This article is based on a presentation I made to the ARIN 56 meeting in October 2025.

Here, I’d like to elevate the typical Regional Internet Registry (RIR) policy conversations above the day-to-day work of address allocation policies with its vocabulary of address block sizes and needs-based justifications, fairness, and efficiency. Instead, we’ll look more broadly at the context of the industry we operate in, and try to gain an understanding of where we are right now, and speculate on where it’s all going.

Where we came from

We are caught up in one of the most productive marriages in all human history, namely the marriage of computing and communications. Since the invention of the transistor in December 1947 and the integrated circuit in 1958, the world has been fundamentally changed.

Until then, all human endeavours had been limited by their geography. The industrial age and the explosive introduction of the railway in the mid-19th century caused a significant societal change as economies shifted their foundations of wealth and power from land and agricultural activities to industrial production. Facilitating these changes was first the telegraph, and then the telephone. These communication technologies allowed industrial companies to project their power and influence across larger areas, harnessing greater volumes of production and increasing their wealth.

When computers were added to the communications realm, the rate of change increased dramatically. It took a decade from the invention of the transistor to that of the integrated circuit, and a further decade for computers to become an essential tool for data processing and communications.

In the late 1960s, Bell Labs programmers Ken Thompson and Dennis Ritchie had devised the Unix operating system. It was written in the C language — a ‘high-level language’ — and its associated compiler could produce assembler code for a variety of computers. Unix was one of the first of the ‘open’ operating systems of the age.

This was not exactly an unforced choice. Under a 1956 consent decree in settlement of an antitrust case, the Bell Systems (the parent organization of Bell Labs) was forbidden from entering any business other than common carrier communications services. It was also required to license any patents it had upon request. Unix could not, therefore, be turned into a product.

Bell Labs shipped the source code of the Unix system to those who asked for the cost of media and shipping, allowing universities and other organizations to modify and extend it. This open model, both of the code itself and the refinement and development of the system, facilitated the development of many variants in the ensuing years, including the highly influential Berkeley Software Distribution (BSD).

In 1973, Bob Metcalf, working at the Xerox Palo Alto Research Centre, published a memo describing ‘X-Wire’, a three million bit-per-second (Mbps) common bus local area network, which became known as ‘Ethernet’. It was notable because Ethernet was the absolute minimal form of computer networking. It really was just a wire.

It had no internal switch, no packet framing, no clock, no controller, no network state. Nothing! Just a wire. And in many ways, Ethernet changed the way we thought about computer networks and communications architectures for computers. Ethernet took off because every conventional function of network control was pushed off the network into the connected computers and implemented as a collection of distributed algorithms.

The network itself, this wire, was only used for transmission. ‘Dumb network, smart devices,’ as we constantly reminded ourselves. This also meant that the money to operate the network wasn’t the network’s problem.

Figure 1 — X-Wire framework. Hand drawn by Bob Metcalfe and photographed by Dave R. Boggs in 1976 to produce a 35mm slide used to present Ethernet to the National Computer Conference. Source: IEEE802.

The collective ‘intelligence’ for an Ethernet Local Area Network (LAN) was placed into the computers that connected to this common wire. All Ethernet packets were self-clocked, using a 64-bit preamble to allow all receivers to synchronize their receiving data clocks against the data rate of the packet being transmitted. Packets were between 64 and 1,518 bytes in total length.

There was no centralized contention control — each station used a Carrier Sense Multiple Access/Collision Detection (CSMA/CD) protocol to negotiate exclusive sending access to the wire, and all connected computers received all packets that were passed onto the wire. It used a unique station addressing plan in the form of a 48-bit MAC address that is still with us today! But most importantly, like Unix, it was an open standard specification.

Figure 2 — DEC Vax 11/780. Source: Wikipedia.

The next image (Figure 2) takes us back to 1977, with the release of the VAX 11/780 computer, made by the Digital Equipment Corporation. This was a medium-sized mainframe computer capable of executing one million instructions per second (MIPs). This was a very popular computer used in many corporate environments, as well as universities and research institutions.

If you look at other 50-year-old technology artefacts, many of these items don’t look all that different from what we use today. Yes, cars today are lighter, more efficient, and come with a whole lot less chrome, but they still have a wheel at each corner and some internal engine for propulsion. There are still 50-year-old cars around today, and some of them are still on the road. But this is not the case for computers. The very few 50-year-old computers left in today’s world exist in museums. Why has the computer industry been so ready to cast off its past lives and embrace change so readily?

