Challenge
A growing inference fleet on a carbon-heavy grid, with customers asking for the numbers.
Solution
512 H200s on N-01 behind the same Kubernetes API, migrated region by region.
Result
61% less CO₂ per query, same latency targets, no application changes.
Inference that never sleeps
Helix AI runs protein-structure prediction as a service for pharmaceutical and academic labs. Unlike training, the load never stops: queries arrive from Boston in the morning and Singapore at night, and the fleet runs close to full all day.
By late 2025 the fleet had grown to 512 GPUs on a grid where most of the power still came from gas. Customers started asking for the carbon behind each prediction, and the answer was not one Helix wanted to publish.
“Our customers measure everything. It was time we could show them the power, too.”
Same API, different latitude
Helix already ran on Kubernetes with the GPU operator, so the move was a matter of capacity, not code. Nordlys racked 512 H200s on N-01 and exposed them through the same cluster API. Helix shifted traffic one region at a time over two weeks, with the old fleet as a fallback.
What made it easy
Bare-metal nodes with the drivers and container runtime Helix already used.
A private 100 Gb/s link from Narvik to their edge in Frankfurt.
Per-minute power and PUE data, exported straight into their reporting.
Six months on
CO₂ per query is down 61%, and median latency from Frankfurt is 41 ms, well inside the 80 ms target. Helix now publishes the energy behind every prediction on its status page, measured at the meter in Hall A.
