`(cid:3)
`Trial
`INDEX Project Report #99-010W
`
`Richard J. Edell
`
`Pravin P. Varaiya
`
`April 16, 1999
`
`© 1999 U.C. Regents. All rights reserved. Do not redistribute this document
`electronically. This document may be found on the web at:
`http://www.INDEX.Berkeley.EDU/reports/99-010W
`
`Abstract
`The Internet Demand Experiment or INDEX is a market and technology trial. Its objective is to determine
`how much users value different qualities of service for Interent access. Findings from the trial imply that
`today’s system of flat-rate pricing by ISPs is very inefficient. Flat-rate pricing wastes resources, requires light
`users to subsidize heavy users, and hinders deployment of broadband access. INDEX is a prototype of an
`alternative ISP model that offers differentiated-quality service on demand, with prices that reflect resource
`cost. In this alternative ISP consumers pay less, suppliers increase profits, and the deployment of broadband
`access is facilitated.
`
`The INDEX Project is supported by the National Science Foundation, Cisco Systems, SBC Communications and the
`California State MICRO Grant program.
`(cid:3)
`We thank J¨orn Altmann, Karyen Chu and Hal Varian for use of their unpublished results. A version of this paper was
`presented as the keynote talk at Infocom ’99.
`
`1
`
`GUEST TEK EXHIBIT 1009
`Guest Tek v. Nomadix, IPR2019-00253
`
`
`
`(cid:3)
`Providing Internet Access: What We Learn From INDEX
`
`Richard Edell and Pravin Varaiya
`edell@eecs.berkeley.edu, varaiya@eecs.berkeley.edu
`Department of Electrical Engineering & Computer Science
`University of California, Berkeley, CA 94720
`
`Abstract
`
`The Internet Demand Experiment or INDEX is a market and technology trial. Its objective
`is to determine how much users value different qualities of service for Interent access. Find-
`ings from the trial imply that today’s system of flat-rate pricing by ISPs is very inefficient.
`Flat-rate pricing wastes resources, requires light users to subsidize heavy users, and hinders
`deployment of broadband access. INDEX is a prototype of an alternative ISP model that offers
`differentiated-quality service on demand, with prices that reflect resource cost. In this alterna-
`tive ISP consumers pay less, suppliers increase profits, and the deployment of broadband access
`is facilitated.
`
`1 The unavailability of differentiated quality service
`
`A major contribution of network engineering is the development of techniques (signalling proto-
`cols and associated algorithms) for using the same set of network resources (links and switches
`or routers) to simultaneously provide different services to end users. This is sometimes called an
`integrated services model to emphasize network support of different applications such as real-time
`voice and video and non-real time data transfer. One may also say that these techniques allow
`provisioning of differentiated quality service. The latter characterization emphasizes the flexible
`transport capabilities of the network and frees users to select the best way to match service quality
`to the demands of their application, time, and budget.
`
`Although it is possible to provide it, differentiated quality service is not sold in the marketplace.
`There may be for two reasons for this, one economic, the other technical. To offer differentiated
`quality requires the design and testing of economically viable alternatives to the flat-rate pricing
`model adopted by virtually all Internet Service Providers (ISPs). It also requires a technology to
`package such a service in forms that users can purchase, provide the means for users to express
`their demand, signal the network to provision the requested quality, and generate accounting and
`billing records. The INDEX trial tests alternative pricing models and a technology to implement
`those models.
`(cid:3)
`Research supported by National Science Foundation, State of California MICRO program, CISCO, and SBC. We
`thank J¨orn Altmann, Karyen Chu and Hal Varian for use of their unpublished results. A version of this paper was
`presented as the keynote talk at Infocom ’99.
`
`1
`
`
`
`2 THEINDEXTRIAL
`
`2
`
`Figure 1: In the first experiment a user instantaneously selects her access speed and pays per minute
`of connect time. The control panel also shows the accumulated charges. Prices are selected ran-
`domly.
