AI models are measured on accuracy, latency, and token usage. But when you run multi-step agent workflows on your own GPUs, one metric stays invisible: how much energy each step actually consumes.
Matcha is building the observability layer that connects GPU hardware telemetry with AI workload traces — giving you energy-per-inference attribution across every model, step, and team.
This playground demonstrates that visibility.
- Click ▶ RUN AGENT — a multi-step AI agent runs a stock research workflow
- Each step appears in Agent Traces with energy (mWh), tokens, latency, and carbon (gCO₂)
- GPU Metrics show real-time power draw, utilization, and temperature
- After the run, click any model name to swap it (e.g. GPT-4o → Mistral 7B)
- Click ↻ RE-RUN to see how the new model changes energy and output
- Run History compares runs side by side — see the savings
Learn more at usematcha.dev