A SIMPLE MODEL OF INTERCONNECTION AND COST

Let us consider a single access network A (e.g., Comcast, Verizon, Time Warner or AT&T) interconnecting with a content delivery network CD (e.g., GoogleZYouTube, Akamai or Limelight).

There is a slight asymmetry to the situation of CD and A, which is worth keeping in mind. Network A has a substantial base cost that existed prior to the growth of high- volume commercial content. For example, network A needs to have both the last-mile drops, distribution network and backbone network to support basic communication and other services, regardless of whether this same infrastructure is shared to deliver video. By assumption, C_a comprises only the incremental, usage-based component of the total cost of A. CD, on the other hand, exists only to deliver this content, and will tend to view their total network-related costs as associated with the delivery of content.

The sources of dollar flows (payments) in this model include the payments that content ownersZproducers make to network CD for content delivery services, and the payments that end users or ‘eyeball’ customers pay to their broadband access provider, network A, for their access and usage. (To keep things simple, we are ignoring other sources of third-party payments such as government subsidies or advertising.)²³ Finally, there is the potential of a payment flow, P, between networks CD and A, as follows:

• P = 0: Traditional revenue-neutral peering, where each network covers all its internal costs from its own customers.

• P > 0: Payment from CD to A to help cover some of the internal costs of A. Such a payment would be part of C_cd and presumably these costs are then passed through by CD to the content producers, who pay more to CD and thus indirectly cover the cost of transporting their content across A. This outcome is fairly common in today’s content delivery network market.

• P < 0: Payment from A to CD. This payment arrangement seems uncommon, but makes sense in certain circumstances. Consider the case where A is a small, rural ISP. If there is no direct connection between CD and A, all of the content from the producers will come into A over a potentially very expensive transit link. Having CD make a direct connection to A may greatly reduce A’s costs. However, if A is small, it may not be cost-effective for CD to connect to A; the connection might actually increase CCD, not reduce it. In this case, it might make sense for A to pay CD.

16.3.1 A Reality Check: The Size of CA

To better understand the current debate over usage costs, it is worth considering the potential magnitude of these costs. Unfortunately, the lack of publicly verifiable informa­tion about the terms and conditions for interconnection agreements makes it difficult to estimate C_a, but we believe it is possible to rely on anecdotal evidence to identify some reasonable bounds on estimates of what these costs may be. In an earlier paper, we esti­mated the cost at $0.10 per gigabyte (GB) (Clark, 2008),²⁴ not including costs related to the access network. Other writers have suggested lower numbers, in the range of $0.07 to $0.10/GB.²⁵ Depending on the extent to which access network costs are allocated as fixed or usage related, the per GB costs may be considerably higher, perhaps $0.20 to $0.30/GB. These estimates are applicable for large urban/suburban broadband wireline access networks with low costs for their transit.²⁶ At one extreme, Netflix estimated that the costs were in the range of $0.01/GB.²⁷ We believe such an estimate is unreasonably low.

For example, Sandvine (2011) reported that average (mean) monthly download usage was 18.6 GB in the spring of 2011. If usage costs $0.20/GB, it would mean that the per customer usage-related monthly costs are close to $4. To watch a 90-minute movie in HD from Netflix (at about 5 megabits per second, or Mbps) would cost about $0.65. By 2013, mean monthly usage had doubled to 38.6 GB.²⁸ Although not precise, these numbers imply that the emergence of high-volume content (video) has generated substantial new costs C_a.

16.3.2 Managing Access Costs CA

There are only four ways that access network A can manage usage-related costs CA:

• Lower the total costs: by careful design of their network, specifically with attention to where the content delivery networks interconnect with them, they may be able to reduce the actual CA incurred.²⁹

• Payments from CD to A (P > 0): network A can negotiate to have the content network CD compensate them for some of these costs.

• Raise retail prices for consumers:

- allocating an equal share of CA to all customers;

- creating mechanisms that will discriminate among users based on some proxy of usage, and increase the price of service for these users.

• Internally subsidize by accepting reduced margins: if the ISPs such as network A are competitive, however, then this is not a sustainable option.

16.3.3 Why Might CD Pay A (P > 0)?

When CD networks connect to large access networks A, payment seems to most com­monly flow from CD to those access networks.

One possible answer is that network A has a better bargaining position, because it holds a terminating monopoly with respect to its customers. However, this pattern of payment may be just as much the result of a persistent norm or common practice that is older than the emergence of high-value commercial content: the presumption that money flows and packet flows go in the same direction. That is, if X is delivering packets to Y, then (if there is payment) X pays Y and not the other way around. Of course, the difficulty of justifying this assumption is what led to revenue-neutral peering in the first place. One rule that ISPs use to determine if they will agree to revenue-neutral peering is whether the traffic in the two directions is roughly in balance. If it becomes unbalanced, one or the other party may employ it as a useful justification for renegotiating the agree­ment. Anecdotally, if there is complaint about the imbalance in costs associated with the asymmetric traffic, it is often the party receiving the excess traffic that complains. But increased traffic (say from X to Y) adds to the costs both for X and Y. So why would X pay Y?

There seems to be an unstated assumption that a transfer is of more benefit to the originator than the receiver, so the sender should be expected to cover more of the delivery costs. This assumption is not always true: for example, when a user downloads a large open-source software package (e.g., a Linux release) the benefit is essentially all to the receiver. However, these cases seem to be ignored as part of the current regime of bargaining.

So long as payment between the parties is an acceptable outcome, the parties will find it profitable to interconnect so long as the net benefit to all parties (including the pay­ments) is positive. Paid peering can lead to more direct connections, which presumably reduces overall system cost and increases total surplus.³⁰

This discussion concerns how costs of delivery are covered: they concern transport payments not content payments. However, the possibility of non-zero payments also raises the possibility that one actor (e.g., the access network A) might have enough market power (e.g., because it is a terminating monopoly with respect to its customers) to demand a payment from CD that exceeds its internal costs CA. In this case, we should assume that the payment is not just a transport payment P_t, but includes a content payment, P_c as well. Regulators and industry observers have worried that access net­works might have enough power to demand content payments, and this would signal the potential for unacceptable discrimination and manipulation in the business of content production. So the obvious question follows: if we allow non-zero values for P, how can we distinguish content payments from transport payments?

16.4