Maybe I'm wrongly assuming, but I thought HDD bottlenecks could be at least somewhat negated by using RAID; at least that was going to be my solution for simultaneous streams.
As HFAT said, maybe. RAID will help, but probably not as much as you think.
A simple example, a two drive RAID0 can do a sequential transfer at about double that of a single drive. However, it cannot transfer two or more sequential reads simultaneously at double that of a single drive.
Access times are not reduced significantly by RAID, so the more streams you run simultaneously the more you are going to be seek limited and not transfer rate limited. Intelligent RAID control/software can mitigate this somewhat, but there is no magic.
A G850 looks nice, but don't SB Pentiums and Celerons lack QuickSync? Wouldn't this hamper video transcoding? Or am I making a bigger deal out of an increase in transcode time.
A big downside of QuickSync and any video card based transcoding is quality. That speed doesn't come for free. There is a reason x264 (generally considered the best quality H264 encoder) doesn't use anything except pure software encoding. You need to decide if the speed/quality tradeoff is worth it to you.
A Zacate build is also intriguing, but it just seems like it would be underpowered. Then again, I have no experience with Zacate.
I've only used a Zacate as a desktop CPU, never in a server. Subjectively it's faster than an Atom, but still no where near even a Celeron SB. If transcoding is in your requirements then I wouldn't use it.
I might be completely off here, but my personal experience with VMs (non bare-metal on an i5-760) is that they work best when you can devote one whole core to each VM you run. Am I wrong about this? I was hoping to go with a quad core to handle the VMs, but the i3 seems to be about the best compromise for what I'm looking for. Can anyone comment on how many VMs can be run smoothly on an i3 2100? And just for kicks how many VMs can be run smoothly on a SB Pentium/Celeron?
As Hfat said, you'll likely run out of RAM long before you run out of CPU unless you are doing something processor intensive on your VMs. Transcoding might count, but you could give it lower priority and it probably wouldn't interfere much.
At work we average over 20 VMs with 2-4 cores allocated each on 8 core machines. That's up to 80 virtual cores on an 8 core box. RAM is always the issue over CPU. Right now the highest CPU load on any VM host is 16%, so just over 1 core fully used.