`
`2 The INDEX trial
`
`INDEX offers its subjects (customers) differentiated quality Internet access at home. The ongoing
`trial started in April 1998, and this paper reports findings based on data from the trial. The 70+
`customers are students, faculty and staff of the University of California at Berkeley. There has been
`some turnover among customers due to changes in residence.
`
`Every customer participates in a sequence of service plans or experiments, each lasting six to ten
`weeks. A service plan is characterized by a menu of service quality/price combinations. So each
`service plan implements a particular pricing model, parametrized by quality choices and prices.
`Customers pay by credit card. The value to a customer of a particular service quality is measured
`by how much money and time she spends consuming it.
`
`INDEX collects detailed statistics on customer behavior. The basic data include byte counts in
`each direction, quality choice, accounting, and connection information. The time granularity is the
`minimum of one minute and change in quality choice. These data, with adequate protection of
`customer privacy, are available for study by researchers. The website: www.INDEX.Berkeley.EDU
`gives the details.
`
`Figure 1 is a snapshot of the ‘Choices’ panel of the INDEX ‘Control Center’ or CC for the first
`experiment. The CC is the interface running on the desktop through which a customer controls
`access to the Internet. In this particular service plan a customer can select one of six different speeds
`(8, 16, 32, 64, 96, 128 Kbps) at a cost ranging from 0 cents for 8 Kbps to 5.7 cents per minute for 128
`Kbps. The selected service quality (speed in this case) is provisioned virtually instantaneously.1 The
`lower left corner of the CC indicates the current status of the connection—“connected at 96 Kbps”
`in this snapshot.
`
`1The highest quality service that INDEX provides is a permanent 128 Kbps ISDN channel to the UC Berkeley campus
`network. There are plans to provide broadband service over ADSL.
`
`
`
`2 THEINDEXTRIAL
`
`3
`
`The lower right corner of the CC shows a ‘spending meter’ that can be toggled to reveal the cost
`incurred until now during the current session, day, or month. (The figure shows the cost of the
`session to be $0.00, indicating the session has just started.) The spending meter is updated each
`minute, so customers can, if they wish, be aware of the cost of the resources they are consuming.
`Flat-rate charges, by contrast, deny users all information about their resource consumption. As a
`result, ISPs have to place inefficient restrictions on subscribers to limit resource consumption.
`
`Another distinguishing feature is that an INDEX customer can instantaneously shift between dif-
`ferent service qualities with no effort (beyond a mouse click). The quality may be changed during a
`session, if the user wishes. By contrast, when an ISP does offer different access speeds, the service
`is segmented into tiers: users must pick a single speed, and it is not possible to change that speed.
`
`How much an INDEX customer values different speeds is measured directly by her purchase of the
`different speed options. The prices that are seen in Figure 1 are randomly selected and varied each
`week. Different customers face different prices. The price variation is large so that the demand
`estimates are robust.
`
`Service quality has many dimensions besides speed, and INDEX service plans are designed to
`explore these dimensions. Furthermore, how a service is packaged into a commodity and priced
`makes a difference in demand, and some plans are designed to explore those differences.
`
`By March 1999 the earliest subscribers participated in six experiments: (1) Symmetric bandwidth,
`(2) Asymmetric bandwidth, (3) Volume pricing, (4) Volume plus capacity charge, (5) Self-selecting
`tariff, (6) Enhanced flat-rate tariff. Figure 1 illustrates the choices in experiment (1), but it must be
`remembered that the prices are randomly selected for each customer and for each week. The aim is
`to determine sensitivity of demand for a certain speed to the prices of that speed and its substitutes.
`From INDEX data we learn that the price sensitivity is very high.
`
`Experiment (2) is similar except that customers separately choose and pay for bandwidth in the up-
`stream and downstream directions. The experiment is motivated by cable TV and DSL access, both
`of which have asymmetric speeds. The objective is to find out if users are aware of the asymmetry
`in their traffic pattern and make use of this asymmetry to reduce their bill. The answer is most are
`aware, and they do reduce their bill.
