Yes, we do offer following discounts:
a. If you are planning to buy datacenter wide license, we offer site license at much discounted price.
b. If you are going to buy more than 500 host licenses, discount will be given
c. If you are going to deploy the solution in developing countries, we give 'emerging markets' discount.

Since you have less than 100 hosts, cost per host per month is $12. Monthly License cost = 20 hosts x $12 per host = $240 per month

No, currently Enterprise subscription is only offered for annual terms.

Cancellation: Your subscription will automatically terminate after the initial term.
Return: Once the license is delivered, the sale is final and cannot be returned. You will have the option to discontinue your subscription after the initial term is over as we do not auto-renew.

yCrash server capacity is totally dependent on how much crash analysis it has to do. Here is the rough estimate you can use.
Infrastructure: yCrash Server is platform agnostic. It can run on any infrastructure (bare metal, cloud, container), any operating system(Unix, Linux, Solaris, Windows,...).
Software: All it needs is Java 8 (or above) version. No other external softwares needs to be installed.
CPU: Doesn't really matter. Any modest CPU capacity is fine.
Memory: You want to allocate 2x the size of your application's JVM heap size (i.e. -Xmx). Let's say you want to monitor multiple applications whose heap size (i.e. -Xmx) varies, then you want to pick the largest heap size. Let's say your largest application's heap size is 2GB, then you want to allocate 4GB as the RAM size requirement for yCrash server. If you would like to ensure high availability, yCrash server can be deployed on multiple nodes behind a load balancer.
Storage: Again storage depends on the amount of crash reports you want to archive. Approximately 100GB. yCrash's analysis results can be stored in local disk, NFS, AWS S3 storage, PureStorage (or any AWS S3 API compatible storages).

yCrash agent adds very negligible overhead (if at all it can be measured), for following reasons:
a. Dormant: yCrash agent doesn't always run your server. It runs only when it's triggered. Once triggered, it captures the dumps, transmits them to yCrash server and then immediately becomes dormant.
b. Duration: yCrash agent runs for less than < 1 minute when triggered. Within this time period it captures all the data and transmits to yCrash server. (Note: if your application's heap size is large then it can take a few more seconds to complete).
c. Non-Intrusive: Unlike APM tools which run within the JVM, yCrash agent run on the server (outside the JVM). They don't intercept every single transaction. Only when a problem happens, it reads and transmits dump files which are written to the disk. Thus It adds very negligible overhead.

Yes for selective few customers like apple we do host the tools exclusively for them. It is called as Hosted Enterprise Edition, where we will host tools exclusively for your organization (like http://youcompany.ycrash.io) in AWS cloud.