122 the Internet hierarchy is only entered into with interconnection partners considered trustworthy. In conclusion, ISPs will enter into peering agreements only when their prospective peering partners have a network of similar geographic extension and have invested into comparable network bandwidth which can guarantee an equivalent level of quality of service. Furthermore, ISPs generally require traffic flows to be roughly similar. For this it is not important to have the same number of customers, only the same type of customers. Lastly, Tier-1 ISPs require peering partners to be particularly trustworthy in the sense that their personnel is knowledgeable in resolving routing problems quickly and correctly.108 The above analysis shows that a transit interconnection requires far less investments into network infrastructure as well as human resources than peering does. Since a transit contract also offers universal connectivity, whereas peering offers only limited coverage of the Internet, a smaller ISP will often find it less costly to pay for transit services in order to reach universal connectivity than to meet the network requirements necessary to peer with several ISPs of higher hierarchy levels. Peering is therefore not always preferred to transit interconnection, even though it generally involves no monetary compensation for the exchange of traffic. Transit fees are justified by the fact that transit givers invest more into their network infrastructure than transit takers. 7.3 Dominance on Tier-1 level The preceding section focused exclusively on an ISP’s decision on whether to interconnect with another ISP via a peering or a transit agreement. It was shown that the differences in the terms for peering or transit are not driven by the number of IPaddresses an interconnection partner offers access to. Rather, factors such as the type of customer mix and the relative geographic extent of the two networks were shown to be important. The focus now turns to an ISP’s principal decision on whether to interconnect with another ISP at all. For this decision the network reach provided by a potential interconnection partner is of fundamental importance. The ultimate goal of network interconnection is to provide universal connectivity. All ISPs not active on the highest level of the Internet hierarchy need at least one transit agreement with a Tier-1 ISP or with an ISP that has such a transit interconnection. Tier-1 ISPs know 108 The above conclusion is supported by the “Settlement-Free Interconnection Principles” which Level 3, considered to be a Tier-1 ISP by several recent sources (i.e. D’Ignazio and Giovannetti, 2006: 18; Ermert, 2005), publishes on line (http://www.level3.com/1511.html, site last visited on Nov. 16, 2006). In these principles the company states that it will consider peering only with firms which have a network similar to that of Level 3 in terms of size, reach, scale, diversity and reliability. More specifically, the guidelines call for a comparable number of private interconnection points, for “a general balance of inbound and outbound traffic” and “sufficient capability, processes and tools to assure adequate interconnection and routing quality.” 123 that lower level ISPs depend on their transport services. It is therefore important to understand whether Tier-1 ISPs could have an incentive and the means to discriminate lower level ISPs. As was discussed above, the demand for Internet backbone services on the logical layer of Internet service provision is a derived demand from the end-user demand for universal connectivity on the retail level of Internet service provision. In the retail market, universal connectivity signifies that all other end-users and content providers on the Internet can be reached via one’s home ISP. In the Internet backbone services market, universal connectivity signifies that an ISP can send and receive data to and from all IP-addresses allocated to public uses in the Internet. The literature on Internet backbone services does not differentiate clearly between universal connectivity on the applications layer and universal connectivity on the logical layer of Internet service provision. The difference is, however, of importance when, as is often the case, the number of “customers” attached to an ISP is used as the measure for the Internet coverage the ISP provides. This is a concept relevant on the applications layer of Internet service provision. On the logical layer a customer of an ISP can, however, be either an end-user, representing only one of millions of Internet-Protocol addresses (IP-addresses) or another ISP, representing an important fraction of all registered IP-addresses. For the purposes of measuring Internet coverage on the logical layer of Internet service provision it is more meaningful to speak of the coverage of IP-addresses which this ISP can offer as a peering partner. Transit services, by definition, offer universal connectivity. Economists have developed models that try to capture the interconnection incentives of ISPs. Theoretical models are of particular relevance in the context of merger policy because competition authorities cannot look at actual market conduct for their analysis. Policy makers depend on predictions derived from economic modeling to understand whether efficiency considerations or attempted exclusionary conduct are at the core of proposed mergers. The model that was influential in the merger proceedings surrounding the MCI and Worldcom merger in 1998 and the attempted merger of the resulting firm MCI/Worldcom and Sprint in 2000 offered initial interesting insights into the interconnection incentives of ISPs with asymmetric installed customer bases. Since then, the literature on interconnection incentives of ISPs has refined this model considerably. The following two subsections shall review the theoretical debate on the interconnection incentives of ISPs in more detail. 