`
`In experiment (3) users can select 128 Kbps and pay so many cents per megabyte of upstream
`and downstream traffic, or select free 8 Kbps service. We can compare user behavior when the
`commodity is megabytes of data transfer versus minutes of connect time.
`It turns out that the
`connect time goes up dramatically under volume pricing compared with connect time charges.
`
`In experiment (4) a user incurs a volume or per megabyte charge plus a per minute connect time
`charge. This form of charge can reflect the bandwidth and buffer resources needed to support the
`user’s traffic. (See [1, Chapter 8].)
`
`In experiment (1) the plan offers a per minute connect time price, and in experiment (3) the plan
`offers a per MB price (at 128 Kbps). In experiment (5) users are allowed to pick a convex combi-
`nation of an offered per minute connect time price for 128 Kbps and an offered per MB price. The
`user must choose the combination at the beginning of each week and the combination is then fixed
`for the rest of the week. If the traffic to be generated during the week can be predicted, the least
`cost option for the user is either the pure volume or the pure connect time charge. The aim of the
`
`
`
`3 TODAY’SISP
`
`4
`
`Figure 2: Customer choice panel for Flat Rate experiment. This customer is choosing to pay $7.50
`for one week of unlimited 64 Kbps access. This choice is made once per week. During the week,
`the customer may select higher speeds at an additional per minute charge. The experiment combines
`flat rate and usage-based charges.
`
`experiment is to find out how well users can predict their traffic pattern.
`
`Experiment (6) is designed to test how much users value flat-rate pricing. At the beginning of each
`week consumers purchase unlimited usage for one week at a particular speed. They may connect at
`a higher speed for an extra cost per minute. Figure 2 shows the CC for a user who was offered one
`week of unlimited usage at a charge ranging from $0 for 8 Kbps to $15.00 for 128 Kbps and chose
`to pay $7.50 for 64 Kbps, while retaining the option of selecting 96 Kbps at 0.8 cents and 128 Kbps
`at 1.6 cents per minute.
`
`Future experiments are aimed at estimating how much users value reduced blocking in modem
`pools, reduced congestion, as well as how much users shift their demand over time in response to
`time-of-use pricing.
`
`The different INDEX service plans are alternatives to the prevailing dominant ISP model of flat-rate
`pricing.
`
`3 Today’s ISP
`
`Figure 3 depicts an ISP network. Subscriber traffic, generated over dial-up 28 Kbps or higher speed
`DSL and cable TV modems, is aggregated at the ISP’s point of presence or PoP. The ISP directs
`user datagrams over the Internet backbone through a Network Access Provider or NAP. Subscribers
`may connect to ISP servers that provide e-mail, news, web caching.
`
`Users are charged a monthly flat rate of $20 for 28 Kbps dial-up access, $50 - $200 for access over
`dedicated 128 Kbps–1.5 Mbps DSL lines, and $40 for shared access over cable TV. More than 90
`percent of residential subscribers use 28 Kbps dial-up modems, and in addition pay $15 per month
`for a telephone line. In case of cable TV, there also is a $25 monthly TV subscription charge.
`
`
`
`3 TODAY’SISP
`
`5
`
`ISP point of presence
`(PoP)
`
`backbone
`network
`
`telephone
`ISDN, DSL
`CATV
`access
`
`network access point
`(NAP)
`
`ISP servers
`
`Figure 3: An ISP aggregates subscriber traffic and directs it to a Network Access Provider or NAP.
`
`The ISP must purchase a modem pool, pay the telephone company for leased trunks to the modem
`pool, and the NAP for access to the Internet backbone. Assuming a monthly cost of $25 for one
`modem and one trunk, and a concentration of 10 subscribers per modem, this amounts to $2.5 per
`month per subscriber. (An ISP that charges $40 per month for access over cable in turn pays $30 to
`the cable operator.) Typical NAP charge is under 1 cent per MB of backbone traffic.
