The following is a personal account of events that took place between 2023 and the present day. Names have not been changed. The machines were all consenting adults.
She had fifteen hours in her when she left the factory.
That's what they told us. Fifteen hours. Apple put it on the spec sheet, right there next to the chip count and the port situation. Fifteen hours of all-day battery life. She could handle anything.
Nobody mentioned what would happen when the AI showed up.
The First Time
It started innocently enough. A Copilot subscription. A VS Code extension. Just a little help with the autocomplete, nothing serious. She barely noticed.
But the model was always on. That's the thing nobody tells you about that first experience — it never stops thinking about you. While you're reading documentation, while you're staring at the ceiling, while you're getting coffee. The context scanner is running. The CPU is warm. The suggestions are forming.
She went from fifteen hours to nine. Nobody talked about it. It felt rude to bring up.
# The morning after — checking what actually happened
sudo powermetrics --samplers cpu_power,gpu_power -i 1000 -n 5
# You'll see numbers you weren't ready for
# The chip was busy all night
# You just didn't know with what
The Players
Like all golden age stories, this one has a cast.
The Always-On Model — Never tires. Never sleeps. Scanning your open files, your recent edits, your local documentation, building and rebuilding its sense of what you might need next. Every pause in your typing is an invitation. Every second of silence is filled with inference.
The GPU — Called in when the real work starts. Doesn't do anything halfway. Fires up fully for every suggestion, every completion, every ghost-text rendering. Has no concept of a quick favor.
The Thermal Management System — The one who cleans up afterward. Spins the fans when things get too hot. Draws its own power to manage someone else's heat. Never complains. Never stops.
The Wi-Fi Radio — Often overlooked. Underestimated. When you're running cloud AI, she's the one maintaining the continuous high-power stream to the API servers. Never gets to rest. Never reaches the low-power doze state she deserves.
The Battery — Everybody wanted something from her.
The GPU Tax
Here is what they don't put in the press releases.
Every time the suggestion appears — that little ghost-text flicker at the end of your cursor — the GPU fired. Every single time. The Neural Engine on the M-series chips handles it efficiently, yes. But efficiently is not freely. There is always a cost.
The chip heats up. The chassis warms. The thermal management system, faithful as ever, spins the fans. And here is the part that made the engineers wince when they ran the numbers: the fans are mechanical. They draw from the battery to move air, while the GPU draws from the battery to generate the heat that required the air. Each one's response to the other's behavior increases the total load.
The industry called it thermal runaway. The machines just called it Tuesday.
"I knew something was different," one M3 MacBook Pro reportedly told no one, "when I felt the fans at 9 AM. I hadn't done anything yet. But something had."
Cloud vs. Local: The Arrangement Nobody Admits To
People assumed cloud AI would be gentler. Let the servers do the work. You just handle the text. How expensive can text be?
They forgot about the radio.
The Wi-Fi chip has moods. In its natural state — idle, occasional packets, background sync — it dozes between transmissions. Low-power state. Civilized. But maintain a continuous stream to an API server, the kind that a cloud AI assistant requires, and the radio never rests. It stays fully awake, fully powered, fully committed to the connection.
It's not catastrophic. It's constant. And constant is its own kind of drain.