7.3.1 The Crémer, Rey and Tirole model The reasoning that led the competition authorities to impose severe conditions on the merger of MCI and Worldcom in 1998109 was based to a great extent on one of the earliest theoretical models, which tried to capture the strategic interconnection deci- 109 MCI had to divest its Internet operations before a merger with Worldcom was approved (European Commission, 1998). 124 sion of ISPs. From this model by Crémer, Rey and Tirole (2000) the conclusion was drawn that an ISP that is dominant in terms of attached customer base in the retail market, would have the means to dominate the market for Internet backbone services. It would either refuse to peer with smaller rivals or price-squeeze them out of the market (Kende, 2000: 22-23).110 The model by Crémer, Rey and Tirole builds on the Katz and Shapiro (1985) model of network externalities introduced in section 6.3.3. Crémer, Rey and Tirole model the number of firms in the market as exogenously given and there is no product differentiation. Consumers exhibit different basic willingness to pay for the service but show no technology preferences and express the same evaluation of the network effect. In a first scenario Crémer, Rey and Tirole focus on interconnection decisions in an asymmetric duopoly situation. The existing users of the two networks are assumed to be locked-in. The networks compete à la Cournot over the addition of new customers to their networks. The choice of the quality of interconnection between the networks is introduced as a strategic variable. In the first stage of the game the quality of interconnection is determined by the network which sets the lower quality level. Given the interconnection quality, the networks then choose their capacity and prices. In equilibrium, the network with the larger installed customer base prefers a lower level of interconnectivity than the smaller rival because it can expect to dominate the market for new customers. This equilibrium solution to the model has been the basis for arguing that a dominant Tier-1 ISP would have an incentive to refuse or degrade interconnection with rivals, especially in dynamic markets with high growth potential. Two effects determine the equilibrium outcome. Firstly, lower connectivity levels lead to an overall demand reduction in the market, which negatively impacts all firms. Secondly, reduced interconnectivity introduces an element of quality differentiation between the firms, which in this model, can only differentiate among themselves along the dimension of network size. The network with the initially larger locked-in customer base profits from this quality-differentiation effect because it can offer more benefits from network externalities to new users. The bigger network trades off the negative effect of the demand reduction against the positive effect of the quality differentiation. The incentive to choose a lower level of interconnection quality is the more positive the stronger the network externalities are and the greater the difference in installed bases is. A differential analysis shows that the incentive to increase the level of interconnection quality may rise when the number of locked-in customers is already very large, because then the demand-expansion effect triggered by a larger network becomes so important, that good quality interconnection is preferred. In a second scenario Crémer, Rey and Tirole (2000: 456ff.) analyze a market initially consisting of four equal sized ISPs. As long as all four have the same size, all 110 Crémer, Rey and Tirole argue that a customer in this model can be either an end-user or an ISP. They do not differentiate between the two. 125 are interested in a good quality interconnection because all profit equally from a demand-expansion effect. The elicitor of a quality degradation would suffer the same negative demand reduction as its three rivals without a compensatory gain from a positive quality-differentiation effect. The authors then show how the incentives to interconnect change when two of the ISPs merge and the resulting market of three ISPs then includes one firm with an installed base of at least the size of the combined installed bases of the other two firms. In this scenario the largest firm is not interested in deteriorating the quality of interconnection with both of the rival networks. However, in some circumstances, it can profit from a targeted degradation strategy, in which it refuses good quality interconnection with one of the smaller rivals while it continues good quality interconnection with the other rival. This conclusion depends on the non-targeted firm not offering transit services to the targeted firm.111 The positive quality-differentiation effect will then result in the targeted firm not attracting any new customers while the dominant firm and the non-targeted firm gain more customers (even though the non-targeted rival profits more from the quality-differentiation effect). It was especially this result that competition authorities relied upon in their decision on the merger by MCI and Worldcom in 1998. 