If you're a hobbyist training a model on a small dataset, the current prices of using cloud computing and running a GPU-backed VM for a couple hours on demand will still be cheaper in the long run.
* University run HPC facilities.
* Your university might have a deal with government funded facilities like the NCI (National Computational Infrastructure). Their computer Raijin has GPU's.
* Or, they might be a participating University in AARNET (Australian Academic and Research Network), which next year will have web-based Jupyter Notebooks for smaller stuff.
* I 100% guarantee there are multiple unused snowflake GPU clusters sitting under academics (with more money and influence than sense) desks that were purchased under the auspice of being "absolutely essential for their work" or "its the end of the year and we HAVE to spend this money" and because hey ignored all advice from IT it never works and is unsupported. Find them, be the support guy (if you have time) on the condition you get access.
Google Colaboratory allows for a free K80 GPU; not as a strong as a GTX 1080/2080 mentioned in this article, but good enough to prototype things, as it's normally ~$0.57/hr on GCP. (how textgenrnn uses the free Colaboratory notebook: https://minimaxir.com/2018/05/text-neural-networks/)
In the case of textgenrnn, I use the Colaboratory notebook exclusively nowadays. :P