WattNot

Power is now AI's bottleneck. Most inference GPUs run at stock settings all day, and every answer draws power and cooling water. Under the EU Energy Efficiency Directive (2023/1791) and Delegated Regulation (EU) 2024/1364, data centres with 500 kW or more of IT power must report energy and water KPIs to an EU database by 15 May each year. Why no one has built it: the tools are siloed. Rack monitors see heat but not models. GPU power tools cut watts but never check answers. Serving tools tune throughput but never meter energy. Nothing links a serving change to answer quality, so tuning is too risky and GPUs stay at stock. What's new: WattNot closes that loop. It meters the joules behind every request, tries one serving change at a time (power cap, FP8, KV cache, batching, reasoning effort), and a fixed gate on answer quality and p99 latency keeps or rolls back each one. Every kept saving becomes an Ed25519-signed, hash-chained receipt that anyone can verify. It runs as a test harness or as a gateway plugin in front of live traffic, and judges can replay every recorded run with no keys. EU report: one click turns a signed receipt into an HTML and CSV export framed on the 2024/1364 fields: IT energy, energy per request vs. baseline, the saving with its p-value, idle power and joules per correct answer. Every field carries a value, unit, label and source. Facility metrics a GPU tool can't see, such as PUE and renewable share, are marked NOT_AVAILABLE. It supports reporting; it is not a compliance claim. CSRD reporting is on the roadmap. Headline result: on an H100 serving Meta's open-weight Muse Glimmer 30B, live traffic went from 1,124.5 J to 979.6 J per answer (-12.9%, MEASURED), with 54 of 60 correct in both arms (Welch p = 0.027 on energy). The sealed holdout, the FP8 rollback and our Modal and Baseten runs are covered in the next two answers. Impact: saved energy becomes fleet capacity. An operator gets more answers from the GPUs and grid connection it already has, so AI can grow without the power and water bill growing as fast, and gets numbers it can put in a report. Business: wattnot would charge a gain-share, a cut of the savings the signed receipts prove. If the gate keeps nothing, the customer pays nothing.