`
`So the subscriber’s cost of data transfer is high. The average amount of data transfer via 28 Kbps
`modems is 60 MB per month, so at a monthly cost of $20, the subscriber is paying 33 cents/MB. If
`you include the monthly $15 for a telephone line, the subscriber is paying 58 cents/MB. The light
`user who transfers a fifth of the average traffic pays an exorbitant $1.65–$2.91 per MB, and the
`heavy user who transfers fives times the average, pays 6.6–11.6 cents/MB. So light users subsidize
`heavy users. Data show that the 10 percent of heaviest users generate 30 times as much traffic as
`the 10 percent of the lightest users, so the subsidy is large.2
`
`Table 1 compares the distribution by protocol of subscriber traffic in one ISP and in INDEX. The
`first column is the rank of the protocol used by the ISP’s subscribers. The third column is the
`percentage of the traffic by protocol for the ISP, and the fourth column is for INDEX. The most
`obvious difference is that the top four applications (HTTP, NNTP, e-mail or POP3+SMTP, FTP)
`account for 96 percent of all ISP traffic, whereas INDEX users use a much wider set of protocols.
`The difference is explained by the fact that INDEX customers are more familiar with the Internet
`and they are more experienced computer users than the average Internet user interviewed in the
`national Nielsen-CommerceNet telephone survey. In a few years more Internet users will resemble
`today’s INDEX subjects.
`
`We now present the case against flat-rate pricing, beginning with the theoretical argument.
`
`2The average connect time of AOL’s subscribers is 21 hours per month, or $1 per hour. The median subscriber, whose
`connect time is much lower, is paying correspondingly more.
`
`
`
`4 WHAT’SWRONGWITHFLAT-RATEPRICING-THEORY
`
`6
`
`Rank Protocol
`1 HTTP
`2 NNTP
`3 POP3
`4 FTP
`5 SMTP
`6 HTTPS
`18 Telnet/SSH
`N/A X-Windows
`
`ISP % INDEX %
`76
`43
`12
`2
`3
`3
`3
`11
`2
`0.5
`2
`2
`0
`4
`N/A
`2
`
`Table 1: Protocols used in ISP vs INDEX
`
`p
`
`c
`
`D(p)
`
`p
`
`c
`
`waste
`
`xu
`
`xf
`
`usage
`
`xu
`
`xf
`
`Figure 4: At a unit price of p, a customer will consume D(p) = xu units; under a flat-rate charge
`she will consume xf . The shaded area is the waste.
`
`4 What’s wrong with flat-rate pricing - theory
`
`We analyze a model of consumer demand and show that flat-rate pricing (1) encourages waste and
`increases cost, (2) forces light users to subsidize heavy users, and (3) can introduce differentiated
`service quality only by inefficient segmentation in quality tiers. The next section presents empirical
`evidence supporting the model.
`
`Waste
`
`Suppose Internet access is sold at a price of p per unit of usage, measured in MB of data transferred
`or minutes of connect time. Then a user’s demand is modeled as a function D(p). This is the
`decreasing curve in Figure 4. The meaning of the demand curve is that each of the D(p) units of
`access is worth at least p to the user, and every additional unit is worth less than p.
`
`If access is sold at the ISP’s incremental cost of c per unit, the user will purchase xu = D(c) units.
`The value to her of consuming D(c) units minus her cost, cD(c), equals the area of the triangle
`above the horizontal line, p = c. This area is called the consumer surplus.
`
`
`
`4 WHAT’SWRONGWITHFLAT-RATEPRICING-THEORY
`
`7
`
`light user benefit
`
`average user benefit, B
`heavy user benefit
`
`c
`
`flat rate charge
`
`xf (low)
`
`xf (av)
`
`xf (high)
`
`Figure 5: At a unit cost of c, the flat-rate charge is the rectangle, the small triangle is the value to
`the light user and the large triangle is the value to the heavy user.
`
`But if the ISP charges a flat rate for unlimited use, the user’s cost of additional consumption is zero,
`and she will consume xf = D(0). The shaded area is waste: it is the amount by which the cost of
`R c
`providing xf − xu units, namely c(xf − xu), exceeds their value to the user,
`0 [D(p) − D(c)]dp.
