If you were to build an AI/ML Model, what platform would you choose?

Managed services (GCP , AWS or Azure) for training and inference are costly. And using them will depend on your ability to price in the cost of the APIs & serving infra....

Do check out the new AWS Lambda configuration which allows for container images of upto 10GB in size. This allows you to bundle the model weights & all peripheral packages in a single lambda function.

This approach gives a good trade-off from using managed services vs using bare-bone servers. And can turn out to be cheaper to run in the long run.

Well depending on how you hosting your SaaS service what infrastructure you use and how much processing beef do you have available / going to need, how many requests on the SaaS you are expecting and what is the maximum response time you can get away with.

Well these are general questions on deployment.

Other questions is how complex is the ML/AI model going to be? Is it a DeepLearning-model? What is the amount of data is it just text, formated text and what is the average amount of data going into the model for a inference.

What is the amount of training data you have

Is this AI/ML module the core of your SaaS?

Does an existing third-party service already cover what you need? How good is it? How expansive is it? If there is one could you or would it make sense to compete against it by doing it yourself or not? Is it generic NLP-based it is very unlikely you can compete against an existing generic solution unless it covers a specific niche and therefore the amount and time to build up something yourself is justified or you have any other good reasons like data ownership/protection and privacy.

Depends on :

1. How you want to use your mode- in a SaaS, as an API, etc.
2. Is it pre-trained or you plan to train on the server/ in production.
3. How heavy is it? Are you doing full-on video enhancing which would take 2 minutes to process one request or just classifying images?

I'm biased as a programmer that I just have my models running on barebones servers or Linux instances that I can manage. If you don't want to manage such things on your own, use Google Cloud AI. But in my opinion, I found Google Cloud AI more difficult (i.e. annoying), pricey, etc. than a single server instance. But answering the above questions would help you finding the right solution.