Home › Case studies › Case 07
Case 07 · No AI. No GPU. No data center. No power draw.
Detecting AI text should not require a data center.
Every competitor runs on GPUs or hosted APIs. We run on any CPU.
The failure · on the record
Binoculars requires two Falcon-7B instances (~14 GB GPU RAM per query). FastDetectGPT needs GPT-Neo. DeBERTa-ConPara needs an 800 MB transformer. GPTZero, Grammarly, Trinka bill per token or per document over an API. Ours: a CPU calibration against a public human-writing reference corpus. Zero GPU-seconds. Zero data-center water.
The numbers
| Detector | Hardware per query | Cost / 1M docs | Power / 1M docs |
|---|---|---|---|
| Binoculars | 14 GB GPU RAM | ~$4,000 (cloud GPU-hour) | ~180 kWh |
| GPTZero API | Hosted | ~$5,000–$10,000 | opaque (theirs) |
| Grammarly Enterprise | Hosted | quote-only | opaque |
| Turnitin | Hosted | quote-only | opaque |
| Truth-in-AI | any CPU | $0 marginal (flat license) | ~0.2 kWh (CPU) |
What we do differently
A university grading 500,000 essays a semester pays flat license, not per-query. No GPU quota. No API bill. No carbon report.
The environmental pitch is also the finance pitch.
Proof
~0.2 kWh
Marginal power draw per 1 million detections · competitors:
Binoculars: ~180 kWh · GPT-based detectors: opaqueTry it yourself
Time it on your own hardware.
time truth-in-ai-detect --file input.txt --iterations 10000
Bounty · put money where the claim is
Bring your own procurement number. We’ll cost out a 5-year TCO comparison, signed.
Sources