`The consumer surplus is now lower: it is the large triangle (the value of consuming xf units) minus
`the cost cxf . The surplus may even become negative.
`
`Different users have different demand functions. Let D(p) be the average demand function, and
`suppose a particular user’s demand function is (cid:11)D(p). So (cid:11) < 1 for a light user, and (cid:11) > 1 for a
`heavy user. The waste is proportional to (cid:11): heavy users cause more waste.
`Thirdly, the waste will be larger if demand is more sensitive to price. If the price elasticity, p=D(p)(cid:2)
`j@D=@pj, is high, the demand curve will have a long “tail” and the shaded area will be larger.
`INDEX data indicate that the price elasticity for connect time is very high.
`
`User cross-subsidy
`If the flat-rate charge equals the cost of serving the average user this charge must be c (cid:2) xf (av) =
`c (cid:2) D(0). This is the rectangular area in Figure 5. Every subscriber pays this amount.
`The benefit or value to the average user of consuming D(0) is the area under her demand curve.
`This is the middle triangluar area. Suppose this area equals B. If the demand curve is (cid:11)D(p), the
`user’s benefit is (cid:11)B. So light users ((cid:11) < 1) subsidize heavy users ((cid:11) > 1). The larger is the spread
`of (cid:11), the greater is the subsidy. Data indicate that this spread between demands of heavy and light
`users is on the order of 30.
`The benefit to very light users, (cid:11)B, may be smaller than the flat rate charge c (cid:2) D(0), and so they
`will not subscribe. If the ISP wishes to retain these light subscribers it must either set the charge
`below average cost (and incur an operating loss) or it must restrict usage.
`ISPs engage in both
`
`
`
`4 WHAT’SWRONGWITHFLAT-RATEPRICING-THEORY
`
`8
`
`p
`
`c2c2
`
`c1
`
`D(p, q1)
`
`D(p, q2)
`
`flat rate (q1)
`
`xf (q1)
`
`xf (q2)
`
`flat rate (q2)
`
`Figure 6: Demand increases with quality. Light users will not subscribe to higher-quality service
`tiers.
`
`practices.
`
`Suppose the ISP charges a usage-based price c instead of a flat rate. Then usage will decrease, all
`except the heaviest users will pay less and enjoy greater net benefits, even light users will subscribe,
`and the ISP’s net revenue will increase.
`
`Tiered quality
`
`Suppose an ISP offers service of different quality, q, with larger q indicating better quality. (To fix
`ideas, think of quality as access speed.) The demand function D(p; q) will depend on q and it will
`increase with q. This is shown in Figure 6 where the higher quality demand D(p; q2) is to the right
`of D(p; q1).
`
`Suppose the ISP offers both service qualities at a flat rate. Then the service must be segmented into
`tiers: users can only subscribe to a single quality. The flat rates for the two qualities are the two
`rectangular areas. If a subscriber chooses the higher quality tier, she will consume xf (q2) units,
`her benefit will be the area of the large triangle and her cost will equal the large rectangle. If she
`chooses the lower quality tier, her benefit will be the small triangle and her cost will be the small
`rectangle. If these areas area as shown, this subscriber will choose the lower quality tier.
`
`However, if the ISP offered both qualities at unit prices c1, c2, as in the INDEX trial, the subscriber
`would consume both service qualities.
`
`
`
`5 WHAT’SWRONGWITHFLAT-RATEPRICING-EVIDENCE
`
`9
`
`Thus the introduction of quality tiers leads to two kinds of inefficiency. First, the high quality tier
`subscribers will tend to be the heavier users and the waste will be higher. The larger waste increases
`the high quality tier charge (the large rectangle). So light users will tend to exclude themselves from
`the high quality tier. The excluded users lose benefits and the ISP loses revenues. One consequence
`is that the market for higher speed access will be unnecessarily limited. Second, both types of
`customers lose because their options are limited to one or the other tier.
