The Internet Singularity, Delayed: Why Limits in Internet Capacity Will Stifle Innovation on the Web

Executive Summary

In this research study, Nemertes performed an independent in-depth analysis of Internet and IP infrastructure (which we call capacity) and current and projected traffic (which we call demand) with the goal of understanding how each has changed over time, and determining if there will ever be a point at which demand exceeds capacity.

To assess infrastructure capacity, we reviewed details of carrier expenditures and vendor revenues, and compared these against market research studies. To compute demand, we took a unique approach: Instead of modeling user behavior based on measuring the application portfolios that users had currently deployed, and projecting deployment of those applications in future, we looked directly at how user consumption of available bandwidth has changed over time.

Read the study below, then read the FAQ!

Participate in the ongoing discussion at the bottom of this page. We want your feedback. Comments are only moderated for tone (vulgarity, ad-hominem), not content.

More media:

Radio Interview on AM900 CHMLAM

Radio Interview on AM770 CHQR (requires registration - audio slot Tue 27 Nov, 10am, forward to 34 minutes in the audio)

3 Overall Framework: Demand, Infrastructure, and Investment

4 Modeling User Demand

4.1 The Application-Centric Demand Model

4.2 The Innovation-Centric Demand Model

4.3 Moore's Law for Internet Applications?

4.4 Behavior and Biological Curves

4.5 The Emerging Virtual Generation

5 Modeling Supply

5.1 Optical

5.1.1 Optical Methodology

5.1.2 North America

5.2 Switching and Routing

5.2.1 Protocols and Layers

5.2.2 Methodology

5.2.3 Core

5.2.4 Connectivity

5.2.5 One Box Two Trunks

5.2.6 Access

5.2.7 Wireless: Building Footpaths Across the Digital Divide

6 Investment

6.1 Methodology for Determining Investment

7 Key Findings: The Coming Bandwidth Crunch

7.1 Global and NA Supply and Demand Curves

7.1.1 Demand vs. Supply Overall

7.1.2 When Is Supply Not Really Supply?

7.2 Investment Gap: What It Takes To Prevent The Crunch

7.3 Sensitivity Analysis

7.4 ComparisonWith Other Studies

8 Does the Internet Ever Break?

8.1 Access Circuit Saturation

8.2 Router Issues

8.2.1 Router Congestion

8.2.2 Addressing and Route Table Expansion

9 Conclusions and Recommendations

10 Appendix A: Detailed Methodology

10.1 Detailed Demand Methodology

11 Bibliography and Sources

11.1 Sources

11.2 Bibliography

Table of Figures

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 1: Nemertes Internet Model Influence Diagram............................. 9

Figure 2: Global Internet Capacity Curves............... 10

Figure 3: North American Demand........................... 16

Figure 4: Global Optical Revenues............................. 21

Figure 5: Incremental Global Optical Capacity........ 22

Figure 6: Projected Global Incremental Optical Investment................... 23

Figure 7: Global Incremental Optical Investment.. 23

Figure 8: Global Optical Capacity............................... 24

Figure 9: North American Optical Capacity.............. 25

Figure 10: North American Switching Capacity...... 30

Figure 11: Growth In Global Switching Capacity..... 31

Figure 12: North American Broadband Access Lines................................. 32

Figure 13: Projected North American Access Lines................................... 33

Figure 14: Total North American Access Lines....... 34

Figure 15: Global Access Lines....................................34

Figure 16: North American Access Capacity............. 35

Figure 17: World Access Capacity............................... 35

Figure 18: Global Internet Infrastructure Investment............................. 38

Figure 19: Global Investment In Infrastructure..... 38

Figure 20: North American Investment In Infrastructure....................... 39

Figure 21: North American Investment In Infrastructure....................... 40

Figure 22: North American Capacity Versus Demand................................ 42

Figure 23: Global Capacity Versus Demand............. 43

Figure 24: North American Demand Compared To Access Limits.......... 44

Figure 25: Utilization Sensitivity............................... 46

Figure 26: Measured Vs. Modeled Data................. 48

Figure 27: Internet User Population.......................... 55

Figure 28: Total Internet Capable Devices............... 56

Figure 29: Internet Capable Devices by Region....... 56

Figure 30: Total North American Demand.............. 57

Figure 31: North American Nominal Demand........ 58

Acknowledgements

Nemertes would like to thank the many outside researchers who have generously discussed their insight and findings with us or reviewed our approach, models, or data, including:

-- Noel Chiappa, Internet researcher and an inventor of the multiprotocol router (1981)

-- kc claffy, PhD, principal investigator for the distributed Cooperative Association for Internet Data Analysis (CAIDA), and resident research scientist based at the University of California's San Diego Supercomputer Center.

-- Jeffrey Cole, PhD, director and Michael Suman at The Center for the Digital Future at The USC Annenberg School

-- Andrew Odlyzko, PhD, Director, The Digital Technology Center at The University of Minnesota.

-- Arielle Summits, Wilson Craig and the rest of the developers of the Cisco Global IP Traffic Forecast and Methodology, 2006-2011, Cisco Systems Inc.

-- The 70 subject-matter experts at major enterprise organizations, service providers, equipment vendors, venture capital and financial firms, and content providers who shared details of their companies’ traffic patterns, revenue breakdowns, and future investment strategies in confidence with us.

Your assistance has been invaluable. Thank you.

In addition to the above, we crosschecked our research against statistics provided by the Pew Internet Research and Internet Gatekeepers Inc. as well as against statistics generated by market research firms and provided as part of other research projects (see the bibliography for details). We also drew upon five years of results from Nemertes’ proprietary benchmark studies that involved the personal participation of IT executives at roughly 500 enterprise organizations. We are grateful for your collective insight and wisdom.

Please note that providing us with research or reviewing our findings does not in any way imply agreement with our findings or confer responsibility for any errors. Our findings, and any errors, are solely our own.

Finally, the authors of this report would like to thank our fellow Nemerteans and our families for their insight, support, and patience during this project.

2 Executive Summary

In this research study, Nemertes performed an in-depth analysis of Internet and IP infrastructure (which we call capacity) and current and projected traffic (which we call demand) with the goal of understanding how each has changed over time, and determining if there will ever be a point at which demand exceeds capacity.

In other words, we assumed that users had consumed, or would consume, a certain amount of bandwidth, and that the rate of change of that bandwidth consumption was the metric that mattered, rather than the specific portfolio of applications. This is similar to the way that Moore’s Law focused on the rate of improvement of processing power, rather than the specific portfolio of technology innovations that enabled that rate to occur. We call this the innovation-centric demand model, and we believe it’s the only reliable way of predicting the unpredictable” obtaining reasonably accurate projections about future scenarios without needing to know the specific innovations that made it possible.

