Cloud LLM vs Local LLM Smart Home: Cost + Privacy

Cloud LLM API pricing for a typical UK household using Home Assistant's voice pipeline with 30-50 daily voice interactions plus 5-10 weekly automation tweaks via natural language:

• Anthropic Claude 4.7 (Sonnet tier): £8-15 per month. Most of the cost is conversation-history tokens (prompt caching cuts this 50-70%, so enable it).
• OpenAI GPT-4o: £6-12 per month. Slightly cheaper per token than Claude; the difference is small.
• Google Gemini 2.0 Pro: £0 on free tier (60 requests/minute, 1500 daily) for light users; £3-8 per month if you exceed free tier.
• Anthropic Claude 4.7 (Opus tier): £25-50 per month - overkill for smart-home use; reserve for routine generation only.

Setup cost: £0. Existing Home Assistant install + an API key is all you need. Latency: typically 1-3 seconds round-trip for voice interactions.

Local LLM hardware + operating cost for a working setup running Llama 3.3 8B (the practical sweet spot):

• Mac mini M4 (24GB unified memory): £999 new from Apple UK. The most accessible local LLM platform. Quiet, low-power, runs HA + Ollama + Whisper STT in one box.
• Beelink SER8 / Minisforum HX99G (32GB + iGPU): £700-900. Linux + AMD ROCm + Ollama. More setup work than Mac mini but more flexible.
• Nvidia RTX 4060 Ti 16GB (DIY build): £400-450 for the GPU + £350-500 for the rest of a host PC. Best raw inference speed but loudest and highest-power.
• Older / used hardware: RTX 3060 12GB at £200-280 used. Works for Llama 3.3 8B with tighter quantisation. Cheapest entry but expect tuning work.

Electricity: ~30W idle, 200-300W under inference. At UK 2026 electricity rates (~28p/kWh) and ~3-4 hours/day active inference: £4-12 per month depending on platform. Mac mini sits at the low end; DIY desktop with discrete GPU at the high end.

Straight cost break-even using a Mac mini M4 (£999) vs. Claude 4.7 cloud (£12/month average) and counting the local electricity (£8/month average) as a delta of £4/month vs zero cloud cost:

• Cloud total over 24 months: £288.
• Local total over 24 months: £999 + £192 electricity = £1,191.
• Break-even point on cost alone: Never within 5 years. Cloud is materially cheaper if you only count money.

Adjusted break-even ignoring the hardware cost (treating the Mac mini as an asset that does many other things, not just AI):

• Local marginal cost over 24 months: £192 (electricity).
• Cloud cost over 24 months: £288.
• Break-even point: About 18 months at £12/month cloud spend.

If you'd buy the Mac mini anyway for other workloads (Plex, Time Machine, Frigate, Home Assistant), local LLM becomes the cheaper option from month 18-24 onwards. If you'd buy it solely for the LLM, cloud is cheaper essentially forever at typical usage rates.

The capability gap matters, and it's bigger than the cost gap for most users. Frontier cloud models (Claude 4.7, GPT-4o, Gemini 2.0 Pro) are roughly 18-24 months ahead of the best openly-released local models on most benchmarks. In smart-home context:

For most households considering the switch from cloud to local LLM, privacy is the binding constraint, not cost. The honest picture in 2026:

• OpenAI API: API traffic is excluded from training data by default. 30-day retention for abuse monitoring. SOC 2 audited; trustworthy by mainstream commercial standards.
• Anthropic Claude API: Similar terms - no training on API data, retained for abuse review. Similar trust posture.
• Google Gemini: Excludes API traffic from training (paid tier). Free tier has slightly looser terms - check the latest policy.
• Local Ollama: Nothing leaves your network. All conversation data stays on the host. Voice transcripts (if using Whisper locally) also stay local.

For most households, cloud terms are acceptable - you're not transmitting medical data or commercially sensitive material through your smart-home conversation agent, and the providers have material brand incentive to honour the privacy policies. For households with strict privacy requirements (medical professionals, security-sensitive roles, families managing children's data), local is the only fully-on-prem route.

Yes, and it's the sensible advanced setup. Home Assistant supports multiple Conversation integrations simultaneously. The practical pattern:

• Local Ollama as primary for device control + simple voice commands (sub-1s latency, full privacy).
• Cloud LLM as fallback for complex automation drafting + multi-step troubleshooting (the cloud model's capability ceiling).

Route via the Assist pipeline's agent priority order. Most local-first households route 80-90% of interactions to Ollama and only 10-20% (the harder ones) to cloud. This typically cuts cloud spend to £1-3/month and gives you the best of both routes.

For most UK households starting out with AI in the smart home in 2026, cloud LLM (Anthropic Claude 4.7 or OpenAI GPT-4o) is the right choice - £5-15/month for materially better capability than any £1000 local setup delivers today. Use the saved money to build out the rest of your smart-home stack rather than committing to a homelab. The cloud route also gives you cleaner upgrade paths as model capability improves through 2027.

For households with binding privacy requirements, or those who already own / want a Mac mini for other workloads, local LLM via Ollama is excellent and getting better fast. Llama 3.3 8B is the current sweet spot; Qwen 3.5 is the alternative if you want slightly stronger reasoning. Expect parity with cloud frontier in 18-24 months; until then accept the capability gap as the price of privacy.

The honest framing: cloud wins on capability and convenience; local wins on privacy and offline operation. Cost is a wash for any serious user. Pick by which you value more.