Building a personal AI agent can run on two fronts: a local Mac Mini with Ollama and open-source LLMs or a cloud server tapping into Claude or Gemini APIs. A Mac Mini costs about $1,200 upfront, draws roughly 20 W around the clock (≈15 kWh/month), so at ₹8/kWh you’re spending around ₹120 ($1.50) monthly on electricity. Spread the hardware cost over three years, and you’re at roughly $33 per month. Add the power bill, and you’re paying about $35/month for unlimited local inference. On the cloud side, a Hetzner VPS runs you about $45/month. Then come API fees: if your agent processes 10 million tokens a month, Claude will set you back roughly $78 for 6 M input and 4 M output tokens. Gemini Pro is cheaper—around $28 for the same workload. Total monthly cost ends up near $123 with Claude or $73 with Gemini. Over a year, that gap widens: local stays at about $420, Gemini clouds cost $876, and Claude clouds around $1,476. Beyond dollars, you trade quality for independence. Local models are solid at summarization, document search and handling private data. They struggle more with coding agents, complex reasoning or long-context tasks where Claude and Gemini lead. If you invoke your agent occasionally, cloud fits best—you only pay when you run. But for daily or nonstop operation, the Mac Mini path cuts expenses in half against Gemini and quarters against Claude.