We then validated our findings with the best available data. To gather that data we consulted academic and industry research organizations and conducted primary interviews with more than 70 enterprises, vendors, service providers and investment companies, as well as drawing on our established base of five years of several hundred benchmarked enterprise organizations.

This resulted in the first-ever study that assessed both infrastructure investment and current/projected traffic patterns independently, and compared the two. It is also the first study to apply Moore’s Law (or something very like it) to the pace of application innovation on the “Net” and validate that it appears to conform to the available data so far.

Our findings indicate that although core fiber and switching/routing resources will scale nicely to support virtually any conceivable user demand, Internet access infrastructure, specifically in North America, will likely cease to be adequate for supporting demand within the next three to five years. We estimate the financial investment required by access providers to bridge the gap between demand and capacity ranges from $42 billion to $55 billion, or roughly 60%-70% more than service providers currently plan to invest.

It’s important to stress that failing to make that investment will not cause the Internet to collapse. Instead, the primary impact of the lack of investment will be to throttle innovation” both the technical innovation that leads to increasingly newer and better applications, and the business innovation that relies on those technical innovations and applications to generate value. The next Google, YouTube, or Amazon might not arise, not because of a lack of demand, but due to an inability to fulfill that demand. Rather like osteoporosis, the underinvestment in infrastructure will painlessly and invisibly leach competitiveness out of the economy.

One could even whimsically speculate” as we did in the title--that the lack of investment could be holding back the time at which the Internet reaches a singularity (a point at which accelerating change creates an unpredictable outcome, such as the Internet becoming independently sentient).

More seriously, we did not set out with an agenda, or to prove or disprove a particular point (singularity or otherwise). We modeled capacity and demand using the best tools at our disposal and validated the findings as fully as possible against the best available data. The result is, as results of all models necessarily are, a projection, the accuracy of which we can only improve with better data and more refined modeling techniques, and we welcome suggestions on the latter and access to the former.

We did come away with the overwhelming conviction that there is a deep industry need for better (more comprehensive and accurate) data in this area. The Internet is almost opaque to serious researchers, even those with the necessary technical skills, integrity and desire, for the simple reason that carriers and content providers refuse to reveal their inner workings. The reasons for this are good” service and content providers are reluctant to reveal their proprietary competitive advantages, or accidentally breach their customers’ privacy” but the need for open, honest, and comprehensive information exchange is acute. So we conclude by urging content and service providers to cooperate with researchers in sharing data.

3 Overall Framework: Demand, Infrastructure, and Investment

Internet modeling exercises have typically concentrated on demand and neglected capacity (or vice versa). Moreover, traffic demand and growth measurements almost inevitably assess traffic that’s already on the network, for the very logical reason that it’s much easier to capture aggregated traffic statistics from devices in the core, rather than monitoring devices at the edges attempting to inject traffic into the network to see how well they succeed. (One is reminded of the old joke about the drunk looking for his keys under the streetlight, not because that’s where he left them but because it’s where the light is.)

While this approach can provide useful predictions of the ways in which loads evolve over time, it does not really give any insight into the ways in which lack of capacity degrades service, or actively limits demand. Moreover, it provides virtually no insight into how demand is generated” thus missing one of the most critical pieces of the demand/capacity equation. Since one of the hypotheses Nemertes wanted to explore was the possibility that capacity might be limiting demand either at present or in the future, we needed an approach that allowed us to do so.

We therefore modeled demand and capacity independently. This independence is important because it allows us to decouple the impact of capacity on demand. If they are not decoupled, the model will fail to capture a scenario in which capacity is limiting demand. Researchers may be able to note that demand has slowed, but won’t be able to tell if this is a true measure of actual demand (users conveniently want no more capacity than happens to be available to them) or a case in which capacity is in fact the limiting factor. Our model was designed to capture reality as exactly as possible, given the rather considerable constraints inherent in a situation in which there is limited reliable data, and future projections are heavily dependent on the emergence of unknown applications and capabilities.

Nemertes approached its analysis by considering the Internet as an exercise in supply (expressed by infrastructure) and demand (expressed in terms of the Internet usage patterns). The first step in doing this was to build an influence model that would serve to direct our data gathering (Please see Figure 1: Nemertes Internet Model Influence Diagram).

The demand side of the model is driven by users, which we segregated by geography. For the purposes of this analysis, we grouped them by North America, Europe, Latin America, Asia Pacific and Africa Middle East. Each of these groups has differing levels of access to a series of Internet-attached devices, each of which run a range of applications. The degree to which a device can generate load is proportional to the amount of time the user desires to do so as well as the degree to which the device in question is physically capable of pushing data. Each of these applications, then, drives data through the five geographic areas and ultimately generates a load that is felt by the Internet as a whole.

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 1: Nemertes Internet Model Influence Diagram

It bears noting that Nemertes did not try to model a typical user for each geography. However, the approach selected allowed us to model a maximal demand profile for each region , which sets the upper bound for demand, without having to survey thousands of users. In telephony terms, Nemertes’ approach focused on the absolute busy hour characteristics of a user group rather than trying for a comprehensive, user demand profile over time.

The capacity side posed its own problems. Most Internet modeling ignores the supply side entirely, or simplifies it considerably, not without reason. True capacity is defined as the maximum throughput measured over some time period. It turns out that this is a complex undertaking when the thing whose throughput you are measuring is characterized by billions of nodes with billions of potential paths from one node to another node. The degree of complexity is compounded when you realize that many different routing protocols are being used such that determining the path that may be used is somewhat indeterminate.

Virtually all the available research literature that attempted to model such a problem was concerned with deriving an algorithm that could be used, rather than one that worked in a practical sense for sizing the Internet. As a consequence, in every case that we examined, the algorithm was far too complex to actually solve. Instead, we opted for a more simplistic approach that fundamentally treats the Internet as a series of containers for holding bits. This approach allows us to simply count up the devices that generate bits, multiply them by their maximal bit rate and then add up the capacities to obtain the total capacity.

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 2: Global Internet Capacity Curves

The model approaches demand from the perspective of several domains: core switching, optical backbone, access lines and connectivity switching (Please see Figure 2: Global Internet Capacity Curves). With the exception of access lines, which are physical and wireless transmission media, the balance are represented by investment in electronics. In each case, we researched the annual capital equipment expenditure, translated that investment into capacity, and computed a total incremental investment/capacity. These figures, then, became the discrete capacities against which we compared demand.

We then plotted each connection modality against the respective geographical regions, and summed the resulting capacities to arrive at the total supply. We plotted both supply and demand as petabytes per month. Although a petabyte per month is a misleading measurement, since it tends to mask things like peak usage and imply a continuous use dynamic, we nevertheless adopted it since most of the literature has settled on this convention.