7.3.2 Critique of the model by Crémer, Rey and Tirole and alternative modeling In chapter 6 it was discussed that the results of Katz and Shapiro (1985) depend critically on the additional assumptions in the model besides the network externalities. These additional assumptions introduce elements of market power such that the model cannot be used to prove that network externalities lend market power to large network operators. The same holds true for the adaptation of the Katz and Shapiro model by Crémer, Rey and Tirole. Here also there are significant additional assumptions in the modeling set-up which lead to the result that the largest firm prefers a lower level of interconnection quality compared to its smaller rivals. Below it is discussed whether these assumptions are relevant for the market for Internet backbone services. Market entry conditions First, consider the assumption of a fixed number of firms in the market. This assumption does not correspond well to the thousands of active ISPs observable in reality. If at all, then this assumption may apply to the market for Tier-1 ISP services in which only 5 to 10 ISPs are active. But whether this market has structural barriers 111 Crémer, Rey and Tirole (ibid., 458) argue that the dominant firm can limit the capacity of the interface with the non-targeted network to such an extent that the capacity is only sufficient to provide good quality interconnection for the traffic of the non-targeted network but would result in very bad interconnection quality if the traffic should grow to encompass also the traffic of the targeted network. 125 are interested in a good quality interconnection because all profit equally from a de and-expansion effect. The elicitor of a quality degradation would suffer the same negative demand reduction as its three rivals without a compensatory gain from a positive quality-differentiation effect. The authors then show how the incentives to interconnect change when t o of the ISPs merge and the resulting market of three ISPs then includes one firm with an installed base of at least the size of the combined installed bases of the other two firms. In this scenario the largest firm is not interested in deteriorating the quality of interconnection with both of the rival networks. However, in some circumstances, it can profit from a targeted degradation strategy, in which it refuses good quality interconnection with one of the smaller rivals while it continues good quality interconnection with the other rival. This conclusion depends on the non-targeted firm not offering transit services to the targeted firm.111 The positive quality-differentiation effect will then result in the targeted fir not attracting any new customers while the dominant firm and the non-targeted firm gain more customers (even though the non-targeted rival profits more from the quality-differentiation effect). It was especially this result that competition authorities relied upon in their decision on the merger by MCI and Worldcom in 1998. 7.3.2 Critique of the model by Crémer, Rey and Tirole and alternative modeling In chapter 6 it was discussed that the results of Katz and Shapiro (1985) depend critically on the additional assumptions in the model besides the network externalities. These additional assumptions introduce elements of market power such that the model cannot be used to prove that network externalities lend market power to large network operators. The same holds true for the adaptation of the Katz and Shapiro model by Crémer, Rey and Tirole. Here also there are significant additional assumptions in the modeling set-up which lead to the result that the largest firm prefers a lower level of interconnection quality compared to its smaller rivals. Below it is discussed whether these assumptions are relevant for the market for Internet backbone services. Market entry conditions First, consider the assumption of a fixed number of firms in the market. This assumption does not correspond well to the thousands of active ISPs observable in reality. If at all, then this assumption may apply to the market for Tier-1 ISP services in which only 5 to 10 ISPs are active. But whether this market has structural barriers 111 Crémer, Rey and Tirole (ibid., 458) a gue that the dominant firm n limit the capacity of the interface with the non-targeted network to such an extent that the capacity is only sufficient to provide good quality interconnection for the traf ic of the non-targeted network but would resul in very bad interconnection quality if the traffic should grow to encompass also the traffic of the targeted network.