The Subtraction Dividend: Helix Runs Inference on AWS Trainium

Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away. - Antoine de Saint-Exupéry

For over a year we had a clever bit of software at the heart of Helix. Its one job was to decide which chip should run which AI model. It was hard to build, we were proud of it, and by the spring it had quietly become useless: our customers had turned up with their own setups and stopped needing it. But it sat there anyway, because deleting something that still works feels like vandalism.

So we deleted it. An enormous amount of it, gone in an afternoon. It was terrifying.

Here is what nobody tells you about throwing things away: it makes room. A week later a customer asked whether we could run their AI on Amazon’s own chips, a range called Trainium, instead of Nvidia’s. Supporting them means our customers are no longer hostage to a single chip supplier: they can pick whatever runs their AI best. Under the old, cluttered system that would have been a slog. On the cleaned-up one it took a few days.

We just stopped hoarding, and a whole new option fell out the other side. That is the subtraction dividend, and I suspect most teams are sitting on one without realising it.

Read the full story: https://helix.ml/blog/the-subtraction-dividend