The real breakdown there is whether or not you are paying for datacenter space too. If you have to pay for power, HVAC, etc out of your budget, then yes, rent from the cloud. If you're like most small research groups, you're part of a larger university or industry department, and you're already paying building overhead costs, which include some sort of server room space. For that sort of budget, I'd be astonished if you needed more than about 10 tiles of server room space, and probably a lot less than that. This is extremely workable in the sort of environment I described.
About a year ago, my research group (I'm a grad student in high-energy physics at a major American university) received 30k$ in supplemental funding via the American Recovery and Reinvestment Act. I used the money to build our group a 39-node (156 core * hyperthreading) compute cluster. For that price, we could have rented 94 days of 156 small EC2 instances. If you try to factor in maintenance and development time, you can get a little bit longer of course, but I certainly spent less than a year's worth of work on it, so you'd less than double the EC2 rental length at grad student pay rates...
More technical details: HEP is embarrassingly parallel tasks with IO needs ranging from very heavy to very light, and with fairly large data storage needs. I bought commodity desktop hardware (core i7 930, fairly cheap X58 chipset mobos, low-end name-brand memory, mid-range PSUs, cheap as near free chassis, 1.5TB WD HDD per node), and a 48-port Cisco gigabit switch (with our IO throughput needs for some jobs, and the OS install scheme, didn't want to skimp at all on the networking). We already had a general purpose server machine (file, NIS, electronic logbook, etc), which became the headnode.
Each node PXE boot + NFSROOT from the headnode, which reduces the maintenance dramatically. The disks in each node contain swap and data storage, and are tied together with Hadoop's HDFS or gfarm (admittedly, this part is still in flux, because extremely high inter-node network traffic causes nodes to hang spontaneously.... Still working on that aspect) which will, when/if it works, give us much greater data throughput by eliminating the bottleneck to the headnode for data.
If you'd like to chat more about building a small compute cluster on a shoestring budget according to this sort of model, please feel free to email me at jay ess double-u at eff enn ay ell dot gov.