (96GB, 14B max, ~12 tok/s)
Mac Mini M4 (24GB) — EUR 920 (24GB, 14B max, ~20 tok/s)
Mac Mini M4 Pro (48GB) — EUR 1840 (48GB, 70B max, ~25 tok/s)
Mac Studio M4 Max (48GB) — EUR 2300 (48GB, 70B max, ~18 tok/s)
Jetson AGX Orin 64GB — EUR 1840 (64GB, 20B max, ~18 tok/s)
RTX 4060 Ti 16GB Build — EUR 1000 (16GB, 14B max, ~55 tok/s)
RTX 4090 24GB Build — EUR 2500 (24GB, 30B max, ~130 tok/s)
RTX 3090 24GB (Used) Build — EUR 1200 (24GB, 30B max, ~80 tok/s)
AMD W7900 48GB Build — EUR 4000 (48GB, 70B max, ~50 tok/s)
Jetson AGX Thor (128GB) — EUR 3499 (128GB, 70B FP16 max, ~80 tok/s)
Jetson T4000 (64GB) — EUR 1999 (64GB, 32B FP16 max, ~60 tok/s)
Cloud A100 80GB (RunPod) — EUR from 1.29 (80GB, 70B FP16 max, ~120 tok/s)
Cloud H100 80GB (RunPod) — EUR from 2.49 (80GB, 180B FP16 max, ~200 tok/s)
MacBook Pro M4 Pro (24GB) — EUR 2199 (24GB, 14B Q4 max, ~22 tok/s)
Lenovo ThinkPad P16s Gen 3 (RTX A500) — EUR 1650 (32GB, 7B Q8 max, ~12 tok/s)
Framework 16 (Radeon RX 7700S) — EUR 1900 (32GB, 14B Q4 max, ~15 tok/s)
Supermicro 1U Server (L40S 48GB) — EUR 12000 (48GB, 70B Q4 max, ~65 tok/s)
Dell PowerEdge R760xa (2x RTX A6000) — EUR 18500 (96GB, 120B Q4 max, ~100 tok/s)
Fuentes y supuestos del cálculo Precios API: AgentOps/tokencost (GitHub), verificados abril 2026
Ratio input/output: 60% input, 40% output (típico para chat/RAG)
Electricidad: configurable (defecto EUR 0.25/kWh, media España)
Hardware: precio de compra + electricidad mensual. Sin soporte ni mantenimiento.
Conversión: $1 = EUR 0.92
Metodología: TCO (Total Cost of Ownership) comparando coste cloud (API tokens) vs coste local (hardware + electricidad + soporte opcional).