HN Explorer

Ask HN: Are hourly priced dedicated servers for AI/ML a good idea?

2 points by red0point ↗ HN
On Hetzner or other providers, one can find lots of high performance dedicated servers, some even with graphic cards as powerful as a GTX 1080.

The problem is, you have to rent them for 100 EUR minimum per month plus set up - which doesn't make sense if you're a ML/AI engineer who just needs to train a model or two for a few hours.

Thus the plan is to buy some of those servers upfront and rent them out on a cheap hourly basis to ML/AI researchers, or even some cloud gaming companies. I've done some research on the technical side of things and I think this should be entirely possible to do.

I calculated that with about 50% usage one could sell an instance with a Nvidia GTX 1080 with an i7-6700, 64GB RAM and 500GB SSD for about 0.5 USD/h - which is significantly cheaper than AWS/GCE and probably ideally suited for AI/ML tasks.

But is this really a good idea? I'm always very excited about new side projects to do, but often overlook critical flaws in my ideas. For instance, why are companies like Hetzner not already doing this?

Thanks!

4 comments

[ 3.9 ms ] story [ 18.7 ms ] thread
minimaxir
This is literally the model that AWS/GCE does.

Due to their economy of scale, you won't be able to outprice them for on-demand usage unless you have an added value/gimmick.

some_account
I agree with previous poster. You will never be able to rent out machines cheaper than the cloud platforms without making a huge loss.
namanyayg
Another commenter answered succinctly, but for the next time a side project can be evaluated for viability by also calculating estimates for fixed and marginal costs & doing more competitor research. Has always helped me.
red0point
Thank you for your feedback.

I will do such research in more detail in the future.

Back of the napkin calculation results in 0.5 USD / h for an instance with a GTX1080 - compared with an equally GPU performant EC2 instance g3.8xlarge clocking in on about 2.28 USD / h.