It turns out that taking this approach is not a bad match to the approach we used for assessing demand. Since we approached both from a maximal perspective, any difficulties in servicing demand would be expressed at the margin and allow easy testing of assumptions and sensitivity. We reasoned that if the model indicated a problem developing in a situation of maximal demand serviced by maximal capacity, it would be a simple thing to back off demand to see at what point the demand could be serviced, then compare the findings to reality.

4 Modeling User Demand

To model user demand, we first looked at the maximum possible demand, which essentially measures how much data users could hypothetically generate on their Internet-attached devices, given the following:

a) the types of Internet-connected devices (home PCs, work PCs, mobile Internet devices, interactive gaming consoles, and IP-enabled TV)

b) the port speeds of those devices (e.g. the maximum possible data rate of the network interfaces)

c) the number of devices.

To create a maximum possible demand curve, we assessed the number of each type of device available to each user for the years 2000 through 2012 based on publicly available data and projections. (Please see Appendix A: Methodology, for further details). We found that, counting all devices and all port interfaces, an Internet-attached North American user in 2007 is theoretically capable of generating approximately 61 Mbit/s of traffic, equating to 20 Petabytes of traffic per month.

It’s important to note that by this measure, not only does maximum possible demand exceed Internet capacity today, but it has always exceeded capacity and very likely always will, because of the fact that port speeds on Internet-attached devices tend to be within an order of magnitude or so of Internet circuits themselves (and there are far more devices than circuits).

For example, back in the 1990s when the Internet backbone consisted of T1 (1.5 Mbit/s) and T3 (45 Mbit/s) circuits, the LAN speed of the typical Internet-connected host was 10 Mbit/s (an order of magnitude larger than the backbone). Today, a typical Internet-connected PC has a 100 Mbit/s or Gbit/s Ethernet link, and the Internet backbone circuits are generally OC-768 (38 Gbit/s).

As Andrew Odlyzko and K.G. Coffman noted back in 2001, For the foreseeable future a handful of workstations will in principle be able to saturate any given Internet link. A few thousand machines will continue to be capable of saturating the entire Internet. (Coffman and Odlyzko, 2001).

Moreover, access circuits (the last mile typically connecting Internet-attached devices to the Internet) tend to run at speeds several orders of magnitude lower than the devices they serve. A typical DSL circuit, for example, delivers a maximum of 1.5 Mbit/s to one or more Internet-attached devices in each household (and each device, keep in mind, has a port speed of 100 Mbit/s or up). This means that the Internet-attached devices in a typical household are fully capable of saturating the household’s Internet connection; the same holds true for business sites.

However, simply because the devices can saturate the “Net doesn’t mean they will. In practice, network-attached devices very rarely (in fact, almost never) generate traffic at 100% of port capacity for a sustained period of time. The actual port utilization depends heavily on the machine’s CPU, the types of applications running on the machine and the habits of the user.

So, the second step in constructing a demand model lies in determining how much of that hypothetical capacity is being used. To gain this perspective, we consider the fact that applications are running on devices and, more importantly, people are operating the devices. This will not always be the case; as many researchers have observed, machine-to-machine traffic will begin to become significant over time. However, in the window of this study (2000 â€“ 2012) we still foresee the majority of Internet traffic being associated with people operating devices and driving applications.

4.1 The Application-Centric Demand Model

There are fundamentally two approaches to constructing an Internet traffic demand model. The first approach is to try to create one or more typical user profiles that describe which applications each user is running, for how long, and how much bandwidth a typical application consumes as it’s being run by a typical user. (This is more difficult than it sounds, because application usage can often be highly idiosyncratic, particularly for consumer applications: Some people change channels 50 times in an hour and download many more movies than they can possibly watch; others watch nothing but their daily half-hour news shows.)

Once the researcher creates user profiles, he or she then attempts to measure historical usage patterns for the application to arrive at total traffic volume. Finally, the researcher projects current trends into the future on an application-by-application basis to estimate how traffic volumes might evolve.

This approach has a couple of significant advantages, including granularity (researchers can distinguish among different application types), defensibility, and comprehensiveness (if the list of applications is comprehensive, researchers can feel relatively certain they’ve accurately captured the majority of the traffic that’s out there, assuming the profile modeling is correct).

Most likely for these reasons, this approach is favored by several consequential researchers in this space, most recently an internal research team at Cisco Systems Inc., which deployed this approach to model consumer application deployment in (Cisco, Global IP Traffic Forecast and Methodology, 2006-2011, 2007). The Cisco model assessed all known consumer applications generating significant amounts of Internet and IP traffic in the summer of 2007, and extrapolated these out through 2011. (Cisco also looked at enterprise applications, but with less granularity, which is reasonable as these tend to be more difficult to classify meaningfully into categories).

There is one significant problem with this approach, which is that by definition, it can’t easily account for innovation. In particular, it does a poor job of meaningfully predicting sudden shifts in user behavior, or the emergence and rapid adoption of new applications. As J. Licklider has stated: People tend to overestimate what can be done in one year and to underestimate what can be done in five or 10 years.( [Licklider] J. C. R. Licklider, Libraries of the Future, MIT Press, 1965.)

This is because in order to factor these in to such a model, researchers have to have some idea in advance of what these shifts and applications might be. This essentially forces researchers to become futurists. Yet predicting the future is notoriously difficult, even for those whose livelihoods depend on doing so, and the difficulty goes up dramatically with the specificity of the required prediction (it’s easier to correctly predict that the economy will go up sometime within the next five years than that a specific technology will emerge in, say, autumn 2010).

Moreover, getting into the game of predicting specific applications weakens the predictive effectiveness of the model, because if an application doesn’t emerge precisely as planned (something that’s highly likely) the model will be less accurate. An application-centric approach wouldn’t have predicted the rise of YouTube, for example, which emerged in 2005, and which Cisco says was already responsible for roughly 27 Petabytes/month in 2006” about as much traffic as traveled on the Internet in total in the year 2000.

Skeptics may point out that as impressive as that sounds, 27 Petabytes/month represents a relatively small percentage of current Internet traffic, and that’s certainly true. As we’ll discuss further in Section 7, however, the stage is already set for technologies to emerge as rapidly as YouTube (well within the window of our study horizon, in fact) which may have a significant effect on overall Internet traffic.

4.2 The Innovation-Centric Demand Model

The second approach is to assume, a priori, that innovation occurs. That is, user pattern shifts and new applications will emerge, and the rate and impact of these new behaviors and applications can be both modeled historically and projected into the future without knowing the specific details of these changes in advance. This approach is particularly useful when modeling technical innovation” it allows researchers to predict that a certain event will occur without requiring them to predict precisely how this will happen.

The well-known Moore’s Law is an example of this type of predictive model: Back in 1965, Intel co-founder Gordon Moore observed that the number of transistors that can be inexpensively placed on an integrated circuit is increasing exponentially, doubling approximately every two years. The trend has continued for more than half a century and is not expected to stop for at least the next decade.

