Colocation without being physically nearby can be costly and a logistical nightmare. So-called "remote hands" can cost upwards of $100/hour, though initial inst

Colocation without being physically nearby can be costly and a logistical nightmare. So-called "remote hands" can cost upwards of $100/hour, though initial installation might be included. Nobody wants to be in the position of helping someone fix their broken hardware (or what they think is broken hardware) as it's an absolute time sink and a sure way for the customer relationship to sour when things go sideways. If you're doing this remotely, you need to send at least one backup, which likely means paying 2x. When things can't be fixed over IPMI or the BMC (presuming you trust keeping those connected to a network, internal or otherwise), just cross ship a new one with the old. But be sure to keep at least one backup on site as logistical problems can quickly crop up, and that's assuming you're prepared to respond immediately and remain engaged. Regarding "very fast server", note that the biggest expense to colocation isn't space but power. There are lots of smaller colocation providers who would be happy to sell 4U of space rather cheaply, but you won't be able to power 4 x Xeon E5 or E7 machines. You'd be lucky to power just one. I've been running several Xeon E3-1230s (v2 and v3) which top out at about 65W or less--65W for the v3 and 55W for the v2, IIRC. That's total--the peak power draw under load at the power outlet is less than the published TDP for the CPU alone! I recently built a Supermicro EPYC 3201. AMD publishes 30W as the EPYC 3201's TDP, but peak draw under load at the outlet is 45W. The EPYC 3201 compiled Ubuntu's GCC package 11 minutes slower than my E3-1230v3 (219m vs 208m). The performance isn't surprising, but I'm a little disappointed at the 45W power draw. The E3s spoiled me. I'm still eager to switch to the EPYC simply because of the Spectre fiasco. It's why I got the EPYC 3201, which is 8 cores without SMT, instead of the 3251 with 8 cores and 16 threads.[1] Plus, without load the EPYC box only draws about 20W, IIRC, which is nice even if irrelevant to my immediate costs. Probably these types of machines are not as powerful as you had in mind. But a colo provider may not let you plug in multiple servers which could exceed your power budget even if you promised not to run all of them under load simultaneously. When you have multiple machines power becomes a significant constraint. [1] The TDP for the 3251 is 35W. The 3201 clocks memory at 2133MHz, but the 3251 clocks memory at 2666MHz. I think the extra 5W comes more from the increased memory speed than the SMT. The DIMMs themselves would also draw more power, so I would expect peak power draw at the outlet for the 3251 to be greater than 50W (45W + (5W TDP difference)), closer to 55W or even higher. (My E3-1230v3s clock memory at 1600MHz, which makes think DIMMs are actually a significant power draw. I think the v2s use 1333MHz, which may explain the ~10W lower power draw--also the Intel changing the voltage regulatory between v2 and v3.)