`
`INDEX data show that all users purchase high quality service sometimes if the cost is proportional
`to usage. ISP flat-rate charges for broadband access (384 Kpbs-1.5 Mbps) are about 5 to 10 times
`higher than for 28 Kbps. Such a large price differential is likely to discourage most users from
`subscribing to the high quality tier.
`
`5 What’s wrong with flat-rate pricing - evidence
`
`We now present findings from INDEX that support the theoretical arguments above.
`
`Responsiveness to prices
`
`In the first experiment (see Figure 1) users select their instantaneous speed and pay per minute of
`connect time. The following demand model is estimated:
`i; j 2 f16; 32; 64; 96; 128g;
`aj
`i log Pj;
`
`(1)
`
`X j
`
`log Xi = bi +
`
`where Xi is the number of minutes of connect time of speed i Kbps purchased by a consumer during
`a week when facing a price of Pj cents per minute for speed j. With this ‘log-log’ model, the
`
`
`coefficient aii is the own-price elasticity, i.e. aii is the percent change in demand Xi for speed i due
`to a one percent change in its price Pi; and aj
`i is the cross-price elasticity, i.e. the percentage change
`in the demand Xi due to a one percent change in the price Pj of speed j. The prior expectation
`i < 0, demand for speed i will drop if its price increases, and aj
`is that ai
`i > 0, demand for i will
`increase if the price of a substitute speed increases.
`
`The estimated demand equation is
`x128 = 4:1 −1:65p128 +0:44p96 +0:55p64 −0:12p32 +0:00p16
`x96 = 2:6 +1:23p128 −3:34p96 +1:17p64 +0:23p32 +0:00p16
`x64 = 2:7 +0:08p128 +0:84p96 −1:71p64 +0:47p32 +0:55p16
`x32 = 2:33 +0:48p128 −0:58p96 +0:88p64 −1:10p32 +0:08p16
`x16 = 0:52 +0:42p128 −0:26p96 +0:18p64 +0:97p32 −1:29p16
`where xi = log Xi, pi = log Pi. The estimated coefficients in (2) whose t-static (i.e. estimate
`divided by its standard error) is larger than 3 are written in bold.3 Observe that the demand for a
`particular speed is very sensitive to its own price and to the price of the next higher speed. The own-
`price elasticity is between -1 and -3 and the cross-price elasticity is around 1. The high own-price
`elasticity implies that when usage is connect-time, waste induced by flat-rate pricing is large.
`
`(2)
`
`3The demand estimates are due to Karyen Chu and Hal Varian.
`
`
`
`5 WHAT’SWRONGWITHFLAT-RATEPRICING-EVIDENCE
`
`10
`
`Second, the large cross-price elasticity indicates that if users are offered differentiated quality of
`service, they would indeed demand more than one service quality. Direct evidence of this in INDEX
`data also comes from the fact that on average, each customer selects 3.5 out of the 5 available priced
`speeds (in addition to the free 8 Kbps speed) each week. Since consumers do value multiple service
`qualities, there is a loss to providers and consumers when ISPs only offer tiered quality service as
`they do today.
`
`In experiment (3) users have two service choices: a free 8 Kbps and a traffic volume-priced 128
`Kbps service. Figure 7 shows the average weekly usage (in megabytes) of each service as function
`of the 128 Kbps price. This figure also shows sensitivity to price, although the elasticity is lower
`than the estimates (2) for experiment 1. The likely explanation for the lower price sensitivity is
`two-fold.
`
`In experiment 1, users are offered quality options that are close substitutes. In experiment 3 the only
`available substitute for 128 Kbps is 8 Kbps, and the very small increase in use of 8 Kbps despite
`a 30-fold increase in price of 128 Kbps, indicates that users don’t consider 8 Kbps an effective
`substitute.