The significance of Moore’s Law is that Moore did not have to describe precisely which technical innovation would lead to this doubling between, say 1988 and 1990” all he had to do was validate that such innovation was in fact occurring at an exponential rate. So long as the law continued to apply, he could be confident the innovation would occur.

The strength of an innovation-centric demand model is that it’s the best available mechanism for modeling the unknown. The main weakness (other than the loss of a certain amount of granularity) is that it applies only under conditions of continuous innovation, and it’s difficult to know in advance what those conditions are, or at what point they will cease to occur.

As has been noted extensively elsewhere (Kurzweil, 2006) innovation-centric models tend to result in exponential increases, and in the real world exponential increases tend to be halted by external environmental forces (otherwise the world would be knee-deep in rabbits, which like many other animals, reproduce exponentially).

The weakness of an innovation-centric approach is that it might not actually apply, which makes validation against measured data extremely important for models of this type. (Unfortunately, as we note later on, through no fault of the researchers, credible measured data is impossible to come by in the area of Internet traffic measurement. This is one of the key issues that researchers must address to enable the industry to make informed decisions).

The challenge in selecting the right model for the demand curve, then, comes down to choosing between the application-centric and innovation-centric approaches. Again, application-centric models, if well-executed and validated, do a fine job capturing the current state and enabling near-term projections of that state. But they do a poor job at projecting future states that could be highly dependent on innovation. We believe that this fundamental weakness of the application-centric approach precluded its use here.

4.3 Moore’s Law for Internet Applications?

For these reasons, the Nemertes model deployed instead takes an innovation-centric approach to projecting user demand. Essentially, we applied a Moore’s Law-like approach not to Internet traffic volumes, but to the applications and devices that generate such traffic.

As noted, we created a user profile that included an aggregate set of Internet-attached devices (each with an associated port speed). The devices available to each user, as well as the port speeds associated with each device, can be reliably validated from 2000 to 2006 (based on market figures for availability of such devices), and projected for the years 2007-2012. It’s worth noting that both the number of Internet-attached devices, and the speeds of those devices, associated with each Internet-connected user increase dramatically over time (see Methodology for details).

Then we assessed the limited data available on current utilization of users accessing the Internet through unlimited-bandwidth pipes. We converted those utilization figures to traffic volumes, and as will be discussed shortly, computed those traffic volumes as a percentage of maximum theoretical throughput.

We then developed a growth rates for these utilization figures, to account for the fact that measured utilization is unlikely to remain constant for the 12-year period under study. In fact, as will be discussed, we have good reason to think that it will rise dramatically. We applied these growth rates retroactively to estimate the volume of Internet traffic for the years 2000-2007. Finally, we projected forward to estimate the amount of user demand for the years 2007 through 2012.

The utilization rate that we computed is based on measurements and assertions that we believe are close to being accurate. In 2001, an evaluation of the Internet conducted by Coffman and Odyzko concluded that the Internet traffic could not be more than 85 Petabytes per month. We also had assertions by the carriers that since 2000 they have seen growth in demand that approached 100% a year.

An additional datum is that several researchers reached independent estimates of Internet usage for 2006 in the 2000 Petabyte-per-month range. If the demand increases that the carriers saw and the measurement seen in 200o and 2006 were correct, the implication was that demand was really growing at rates approaching and in some cases exceeding 100% year over year.

Growths of this magnitude translate into utilization rates that, in 2006 stand at .066 % or about 350 Megabytes per day of usage. This seems imminently reasonable. It also agrees with what one would expect of an average user of the Internet. This is equivalent to downloading about an hour of Internet video, or multiple hours of working, emailing, talking, sharing, uploading, downloading and watching video” often at the same time. Based on the most recent figures from the Annenberg Center for The Digital Future, the typical Internet user spends roughly four hours per day actively using the Internet (as opposed to being merely online), spread across multiple devices, and the majority of that time is spent heavily multitasking, often running streaming multimedia applications in the background while focusing on less bandwidth-intensive applications in the foreground.

As a final step, we projected demand curves into the future based on estimates of utilization and growth that show the influences of Moore’s Law behavior as described above, resulting in the following demand chart for North American traffic (Please see Figure 3: North American Demand).

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 3: North American Demand

4.4 Behavior and Biological Curves

At first blush, this projection might seem a bit eyebrow-raising. Two hundred Exabytes (200,000 Petabytes) after all, equates to an individual user consuming or generating 26 Gbytes/day by 2012. Even spread across multiple devices, this would seem to be extreme” it equates to each and every user deploying nearly seven hours of high-definition interactive video per day.

And before we explain how we arrived at such a projection, and why we believe it’s reasonable, it’s worth stating clearly: We are not predicting that this level of traffic will occur.

In fact, we are virtually certain that it will not, because it can’t. As we’ll discuss at greater length in Section 7, even if such a demand existed, it could not be satisfied, because the infrastructure to support such a demand is lacking.

However, we believe that the demand could occur, and in the absence of constraints imposed by infrastructure, that it would be at least likely, if not guaranteed. Here’s why: As noted, an innovation-centric model is the best possible way to model innovations that can’t be precisely predicted.

The challenge lies in two areas: First, determining the appropriate measures that are increasing (or possibly increasing) exponentially. Second, validating that the exponential innovation model actually applies at a particular point in time (as opposed to somewhere off in the future).

Regarding the first, as noted the measure we settled on was utilization, specifically utilization of the local bandwidth (e.g port speeds) available to the user. This is distinct from utilization of WAN access circuits, which is something we will discuss shortly. Our key finding is that utilization of available port speeds appears to have been increasing exponentially over time (and is in fact responsible for a considerable part of the high traffic growth measured by Odlzko and others in the earlier part of the decade). Given that our analysis matches the best-available data, we can be reasonably sure that it has been accurate at least thus far.

In short, users have been demonstrably using more and more of the capacity that’s available to them” and we believe this is a growth rate that will continue. Coupled with the fact that there’s dramatically more bandwidth available to them going forward (again, talking in terms of port speeds, not necessarily access), this results in the startling growth projections reflected here.

That brings us to the second point: Why do we believe the utilization growth rate will continue, barring external constraints? Why wouldn’t it slow down or even decrease?

This is simply another way of asking ourselves, Does the exponential innovation model apply at present? We believe it does for the following reasons. First, the assumption matches the best available data so far (with a caveat discussed in Section 7). As noted earlier, agreement with measured data is one of the key validations required for this type of model.

