[jcannell]

I'm glad you mostly? liked our service. Here's a better (long overdue) explanation: our billing system attempted to bill you because you had a negative credit balance (owed us) - most likely from forgetting to destroy an instance. However there's a threshold for any credit card charge around $5. I suspect that we later experimented with a different threshold criteria, which then caused the billing system to attempt to charge your card the balance owed - a long while after it was accrued.

We do now have better billing history reports so you can see what caused a charge, although sadly the customer service still isn't much better. Source: I am the CEO/founder and still handle most customer service.

[rckoepke]

Your service helped me immensely when I was getting started with machine learning, but had no budget and just a laptop. For me, everything “just worked”, and I learned a ton about Docker and created my own system to handle the ephemeral nature of the computing systems by automatically backing up training checkpoints, programmatically finding the next best “deal”/“bid” and resuming the training on a different instance.

Was very very cool, and I generally recommend it to anyone I come across who has modest needs (<25 GPUs).

I did find the throughout was not always accurate at the time. This was about 2 years ago. It was fairly frequent that a listing would say “300” to “1000” Mbps Up+Down, but would actually get an order of magnitude less to any of the big cloud services (GCP, AWS, etc). It wasn’t important to me that the speed was low, but it was important that it didn’t match what was “advertised”. For certain workloads I would have gladly paid more for higher throughput but that’s not really an option when the listings couldn’t be fully trusted.

I also heard there may be some opportunities to market your technology/platform for “on-prem”, “on-demand” GPU clouds for large enterprises so that a pool of corporate GPUs could be efficiently used and accurately billed to a variety of internal stakeholders. Could improve asset utilization for capital intensive on premise GPU’s.