The "Dead Battery on Monday Morning" Mystery
You shut your MacBook Pro lid on Friday evening with 88% battery remaining. When you open your laptop on Monday morning at a coffee shop, you're greeted by a flashing red battery icon at 3%. You check macOS battery settings, and the culprit isn't an open Chrome tab or Slack. It is Ollama or LM Studio running silently in the background.
Why does an idle AI model that is generating zero tokens drain your laptop battery? The answer lies in how modern System-on-Chips (SoCs) manage deep C-state sleep and memory bus power gating.
The Hardware Reality of Idle Unified Memory
To retain model weights in Unified Memory or GPU VRAM, the memory controller must continuously supply refresh power to millions of LPDDR5X capacitor cells. Furthermore, because runner processes like ollama_llama_server keep active Metal command queues open, macOS is prevented from clocking down the memory bus or putting CPU cores into deep sleep (C-states C8 through C10).
┌─────────────────────────────────────────────────────────────────┐
│ SOC POWER RESIDENCY: IDLE LOADED MODEL vs. PROACTIVE SLEEP │
└─────────────────────────────────────────────────────────────────┘
IDLE MODEL LOADED (Memory Active):
[ CPU Cores: C1 Active (5.2W) ] ── [ Memory Bus: Pinned Clock (18.4W) ] ──> ~24W Drain
Battery Life: ~3.8 Hours | Chasis Warm | Fans Running Low
AUTOMATED IDLE EVICTED (ContextWarden Sleep Guard):
[ CPU Cores: Deep C-State (0.15W) ] ── [ Memory Bus: Power-Gated (0.8W) ] ──> ~1.2W Drain
Battery Life: ~18.5 Hours | Chasis Cool | Absolute Silence
Measuring the Power Draw: Real Test Data
Using a calibrated hardware power meter on an Apple MacBook Pro 16" (M3 Max, 48GB Unified Memory), we measured total system power draw across four operating states:
| System State | Power Draw (Watts) | Estimated Battery Life | Chassis Temperature |
|---|---|---|---|
| Clean macOS Idle (No LLM) | 2.4 Watts | 22.5 Hours | 28.5°C (Cold) |
| Ollama Idle (14B Model Loaded) | 21.8 Watts | 4.2 Hours | 39.2°C (Warm) |
| Active Token Generation (Streaming) | 48.5 Watts | 1.9 Hours | 84.0°C (Hot) |
| ContextWarden Battery Guard Active | 2.8 Watts | 20.5 Hours | 29.0°C (Cold) |
Configuring Native Keep-Alive Timeouts
To prevent Ollama from hoarding memory indefinitely, configure an aggressive keep-alive timeout. By default, Ollama holds models for 5 minutes, but many custom scripts set this to -1 (never unload):
# Set Ollama default idle unload timeout to 3 minutes
$ launchctl setenv OLLAMA_KEEP_ALIVE "3m"
# In API requests, explicitly request automatic unloading when finished:
curl http://localhost:11434/api/generate -d '{
"model": "qwen2.5-coder:14b",
"prompt": "Refactor this function",
"keep_alive": "2m"
}'The Dilemma: Cold-Start Latency vs. Battery Life
The problem with standard keep-alive eviction is cold-start latency. When your model unloads to disk, your next autocomplete prompt in VS Code or Cursor will stall for 10 to 18 seconds while 12GB of weights are read from NVMe back into Metal memory.
This is where intelligent lifecycle management makes all the difference:
- AC Power Mode: Keep models warm in memory while connected to a MagSafe charger to deliver zero-latency code completions.
- Battery Mode: Proactively freeze and evict idle models after 90 seconds of inactivity to instantly restore your 18-hour battery endurance.
- Lid Close / Sleep Trigger: Automatically flush VRAM the second your display sleeps, preventing battery depletion in your backpack.