4.5 The Emerging Virtual Generation

Second, we believe behavioral conditions favor it. It’s true that there’s often a significant time lag between the point in time when an innovation is introduced and when it reaches widescale deployment. But as has been discussed copiously elsewhere, the millennial generation (18 and under) that is coming of age is comprised of people with demonstrated capacity to adopt networked applications in widespread fashion in record time” 12-18 months or less. (Examples include YouTube, Facebook, etc.) Based on the best available data, including that from the Center for Digital Life, Internet users today, and younger users in particular, are:

- Multitasking (2/3 say they run more than one application at one time)

- Running IP-enabled devices at once (in other words, it’s not either/or it’s both/and)

- Switching preferentially from non-IP applications to IP-based versions of the same applications (for example, listening to radio streams on the “Net rather than via satellite or radio, and watching television shows on the “Net instead of on TV).

All these trends drive up utilization, and all are accelerating.

Third” and quite importantly” we believe that a key cluster of technologies have matured to a point that enables rapid, low-cost development of a broad range of applications. These technologies include Web 2.0 programming and development tools, HD displays, low-cost cameras and recorders, and data storage. According to a recently published IDC/EMC study, the amount of data globally in 2006 was some 161 Exabytes, proceeding to 988 by 2010, with much of the data-generating applications existing on Internet-attached devices (phones, PCs). (IDC, 2006).

Although we’re deliberately trying to avoid predicting specific applications, it’s worth considering the following: MIT researchers have already developed devices that allow users to set up a variety of always on video connections between friends and family members. These can displayed on wall-mounted high-definition displays, or carried as low-definition keychain-sized trinkets (with wireless access to the “Net). Either way, they enable users to set up sustained interactive connections between remote parties.

Why would anyone do this? Consider the following trends:

- Increasing oil prices, raising the cost of travel

- Aging baby boomers caring for elderly parents, with college-age children

- A far-flung, highly distributed population.

A wall-mounted HD display with high-resolution camera and an easy-to-use interface could serve as a virtual window between families and friends. (Users could also use the system to download a changing display of photos and artwork from the “Net, as a high-end restaurant in Dallas already does).

Such a virtual window would enable adults to keep better track of their aging parents and far-off children, and maintain long-distance friendships. This is something that early adopters already do (though typically via low-res Webcam connections from hotel rooms while traveling).

But what really moves this concept from nice-to-have to need-to-have is the rising cost of energy, which is already beginning to limit extraneous travel. A steep increase in oil prices (something that’s certainly not unthinkable) could drive widespread adoption of such an application within the next five years in a tipping point effect. Nemertes is already seeing similar trends in the business world, with over 80% of enterprises defining themselves as virtual workplaces and nearly 40% planning to use video over IP within the next 12 months, with travel avoidance cited as a key driver (Source: Nemertes Benchmarks).

And again, deployment of this sort of virtual window technology deployment is possible well within the timeframe of this study. As noted earlier, the key technologies” low-cost, high-definition displays, inexpensive cameras, copious storage” already exist. Within this context, the idea that the typical user will be running hours of HD interactive video by 2012 suddenly doesn’t seem so farfetched.

As noted, we aren’t in the business of predicting particular applications, and we don’t mean to imply that just because this particular application could happen, it necessarily will. The point, again, is that technologies enabling this type of application, and many others, already coexist with a “Net-savvy generation predisposed to use them and an economic environment that increasingly favors their use.

In sum, we believe that the environment necessary for a Moore’s-law increase in application utilization exists today. Or as Vint Cerf put it recently to the Washington Post: Once you have very high speeds, I guarantee that people will figure out things to do with [them] that they haven't done before. (Cerf 2007). Specific details on how we constructed our demand model are described in the Methodology. But to see how the model plays out in the context of the current capacity environment, we have to first examine infrastructure capacity.

5 Modeling Supply

5.1 Optical

Internet capacity, at the highest level, is defined by how much traffic the optical backbone can carry. A significant part of the Nemertes model is devoted to assessing the capacity of this optical infrastructure.

The optical backbone that makes up the long haul capability and, to an increasing extent the distribution to the user, is composed of optical fibers driven by opto-electronics. These fiber runs, for all practical purposes, have inexhaustible bandwidth. The limit to what can be carried is set by the opto-electronics.

For this reason, determining the carrying capacity of the fiber is less an exercise in counting fibers, than it is in inventorying the optical hardware that drives them then multiplying by the speed at which they can operate.

5.1.1 Optical Methodology

While inventorying hardware in the Internet is not really practical, from a direct examination standpoint, it is nevertheless possible to review the amount of equipment being deployed into the network annually and make an informed assumption as to the degree to which that investment translates into optical capacity.

Various financial filings by optical electronic vendors were examined to determine the annual revenue of the largest optical equipment manufacturers. (Please see Figure 4: Global Optical Revenues), in the year 2006, the total revenue for these manufacturers worldwide amounted to more than $6.3 billion.

While the revenue of the manufacturers is not precisely the same as the investment in the network, it is still true that the investment would necessarily be tightly correlated with the revenue of the manufacturers. The missing piece would be the amount of vendor revenue that is derived from non-equipment sales. This varies by vendor, but tends to be de minimis, when compared to the revenue generated from the sales themselves. For the purposes of this study, Nemertes assumed that, for all practical purposes, they are the same.

The missing portion of the equation, then, is the amount of capacity that a given investment dollar buys. We computed this capacity amount on a historical basis by dividing the investment by the transmission capacity of the devices being purchased. This was distributed among optical equipment operating at the OC 12, OC 48, and OC 192 rates. For the purposes of this analysis, the bandwidth figures that were used to represent the OC transmission rates were 622 megabits per second for an OC 12, 2,488.2 megabits per second for an OC 48, and 9,953.28 megabits per second for an OC 192.

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 4: Global Optical Revenues

These data rates are deployed over various optical transmission/detection technologies, primarily dense wave-division multiplexing (DWDM), multiservice provisioning platforms (MSPP), and reconfigurable add-drop multiplexers (ROADM). DWDM uses different wavelengths of light to carry different transmission channels. In DWDM, this can be as high as 80 channels with 50 GHZ spacing. DWDM is most usually deployed to drive single mode fiber, whose core diameter is 9 micrometers.

MSPP, on the other hand is most usually found in metro rings. MSPP is used to drive SONET (Synchronous Optical Network) protocols to multiple add/drop multiplexers. Such setups frequently use multi-mode fiber with core diameters of up to 62.5 micrometers.

Finally, ROADMs are rapidly being superseded by MSPP and is expected to generally disappear from the metro environment by 2012.

The following (Please see Table 1: Optical Investment by Technology Type) shows the investments in the various technologies. Knowing the relative amount of optical data rates deployed in each technology, allows us to compute the optical capacity for each year. This gives us an incremental worldwide optical carrying capacity of 2.78 petabits per second in the year 2006 (Please see Figure 5: Incremental Global Optical Capacity).

