What does one AI job really cost?
Flowlink estimates the energy, carbon emissions, water use and dollar cost of each AI job run in GPU data centers, from published industry factors and standard methods. It's built for colocation operators and companies running AI on their own hardware. I founded it, I lead it, and I built all of its software.
Founder, CEO & Sole Software Developer
Checking my own numbers
In a summer 2026 audit I traced the water-use default I had been using to a power-plant figure that gets cited as a data-center average. It was roughly four times too high. I'm replacing every factor with a sourced value (PUE from the Uptime Institute, water use from Lawrence Berkeley National Laboratory, grid carbon from EPA eGRID) and logging each correction.
Flowlink's figures are estimates, not meter readings. The next step is a test against metered energy with a university research-computing center, published whatever it shows.
It has no customers yet. The audit also turned into research: two papers, a technical white paper and a trade-press column.
How it was built and pitched
- July 2025. Started Flowlink to answer one question: what does a single AI job cost in energy, carbon, water and money?
- 2025–26. Built the seven modules myself, then pitched the company with my co-founder: Diamond Challenge semifinal (1 of 20 teams invited to the in-person round), Blue Ocean global top 100, Coriell Science Fair and the Delaware Valley Science Fairs.
- Summer 2026. Audited every factor and found the water error; rewrote the method as a 30-page technical white paper.
- Fall 2026. Cold-emailed hundreds of researchers; the most useful replies came from asking one specific question about their own work, and that feedback shaped the paper and led to new research. An economist at Resources for the Future reviewed the white paper in October, and a UNC-Chapel Hill lab is now reviewing its methods.