The idea that dedicated hardware appliances are always better, more reliable and higher-performance than commodity and/or PC hardware - in practice, with today's costs.
It's basically what Cisco wants you to believe; sure, the NPE in the router chassis only has a ~300 MHz ARM processor, but it's got all these ASICs (Application Specific Integrated Circuits) just for fast packet forwarding, FIB routing lookups and so on.
While that may be true, and I generally do favour using proprietary gear of that sort for routers and switches for a variety of administrative and MTBF-related reasons, the fact is that in the era of 3 GHz processors and 8 GB of RAM, often the presence of ASICs and CAM just doesn't matter -- the PC can still smoke that router. Sure, all the stuff is done in CPU instead of offboarded to dedicated hardware, and sure, it's all in processes subject to the ravages of a contended userspace scheduling environment in a general-purpose OS, but when you have 20x the CPU power, sometimes it doesn't matter - it still comes out way ahead, and much cheaper.
I learned this again recently when dealing with a fairly high-end PIX buckling to increasing packet processing loads in a growing VoIP environment (high packets-per-second cripple routers much more so than overall throughput per se, and VoIP audio streams consist of very large amounts of very small packets); the Linux firewall I set up as a stopgap measure for inter-VLAN routing in the meantime blew that thing out of the water.
Ditto for BGP. There is still a vivacious debate in the Cisco world about the minimum router specs needed to hold one or more full BGP views of the ever-growing IPv4 routing table, since so many router models are generally capable of it were they not skimpy on the RAM. Well, you know, Quagga and a solid Linux server with a great NIC and low-interrupt I/O tweaks can do wonders. :-)