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Table 1: Optical Investment by Technology Type

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 5: Incremental Global Optical Capacity

The essential problem, though, is not how much capacity was available in the past or is available now, but how much will be available in the future. While this is not precisely predictable, we can look at the revenue projections for the various manufacturers, look at the projections made by the various industry trade groups and look at recent historical trends.

Additionally, we can look at the predictions being made for relative revenues for each of the optical technologies. Predictions made by Lightwave in 2006 showed Metro MSPP technology investment flattening out in 2009, while ROADM investment would essentially disappear by 2012.

When these predictions are factored into the linear curves that simple trending would produce, the result is the following set of investment lines (Please see Figure 6: Projected Global Incremental Optical Investment). When we add these, the following chart results. (Please see Figure 7: Global Incremental Optical Investment).

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 6: Projected Global Incremental Optical Investment

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 7: Global Incremental Optical Investment

As can be seen, the trend line predicts that by the year 2012 approximately $9.8 billion of new optical investment per year will be made worldwide. This investment, when converted to capacity, shows the following curve (Please see Figure 8: Global Optical Capacity). As shown,, the total capacity of the backbone is approximately 5 million Petabytes per month or 5000 Exabytes.

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 8: Global Optical Capacity

5.1.2 North America

As the discussion above indicates, we can approximate the capacity for the optical backbone of the worldwide network by looking at worldwide investment. This yields an answer that seems reasonable on its surface, but which gives no insight into the situation in North America. To arrive at an approximation for North American capacity, it is necessary to utilize an abstraction. This is just the simple assumption that North American investment is proportional to the percentage of Internet usage generated by North America, and this in turn, is proportional to the relative number of North American users. The notion here is that investment tends to match usage, which users drive.

Fortunately, we know the relative number of North American users. In 2006, the United States Census estimated that out of approximately one 1.2 billion Internet users, 234 million of them were American. This yields a factor that allows us to index the total optical investment and compute the relative investment in North America (Please see Figure 9: North American Optical Capacity).

This approach may overstate the investment in optical capacity in North America, especially in future years, since it does not distinguish between optical distribution such as fiber to the home from optical backbones such as submarine cable. Still, it probably does a pretty good job of estimating the optical capacity available to North American users since an increasing amount of Internet traffic generated by North America is directed to sites overseas, which make use of the optical capacity implicit in the international backbones. Consequently, the optical capacity available to North America is actually higher than a pure investment figure would indicate.

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 9: North American Optical Capacity

5.2 Switching and Routing

5.2.1 Protocols and Layers

The Internet is based on the first four layers of the ISO seven-layer model: physical, data link, network and transport. The previous section on optical infrastructure covered the physical layer and some of the data-link layer. This section focuses on the data link, network and transport layers. Each layer communicates with layer above and below and each layer has its own set of protocols that define the communications interface and behavior of data transport. Most of the action in the Internet is focused on the TCP/IP protocol. TCP, Transmission Control Protocol operates at the transport layer (layer 4) and IP operates at the network layer (layer 3). Despite people talking about them as a single unit, TCP and IP are two distinct communications protocols. Another significant protocol is UDP (User Datagram Protocol) that also operates at level 4. UDP is a significant protocol for VOIP communications, video, chat and some peer-2-peer applications.

Typically, when people refer to switching and routing, they are referring to switching activities taking place at layer 2 - i.e. Ethernet â€“ and, routing activities taking place at layer 3 â€“ IP routing. Unfortunately, there is a lot of overlap in switching/routing equipment since most equipment does some level of both switching and routing.

The goal is to calculate an estimate of the total switching and routing capacity of the Internet infrastructure. Our research indicates that the last time anyone had firm data on Internet capacity was back in the early 1990s when the bulk of the Internet was still shared by research institutions and the core capacity was provided by the National Science Foundation (jointly with IBM, MCI and managed by Merit Networks). To start with the measurements done in the 1990s offers no value to today’s calculations; all of the equipment has been replaced and the NSF component of the network has been dwarfed by newer private capacity connections.

Another way to measure capacity is to interview all of the Internet service providers and add up all of their inter-switch trunks and lines to get an aggregate Internet capacity. This approach, too, is not possible. For this research, we interviewed a group of Internet service providers and asked them to share with us (for this research) their network architectures. None of the ISPs would disclose complete details on their infrastructures, however. We were able to glean tidbits of information from annual reports and briefings but it became a process of trying to put together a jigsaw puzzle with only a few pieces and a second-hand sketch of what the end-puzzle might look like.

We took a novel approach for this research: We postulated that all Internet switching and routing capacity must involve switching/routing equipment. We looked at the switching and routing equipment market and 85-90% of the market share is based on the shipments of Cisco, Juniper, Alcatel and Ericsson (Redback). Therefore, if we could track the shipments of equipment then we could start to build a network capacity model.

5.2.2 Methodology

For our analysis, we divided the switching/routing into two layers: core and connectivity. We picked these not because engineers necessarily think of the Internet in such terms, but because, it is possible to parse the investment in switching equipment into either those functions largely provided by carriers’ backbone networks from those that are largely provided by service providers. In telecom terms, we separated long haul from local traffic.

5.2.3 Core

Core Layer â€“ The core layer refers to the core routers/switches that comprise the service provider backbone networks. These are typically very high capacity units with high-speed interfaces, fault-tolerant operating systems, redundant switching and control planes and redundant power. These routers/switches typically provide a peer-to-peer functionality within the Internet that allows traffic to be handed off between networks and for traffic to be collected and routed efficiently worldwide.

Examples of core switch/routers include the Juniper Networks T-Series and the Cisco Systems 12000 Series. Typical capacity (2007) is in the multi-Terabit (1x10E12) range for throughput with interface support up to OC-768 and 10 Gigabit/s Ethernet. The performance of such switching equipment has benefited greatly from Moore’s Law. The price performance ensures that investments that have generally increased in a linear fashion can show exponentially increasing growth rates of switching capacity.

5.2.4 Connectivity

Connectivity Layer â€“ Many also consider the connectivity layer to be the Edge layer. The reason we differentiate is because edge implies a sharp demarcation point; the edge of the network. In reality, we see this layer as more of a fluid boundary between the core and the access layers. This layer, as distinguished from the long haul or peering traffic, typically routes traffic within metroplexes or between metroplexes within the same carrier network. Typical equipment in the connectivity layer includes:

Routing - Juniper Networks M-Series Multi-service edge switch

Switching - Juniper Networks MX-Series Metro Ethernet Switch

The connectivity layer is where CLECs, ILECs and MSOs drop-off their Internet traffic. For our analysis, we include equipment that interfaces with DSL, Metro Ethernet, Cable and FTTH. While this tends to blur the boundary between access and connectivity somewhat, it lends itself to relatively easy analysis and the capacity involved is easy to discriminate from access line capacity.