`
`Second, under volume pricing, a user is better able to control her expenditure which depends only
`on the web or ftp transfer she executes. Under connect-time charges, her expenditure depends also
`on how much time she is connected to the network while she is thinking but not transferring data,
`and also on the actual rate of data transfer which depends on network congestion as well as on the
`(peak) rate she has selected. These less controllable aspects of connect-time charges are likely to
`lead to greater price sensitivity.
`
`Despite the smaller price elasticity, note that the average volume transferred is 30-40 percent less
`even under the smallest usage-sensitive charge compared with when the transfer is free (as in the
`flat-rate option).
`
`Quality-sensitivity of demand
`
`INDEX data show that the user’s demand increases strongly with quality as Figure 6 suggests. To
`explain the data, rewrite the demand equation (1) as
`
`Xi = ebi (cid:2) Y
`
`aj
`i ;
`
`Pj
`
`(3)
`
`j
`
`so that higher values of the constants bi corresponds to a multiplicative outward shift in the demand
`curve.
`
`From (2) the estimated constants are,
`
`eb128 = 60:3; eb96 = 13:5; eb64 = 14:9; eb32 = 10:3; eb0:52 = 1:7:
`
`Thus increasing the speed from 32 Kbps to 128 Kbps leads to a six-fold increase in average demand
`for connect time. The same order of magnitude is evident in Figure 7 if we compare the volume
`transferred at 8 Kbps versus 128 Kbps during the week when the latter service is free.4
`
`4Demand models have been proposed in which a user selects one of several different qualities to transfer an exoge-
`
`
`
`5 WHAT’SWRONGWITHFLAT-RATEPRICING-EVIDENCE
`
`11
`
`Figure 7: Variation in average weekly usage in MB of data transferred at 128 Kbps and at free 8
`Kbps as a function of price per MB at 128 Kbps.
`
`There seem to be two fundamental reasons for the high quality-sensitivity. One reason is that certain
`applications evoke a better subjective experience at higher speed than at lower speed, and so users
`demand for those applications will increase with access speed. Browsing of some web objects and
`interactive sessions (telnet, X-windows) are examples. A more subtle reason has to do with the fact
`that the full cost to a user is the sum of the money cost of access plus the value of her own time
`spent. So the total cost at the lower speed may well be greater than at the higher speed.5
`
`The large quality-sensitivity of demand implies that consumers and suppliers lose on two counts
`from the current ISP practice of tiered quality service. Since tiered service is flat-rated, and demand
`is quality-sensitive, the waste at higher speeds is greater, and the flat rate charge is correspondingly
`higher. Also, many light users will not subscribe to flat-rated higher quality service, even though
`they would occasionally subscribe if the charge were usage-based. So those users are denied the
`benefits of higher quality service, and producers are denied the revenues from those subscribers. In
`turn, the reduced market for broadband access will lower the pace of equipment cost reduction and
`the diffusion of quality-sensitive applications.
`
`Variation among users
`
`Figure 5 indicates that inter-user subsidy and waste under a flat rate charge increase with the varia-
`tion in demand among users. INDEX data show that the variation is indeed very large. To measure
`
`nously specified amount of data. The results here indicate that the amount of data transferred depends on the available
`quality, i.e the amount is endogenously determined.
`5Karyen Chu and Hal Varian have estimated INDEX users valuation of time.
`
`
`
`5 WHAT’SWRONGWITHFLAT-RATEPRICING-EVIDENCE
`
`12
`
`mean weekly expenditure ($)
`
`Figure 8: Variation in average weekly expenditure among users.
`
`Number of observations
`
`this variation, we re-estimate the connect-time demand equation (2) with a user-dependent constant
`term,
`i; j 2 f16; 32; 64; 96; 128g;
`
`(4)
`
`aj
`i log Pj;
`
`X j
`
`
`
`log X ki = bki +
`
`
`
`where X ki is the consumption of user k, and bki is the constant that depends on k. So user demands
`
`
`
`can differ by a multiplicative constant.
`The R2 of the estimated equation (2) ranges between 0.15 and 0.30, and the R2 of the estimates
`of (4) ranges between 0.97 and 0.98. Moreover the estimates of the coefficients aj
`i do not change
`significantly.6 The results imply a very strong user-specific multiplicative component in the demand.
