Now, my experience with the atoms (and why I'm not selling 'em) is that they use more power than you expect; 50w each or so, down from 75w each for a core2duo with 8GiB ram. So I'm guessing this person is also in a region where power is cheaper than it is for me. (I am in the SF bay area, and thus I pay enough for power that margins on a power-hungry atom system would be rather small.)
You can get the operations overhead per machine down with a good pxeboot/automated provisioning system. some larger places I've worked, provisioning 5,000 machines was a relaxed afternoon of work. Yeah, physical servers are a bit more work than virtuals, but they don't have to be that much more work, if you have a good system.
That said, getting a VPS on modern dual-socket 8 or 12 core boxes with ecc ram usually gives you more bang for your buck, if the security tradeoffs are worth it.
Right, which is why I wouldn't use them for production, other than firewall. I've just oredered a http://www.supermicro.com/products/system/1U/5015/SYS-5015A-... kit for 199 EUR sans VAT, add about 35 EUR for 2 GByte Kingston ValueRAM, a port riser for 10 EUR and a dual-port Intel NIC (e.g. E1G42ETBLK) for about 100 EUR, a little flash DoM with pfSense and you've got a nice firewall.
I'm hoping most of the traffic will just remain on the Intel NIC (I'm running a transparent bridge), which is server-grade, so it will itself be not a source of corruption. The rest will be caught by checksums, hopefully.
Of course you roll your own, e.g with Linux VServer and HA (heartbeat/drbd). If Atoms were really low-power (they aren't, 8 W is noise in comparison to the chipset), offerend ECC support and decent mothboards with Intel NICs and IPMI onboard then they would be interesting hosts for virtual server guests.
As is, they're reasonable as firewalls (which are typically redundant) and SOHO servers. I hope ARM will deliver something competitive here.
You have relatively smaller seats and have to share with other passengers if you take the jet, whereas you can take the whole crappy 50 year old plane to yourself -- it's dedicated!
Well. I'll still take the former, but apparently some people prefer the later.
Sometimes the problem isn't about performance, rather it's about fault-tolerance and redundancy. Granted, if all of these servers are located in the same data center, redundancy takes a hit. But all else being equal, having one's software running on multiple machines reduces down-time.
Also, live migration doesn't help you of the hardware dies. (I mean, if the host is using shared storage, it is easy to bring you up on a new host, but you still crashed. you can run a dedicated off shared storage, if that is all you want.) You can only 'live migrate' off a 'live' host. Now, if you wanna get really expensive, vmware (and xen, if you want to do some patching) has the capability to essentially 'live migrate every Xms' - the idea being that they never bring up the second host unless the primary host fails; but it's ready, and only Xms behind. (they do some things I don't understand to insure that I/O to the shared block device is in sync) This is very expensive, though, because not only do you need to pay for 2x the hardware, you need a vmware license. (or to run a non-standard xen)
A dedicated server with mirrored drives and ecc ram is likely going to be more reliable than a VPS.
If you're looking at this end of the spectrum, just go with a VPS/Cloud system. You'll be paying about the same amount, maybe less. And you get drastically improved reliability if your provider is using failover and isn't using crap hardware.
Basically, would you rather take a business class seat on a jet or charter a 50 year old plane being piloted by a drunk hobo to fly somewhere?