Ethernet Aggregation provides interface at the connectivity layer to cable operators, DSL providers, wireless providers and direct connection to enterprise customers. Ethernet interfaces have been around since the 1980s, Ethernet as a service provider interface is a relatively new phenomena. Acceptance is high since the interface is ubiquitous, high bandwidth and the customer premise equipment is low-cost. Our research indicates a shift from more traditional IP services interfaces to Ethernet interface over the coming years. We reflected this shift in our capacity calculations.

As in core switching, Moore’s Law impacts the price/performance of connectivity switching. Over the course of the historical data and as projected into the future, there is a clearly exponential growth observable. This will become apparent as the investment in switching equipment is translated into capacity.

5.2.5 One Box Two Trunks

After extensive review of network capacity modeling, it’s clear that the level of model complexity increases at a greater rate than the complexity of the network being modeled. Given the fact that there is no network more complex than the Internet, a true capacity model stretches the limits of current capacity modeling systems. In fact, theoreticians can’t even agree on how best to model the capacity of the Internet, let alone actually model the capacity of the Internet, itself. Given this reality, we decided to approach the problem from a very simplistic perspective: one box, two trunks.

We know that every switch and router in the Internet has at least two trunks connecting to other switches and routers. Each trunk operates at a given rate but the rate is not equal to capacity. Network engineers load balance traffic between trunks, and based on discussions with Internet service providers, trunks are typically sized to operate with a peak capacity of no more than 30% loading. As an upper limit, a node with two trunks should have no more than 50% peak loading on either trunk to provide for redundancy.

Our approach is to calculate capacity by assuming that each node has two trunks running at 100% loading. In the real world of the Internet this would be equivalent to the capacity of four trunks running at 50% peak loading, or six trunks running at 33% peak loading, non-stop. This approach, though simple, yields a high-level capacity figure that takes into account the fact that line rate is much higher than actual capacity, yet makes the process of equating line rate to capacity possible.

Estimates for equipment shipped were calculated based on interviews with router manufacturers, researching market research data and analyzing annual reports of both service providers and equipment manufacturers. All information taken together gave us a global estimate of shipments of service provider router/switch equipment, over time.

An estimate of boxes does not directly correlate to Internet capacity for a number of reasons. First, only a percentage of units shipped are ever installed. The rest may be put into spare inventory or disaster recovery inventory; neither ever having an incremental impact on Internet capacity. Second, the switch/router unit is only half of the capacity equation; the other half being the network trunks. Core routers typically have a minimum of two trunks with more trunks added as demand and network routing complexity increase. Finally, there is overlap between enterprise switches and routers and service provider switches and routers. Just looking at total units shipped can be misleading.

To address these issues we did the following:

1. Over time the service provider and enterprise routing equipment has diverged. Today, there are clear product lines destined for service providers and enterprise clients. There is still some overlap where the largest enterprises require low-end carrier class routers and the smallest service providers can use high-end enterprise routing equipment. We believe that this overlap is minimal and not significant to our analysis. For our backward looking analysis, we estimated a percentage of total product shipped was destined for enterprise and we therefore removed it from our capacity calculations.

2. Our goal was to calculate maximum available capacity to match our calculation of maximum potential demand. Therefore, we assumed that all equipment shipped contributes to Internet capacity. This leads to estimation beyond the actual capacity but this over-estimation is mitigated by the maximal demand approach.

3. In reality, the capacity of a given switch/router is directly related to the capacity of the trunks that connect the unit to either peer or higher-level networking components. In fact, the rate of switching capacity - packets per second and backplane throughput- has increased at a greater rate than the increase of speed of the trunks. Therefore, for our calculations, we have assumed that the switch/routers processors are not rate limiting and that capacity is directly related to the capacity of the trunks interconnecting the nodes.

The following table (Please see Table 2: Global Shipments of Core and Connectivity Nodes) shows the estimations of global shipment of core and connectivity nodes (2000 â€“ 2006) and the projected trunk rates for each node. The trunk rates are based on typical connectivity and core trunk rates for each year of the analysis. Given the scope of the assessment, we know that units connected at higher speeds and lower speeds than the rates shown. However, on a global basis, we believe that these rates are a fair estimate of the higher-end trunk rates available at the time.

	2000	2001	2002	2003	2004	2005	2006
Core Units	11400	8300	6660	7744	9594	10666	12071
Connectivity Units	62439	36880	10128	11269	12998	23365	30411
Core Trunk	2.4E+09	4.8+09	9.9E+09	9.9E+09	9.9E+09	1.99E+10	1.99E+10
Connectivity Trunk	1.2E+09	2.4E+09	2.4E+09	4.8E+09	4.8E+09	4.8E+09	9.9E+09

Table 2: Global Shipments of Core and Connectivity Nodes

We based the calculations for Internet capacity on summing the yearly incremental capacity. We recognize that adding 100 core routers to the network doesn’t raise the total number or core routers by 100 â€“ some equipment replaces older equipment. We assume, however, that capacity never decreases. Even when a new router replaces an old router, the aggregate capacity of the new router is greater than the capacity of the existing router since the main reason for upgrade has been, and will continue to be, to increase capacity. Our calculation of incremental capacity for an upgraded node will initially be higher than actual. However, over time as additional trunks and higher speed trunks are turned up, our initial over-calculation will be compensated for.

Our research only goes back to 2000, so we needed to determine a starting point for Internet capacity (Please see Figure 10: North American Switching Capacity). In reality, the amount of bandwidth added since 2000 is much greater than any starting bandwidth position, so that the starting point has surprisingly little impact on the final capacity curve, as will be discussed shortly. But, to assess a starting point we made the assumption, based on input from experts (including Andrew Odlyzko) that the Internet capacity in 2000 was probably roughly equal to Internet demand. It’s important to note that roughly equal does not mean exactly equal since supporting X kbit/s of demand requires n x X kbit/s of switching and routing capacity. This is related to queuing factors and the need to have much more switching capacity than demand to support a given level of demand.

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 10: North American Switching Capacity

On a global basis, we estimate that core and connectivity switching capacity in 2000 was 66,000 Petabytes/month. For North America, the estimate is 40,000 Petabytes/month. Over time, core and connectivity capacity have closely tracked each other, and starting in 2007 we project that connectivity capacity will increase at a faster rate than core capacity. From a network design viewpoint, these curves make sense. As higher speed interfaces are extended toward the edge of the network, the connectivity layer will require more total capacity than the core layer. Also, a significant percent of connectivity traffic stays in the region and never touches the core.

We based forward-looking capacity projections on observation of historical increase in capacity as well as primary research on Internet capacity growth (Please see Figure 11: Growth in Global Switching Capacity). Historically, we see capacity growing at a high rate of 61% year-over-year between 2001 and 2006. During this time the core capacity grew faster than connectivity capacity: 75% and 53% respectively. However, toward the end of the period we see this gap closing and for forward looking projections, we projected a total capacity growth of 50% per year with the continuation of the trend of connectivity growing slightly faster than core capacity.

