Compute Ledger · Measured

Everything I've run through the machines.

A running tally of every token I've burned — across Claude, ChatGPT, Claude Code, and my own GPUs — and what it costs to keep the lights on.

5,439,279,945

tokens processed · all providers, all time (incl. cache reads)

44,295,025 generated · measured since Apr 2026

Private AI · Home lab Connecting

The Spark behind the stories.

Waiting for model telemetry

227,290

input tokens today

18,804

output tokens today

requests today

—

live output tokens / sec

18,804

recorded output tokens

Inside the inference engine

Connecting

Apps. Benchmarks. Late-night experiments. One measured counter.

—

input tokens / sec

—

output tokens / sec

—

active requests

—

seconds of saved activity

OUTPUT TOKENS / SEC · RECENT ACTIVITYWaiting for samples

Waiting for engine telemetry.

GLM engine counts include apps, direct API calls, benchmarks and Showcase tests. Input includes cache reuse reported by the engine. Earlier saved gateway history is retained once. Old direct calls without saved engine records cannot be recovered. Counts survive collector restarts. A model crash between samples can leave a small gap.

Spark 1 · Head Waiting

—

GPU use

—

GPU temperature

—

GPU-reported power

—

Unified memory used / total

Spark 2 · Worker Waiting

—

GPU use

—

GPU temperature

—

GPU-reported power

—

Unified memory used / total

The active model is shown above. Counts retain historical gateway usage and include all GLM engine traffic from October 8, 2026 at 3:26 p.m. ET. Gateway requests are not added again. Today resets at midnight in New York.

Each hardware card reports its own Spark. Unified memory includes the operating system and other processes. GPU-reported power is not wall power. Unavailable or stale node readings are hidden. Live rates measure token-counter changes between samples. They include idle time and are not benchmark scores.

Waiting for the latest reading.

The purchase justification department TOKEN MATH

Was the hardware worth it?

Every local token nudges the Spark toward its financially responsible era.

—

API equivalent today

—

Reference API equivalent since setup

—

remaining against the $4,500 target

Waiting for token counts$4,500 hardware target

Financially responsible era: still buffering.

Recorded input + output tokens × the fixed GLM-5.3-Flash reference rates below. This estimates a hosted API equivalent, not an actual bill or net savings. Counts combine older Qwen gateway history and GLM engine usage without overlap. Electricity is not deducted.

OpenRouter pricing: $0.0232 input / $1.2 output per million tokens · checked 2026-10-12. Current GLM rates apply to all recorded counts, including older Qwen usage. Provider rates can vary.

On-device · running here ● IDLE

Muse Glimmer 30B · 4-bit · Apache 2.0

3.3K

tokens generated

tokens / sec

30B

on Apple M5

private

nothing leaves the box

Meta’s open model, running privately on my own machine — no API key, nothing leaving the device.

Processed · by source

Claude Code5.38B

Anthropic API12.4Mest

ChatGPT9.1Mest

OpenAI·Codex4.6Mest

Field Signal Desk9.7M

Spark (saved history + live GLM)28.1M

Spark usage combines saved gateway history with live GLM engine counts; Claude Code is measured from local logs. Historical API and chat entries marked “est” are estimates. Cache reads count toward processed tokens.

Cloud engine · historical ● IDLE

Nemotron-49B · Lambda 8×B200

—

GPU util

—

tokens/sec

—

VRAM

—

temp

Engine off — last run never (provisioned, not yet run). Gauges light up green during a session.

Lambda credits$7,500 · $0 burned

Spark output · recent days

Recorded daily output from saved gateway history and the GLM engine.

This habit isn't free.

If the ledger's fun to watch, buy me a coffee — it goes straight back into tokens.

☕ Buy me a coffee