`The estimated coefficients bk
`i differ among the heaviest and lightest users by 3 or 4, so that their
`demands differ by a factor of e3 = 20 to e4 = 54.
`
`The difference in demand is also indicated in Figure 8 which shows that the heaviest users spend 15
`to 20 times more than the lightest users.
`
`We can estimate the extent of inter-user subsidy.7 We evaluate the distribution of actual expenditures
`among users over the six weeks of the volume-priced experiment 3. We compare it with what that
`distribution would have been for a flat rate tariff, assuming that user traffic is unchanged and the
`tariff is set to recover the same revenues. Figure 9 is obtained as follows. We rank users by their
`traffic volume, highest users first. For each n we plot the cumulative expenditures of the first n users
`versus their cumulative traffic volume. This gives rise to the upper ‘dots’ in the figure. The lower
`
`6This is expected since prices are varied randomly.
`7These results are due to J¨orn Altmann.
`
`
`
`6 WHYISPSSHUNUSAGE-BASEDPRICING
`
`13
`
`Figure 9: Expenditures versus usage under usage-based and flat rate pricing.
`
`dots give the plot for an ‘equivalent’ flat rate tariff.
`
`Thus, for example, the three heaviest users account for about 3 GB of data and 35 percent of total
`expenditures in the actual experiment. Under a flat rate tarrif, however, these three users would only
`account for 3/60 or 5 percent of the total expenditures, represented by the lower dot.
`
`The volume-based pricing scheme is clearly more fair since users pay almost in proportion to their
`usage.8 Second, the heaviest 30 percent of users would pay less under the flat rate scheme, whereas
`the remaining 70 percent would pay more. In fact, this comparison understates the subsidy involved
`under a flat rate because it assumes that users will generate the same amount of traffic under a flat
`rate system. But we know that in this experiment they generate twice as much traffic (see Figure 7),
`and the additional traffic is of less value. So the percentage of users who will be worse off under a
`flat rate tariff is likely to be 80 percent or more.
`
`6 Why ISPs shun usage-based pricing
`
`In December 1996, AOL changed from usage-based pricing ($9.95 per month, including 5 hours of
`connect time, plus $2.95 for each additional hour) to a $19.95 per month flat rate (later increased to
`$21.95). As a result, average monthly connect time per subscriber jumped from 6.4 hours to 22.1
`hours in 1998, revenue per hour of connect time declined, as did the company’s operating margin
`([2]).
`
`8The upper dots do not lie strictly along a straight line because different users face different, randomly chosen prices
`and because the use of the free 8 Kbps service varies across users.
`
`
`
`7 ANALTERNATIVEISP
`
`14
`
`The shift to flat-rate pricing implies two fundamental changes in an ISP’s business. First, it creates
`an incentive for the ISP to passively or actively degrade service quality, since per subscriber usage
`and cost decrease with worse quality but revenue remains the same. Passive degradation is acheived,
`for example, by limiting the size (increasing blocking probability) and speed of the modem pool.
`Active degradation is acheived (especially at higher speeds) by placing contractual and administra-
`tive limits on the ways customers’ can use network access. For example, ISPs that provide access
`over cable modems limit streaming and access speeds. Other common practices are to prohibit
`users from installing a web server or a LAN in their homes, and to deny permanent IP addresses.
`(Providers also limit speed in order to segment the market into tiers.)
`
`The only incentive to limit service degradation is the threat of loss of subscribers to other ISPs. ISPs
`reduce this threat by increasing the cost of switching to other providers. For example, in order to
`switch, a subscriber would have to reconfigure her computer which she may find difficult to do, and
`her e-mail would not be forwarded. ISPs incur significant costs to recruit subscribers. Typically,
`these costs are promotional free trials, and free services in the form of news clips, bulletin boards.
`
`Thus competitive, flat-rate pricing increases per subscriber recruit