I have found peering to have additive value; a lot of 1-2 Mbps peering sessions can save as much money for you as a single large traffic peer. The more traffic, the stronger the case for peering.
Sadly, this completely ignores the cost of implementing and maintaining peerings. BGP does not exactly configure itself, and current routing technology is somewhat frail -- if it breaks, somebody has to pick up the pieces; if suboptimal routing results from a peering, somebody will have to go and tweak things.
Since the Internet is not exactly static, these things creep up from time to time even after a peering is established.
An exhaustive itemization of the costs of any given peering is a of vital component of a cost-benefit analysis, particularly where much of the benefit is a reduction in monthly usage based billing costs, or a deferral of an upgrade of a flat rate contract, rather than installing new parallel connectivity to meet the demands of traffic growth. While such a list will vary from organization to organization, and some organizations may be tricky to complete, one can consider the primitive case of a transition from being singly homed to a transit provider to bring up an initial peering.
Among the things to be considered: PA address space, BGP (and making sure that the routers can handle the load, and so can the people operating them), avoidance of accidental transit, what happens to the peering traffic when the peering fails from time to time (or fills up with growth traffic over time), NOC-to-NOC coordination, impact on actual and potential SLA offerings, and so on. The immediately subsequent peering may not make much of an impact on many of these however there are real incremental costs. Some of these may not rise linearly in proportion to the number of peers (e.g. there may be a step change),
but rather may be a function of that number, the number of your routers which talk eBGP, and your overall traffic load.
I know a handful of cases where a broad peering strategy had a fairly clear negative impact on the bottom line even though the saving in transit fees was both directly measurable and large. Anecdotes are not general proofs, but do underline the uncertainty in your statement: "a lot of ... peering sessions CAN save as much money ... as a single large traffic peer". [emphasis mine]
Unfortunately, discussions here seem to focus mostly on measuring the reduction in transit fees. Wouldn't it be nice if this could be coupled with reasonable discussion about the increase in other costs, and how for some networks these costs are much higher than for others?