Open-source tool energygraph shipped version 1.3 today—a zero-dependency utility that, for the first time, lets you see the true electricity cost of local AI inference right in your terminal. It logs every piece of hardware's actual power draw, broken down by core, by device, and by memory module.

What this is

energygraph is a local hardware power-monitoring tool. The biggest change in version 1.3 is the addition of support for dedicated GPUs (dGPUs—discrete graphics cards, as opposed to integrated graphics) from NVIDIA, Intel, and AMD. Combined with its existing coverage of CPU cores, integrated graphics, and memory modules, the tool now gives you a near-complete view of your machine's power consumption. It renders dynamic charts inside the terminal, weighs just a few megabytes, and depends on no external libraries.

Its target user is unmistakable: people running open-source large models (Llama, Mistral, Qwen, etc.) at home on consumer-grade GPUs. The developers call it "the electricity meter for local AI inference."

Industry view

On Reddit's LocalLLaMA community (a gathering spot for local LLM enthusiasts), developers are enthusiastically comparing peak wattage readings while their models run. A seemingly niche tool suddenly going viral is itself a signal—we're starting to treat AI inference as "something the meter will bill you for."

There are calmer voices, too. Some argue that running a small 7B (7-billion-parameter) model pulls only tens of watts—nothing compared to a household air conditioner—and that the electricity anxiety is overblown. But for those simultaneously running multiple 4090s on a 70B model, or for enterprises weighing on-prem data center deployment, power cost is no longer a line item you can ignore. One developer put it bluntly: "You think you saved on cloud fees, but you just moved the bill from AWS to the electric company."

Impact on regular people

  • For enterprise IT: On-premise AI may look like it slashes cloud costs, but data center power, cooling, and UPS (uninterruptible power supply) investments must be folded back into total cost of ownership (TCO—the full spend from procurement and ops to electricity). Vendor selection can no longer be driven by hardware sticker price alone.
  • For individual professionals: If your work involves running AI models locally, use this tool to log actual power consumption and build an economic case for your boss. It's far more persuasive than abstract talk about "30% productivity gains"—and a more reliable path to winning hardware budget.
  • For the consumer market: Consumer GPUs (the 4060, 4090, and Arc lines) are now heavily used for AI inference, and electricity will become a recurring cost of long-term ownership. Buying a GPU is no longer just a one-time hardware spend—it's more like buying a car that burns fuel.