Sparsity and non linearities in data needs to be factored in. Then data aging and maturity. The way anomaly gets detected matters based on these. I have seen consistent Alert floods whereas in reality they were temporal fluctuations or temporal correlation between variables. Measurement and sampling time in timeseries across touch points is another important design decision.
1Cloudhub an AWS partner has a very affordable product called Costimize for AWS to help save costs. I will be happy to help. Contact Sankar.nagarajanAT1cloudhub.com
Secondly, you can use AWS spot instances for GPUs which costs less.