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 11: Growth in Global Switching Capacity

As a point of reference, our projections fit well with other analyses. In the 1990s there was talk of explosive growth of Internet capacity leading up to the 2000 technology bubble burst. Starting in 2000 there was a significant contraction of investment in Internet infrastructure. Investment didn’t stop but as seen in the slope of the curves, the capacity growth rate dropped from 124% year-to-year in 2001 down to a low of 26% year-to-year in 2004. In 2004 we see a reversal of this trend with the year-to-year growth increasing to 43% in 2006 and 50% projected in 2007. As Andrew Odlyzko noted, within any given timeframe the Internet may be growing at an explosive rate, but over time the traffic growth rates are approximately 50% per year. Similarily, we believe that at any time capacity may be growing at an explosive rate , but over time the capacity growth rates are approximately 50% per year. As discussed in the Section 4 (Demand), we are projecting a higher year-over-year growth rate for traffic.

Another way of rationalizing our capacity projections is to compare investment to capacity. As discussed in Section 6 (Investment) we are projecting an annual increase in investment of 10% per year. At the same time, Moore’s Law states that the amount of switching capacity per $ should double every 18 months. Considering the 10% increase, a 50% annual increase in switching capacity is a reasonable assumption based on Moore’s Law.

5.2.6 Access

While the Internet is generally thought of as a cloud of connections into which a user plugs, architecturally, the Internet cloud begins where the user accesses that cloud. This access layer is the connection between the user and the ISP (Internet Services Provider) and is usually provided by the local telecommunications carrier.

Access comes in several flavors. Up until the end of the last century, the most common way to access the Internet was through a dial up modem over a telephone line. These dial up connections were limited to 56 kilobits per second and were minimally able to support activities such as Web surfing. Beginning just prior to 2000, though, a significant percentage of users began to acquire broadband access to the Internet, either through DSL (Digital Subscriber Line) technology or cable modems. In the case of the former, the service still traveled over the telephone connection, while the latter was carried over the local cable television cable. Increasingly, newer access technologies based on wireless and optical fiber are being deployed, although their impact on the market is nominal at this time.

It is clear that access plays a significant role in the degree to which users translate their demand to capacity consumption. If the only available access is 56 kilobits per second, then the degree to which packets can be placed on the Internet backbone is far lower than if the connection available is running at a one megabit rate. Because access constraints may throttle demand it is important to determine the actual carrying capacity of this layer.

The analysis of access capacity begins with an inventory of the access lines available. In the case of North America, this is relatively straight forward. The carriers report their broadband access line count routinely to the Federal Communications Commission (FCC). Over time, North American broadband access numbers have followed a relatively linear deployment profile (Please see Figure 12: North American Broadband Access Lines). In 2006, the number reported by carriers was about 99 million.

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 12: North American Broadband Access Lines

Assuming that such deployment continues to follow the same curve, the curve can be projected as shown below (Please see Figure 13: Projected North American Access Lines).

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 13: Projected North American Access Lines

This curve, though, only represents a fraction of the total access capability. In order to be comprehensive, wireless devices must be included as well. Additionally, enterprise access, although small in comparison to the total consumer access needs inclusion as well. When this is done the total broadband access lines very much exceed the total number of potential users, however, this just means that typically users have available to them several different access capabilities simultaneously. When these are included, the curve for North America appears as follows (Please see Figure 14: Total North American Access Lines):

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 14: Total North American Access Lines

This resolves the problem of access lines for North America, but what of the global distribution of access lines? It turns out that North American users as a percentage of total global users has been inventoried by the U.S. Census. When this is combined with data points provided by the EU and countries in Asia and Latin America on the total broadband access lines installed there, it is possible to take the North American access lines and index them to derive the balance of the world’s access lines. When this is done, the following curve is derived for the entire Internet (Please see Figure 15, Global Access Lines).

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 15: Global Access Lines

This information combined with the distribution of access line technology plus the likely data rates provided by that technology, generates the following two capacity curves for North America and the World respectively (Please see Figure 16: North American Access Capacity and Figure 17: World Access Capacity):

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 16: North American Access Capacity

Image Placeholder

The images are only available in the full report.

Contact research@nemertes.com if you have any questions or technical difficulties

Figure 17: World Access Capacity

As can be seen, the broadband access capacity grows at an essentially linear rate over time. This curve, however, depends on two assumptions. The first is that world fiber to the home will be increasing over time. The second is that wireless devices will increasingly become a surrogate for fixed access technologies. While the number of mobile devices worldwide is known and the projected uptake can be surmised, the degree to which these devices will be used for data is uncertain. If the technology implicit in wireless devices continues to improve and data rates available to basic devices increases, it is possible that access line capacity could increase by several times on the tail end of the curve, thus inflecting the curve up or , at the very least, straightening it.

5.2.7 Wireless: Building Footpaths Across the Digital Divide

Another consequence of this modeling effort was the realization that access is now evolving past the limitations of fixed infrastructure. While most users of today’s Internet access it through landlines, either DSL (Digital Subscriber Line) or broadband cable connections, an increasing number of users, mobility workers in particular, are accessing the Internet through either fixed wireless or mobile wireless technologies.

This has been an evolving process. Once wireless service evolved past simple analog technologies (1G or first generation) to digital (2G or second generation), mobile workers wanted to use their instruments to access data services, especially the Internet. 2G technologies, though, were not sufficiently robust to deliver acceptable data rates. This led to the development of so-called 2.5 technologies which allowed for reliable data transfer, albeit at rather slow data rates; typically no more than 144 kilobits per second.

With the introduction of 3G technology and data rates much higher than 144 kilobits per second, mobile access to the Internet begins to be an attractive possibility. Using protocols such as EV-DO which can achieve burst rates as high as 1.8 megabits per second under the REV A standard or many times that under the REV B standard, land line comparable access is possible. 4G technologies, which are still being developed, are expected to push these rates even higher and will likely provide access rates that exceed many fixed access technologies.

The net impact of this evolution is that the load on the Internet will increasingly be driven by wireless access. While that influence does not significantly influence the results of the model, to the extent that such an influence is more than nominal over the study period, it could have profound implications for network demand. If, for example, new services based on WiMAX, a new 3.5 G wireless data technology, become prevalent, demand could be significantly impacted.

6 Investment

6.1 Methodology for Determining Investment

We derived Internet infrastructure investment from three primary sources:

1. Extensive research of investment information fr

Nemertes Research

Top Menu

User login

Keyword search

Guided search

Research Topics

Navigation

Date authored

Related categories

Research Topics

Navigation

Date authored

The Internet Singularity, Delayed: Why Limits in Internet Capacity Will Stifle Innovation on the Web