Sometimes it's cheaper and more practical to just buy the hardware.
I had a similar issue with a company I worked for before, we had an account on Azure and the machine was expensive but also slow and after 4 months we decided to buy a used server for $1000 and closed the Azure account and never looked back.
Once ready I opened up the render queue page in Chrome on it and as soon as a new job comes in it renders it. It get's it's jobs from a SSE connection. I use Three.js as the 3D library. It works very well now but in the beginning it had some hitches because while the GPU is strong the CPU on the Nano is very weak, so decompressing the models and the initial asset loading where not as fast on beefy modern x86-64 but nothing a few quick optimizations could not fix. The final hurdle was occasionally losing context but I just added a listener for those events and restarted the job if it failed. Another fun solution was checking to see if the image were actually rendered instead of just a transparent PNG by checking the file size of the outputted image, if the image is impossibly small it gets sent back to the job queue.
depending on usage, you could start saving money over rented GPU time in a single-digit number of weeks.