Re-introducing Moore’s Law

The essential difference here is the miraculous pace of Moore’s Law — an observation by Intel co-founder Gordon Moore in 1965 — driving the improvements in the production of integrated circuits on silicon wafers. Moore’s Law is the observation that the number of gates on an integrated circuit doubles approximately every two years, while the cost of fabrication of the chip has increased by a far lower factor, if at all. The cost and quantity of compute capability have improved consistently over the same 50 years (Figure 3).

Figure 3 — Moore’s Law transistor count 1970 – 2020. Source: Wikipedia.

Exponential growth is often hard to understand. It has its own momentum and unintended consequences. Initially, this was all about making these lumbering mainframe computers faster, smaller, and cheaper. The mainframe computers of the early 1960s took up an entire data room with multiple racks of circuitry, power and air conditioning. These machines had their useful service life that was best measured in units of months rather than decades!

This dynamic pace of evolution of computing capability has impacted many aspects of the computing environment, including network protocol design. One of the objectives of a computer network was to allow any connected computer to communicate with any other. For that, each computer needed to have a unique identification field, or an ‘address’ to distinguish it from all the other connected computers.

The proprietary network protocol used by the VAX systems was DECnet, and in the 1980s, Phase 3 of this network architecture was in widespread use. It used a 16-bit address field, allowing a maximum of 65,535 computers in a single DECnet environment. In the late 1970s, the concept of a single network populated with tens of thousands of these large lumbering computers was an unaffordable fantasy. But that thinking did not factor in exponential growth through Moore’s Law, and the associated exponential drop in size and cost.

The computing environment rapidly shifted in the 1980s into personal computers. The multi-million dollar mainframes of the 1970s, with their specialized environments of highly curated clean rooms, transformed into small boxes that found a home on every desk, both in the workplace and in the home. The idea of running out of 16-bit DECnet addresses, which seemed impossible in the 1970s, was an obvious reality within just ten years.

But not every computer networking protocol had such a severely constrained design. The folk working on the Internet Protocol (IP) took a far more radical step and used a 32-bit device address architecture. Instead of an inbuilt ceiling of around 65,000 connected devices, the IP protocol had a ceiling of around 4.3 billion connected devices! This was crazy thinking, even during the period when computers escaped from their clean rooms and invaded offices and homes. There weren’t even 4.3 billion people in the world.

The three key ingredients were in place that allowed for the explosive growth of the Internet:

1. Advances in the speed, size and cost of computing.
2. An openly available, highly efficient operating system for these computers in the form of Unix.
3. A freely available computer networking protocol in IP that had this massive inbuilt capacity through its crazy decision to use a 32-bit addressing architecture!

This was why the IP design was so revolutionary. A few years after that photo in Figure 2 was taken, there were a few hundred thousand of these VAX computers in the field. And we looked at the IP address space, with some four billion addresses, and said to ourselves: ‘What’s your problem? We’re here forever!’

Awesomely wrong, as it turned out!

Moore’s Law continued to cause rapid change in this industry. At the same time as the VAX 11/780 was released, over on the West Coast of the US, the Apple Corporation released its first Apple computer. This device was never going to challenge the mainstream computer industry. It cost hundreds of dollars, not millions, it lived on the desktop, and it played games.

Steve Jobs, Apple’s founder, appreciated the true effects of Moore’s Law and searched for untapped markets that could absorb the production capacity of the silicon chip industry. Apple was a pioneer in the consumer market and applied design principles to personal computers. The computers created by the company looked like they were truly part of the home.

But Steve understood that Moore’s Law meant more than ‘cheaper and smaller’. Computing capacity could be used to elevate his machines to interact with human users on human terms. Apple developed the Mac, a computer with an intuitive look and feel that did not need a geeky user manual.

When you powered up an Apple Mac, it displayed a friendly ‘hello!’. The human at the keyboard didn’t need to memorize some arcane Job Control Language to make this machine sing and dance. In the consumer market, this worked! The personal computer market quickly outpaced the mainframe market and dominated the entire computing environment.

Coincidentally, this graphical interface was also developed at the Xerox Palo Alto Research Centre, and while Xerox was highly influential in these landmark inventions, the company stubbornly stuck to being a photocopier company!

From telephony to broadcast TV

Something else was also happening here. In these early days of computer networks, the network was viewed in the same terms as the telephone network. Human users of the telephone both spoke and listened. The telephone network was symmetric and invisible.

In computer networks, connected devices both provide digital services and access the services of others. Connected computers were both client and server in the networking environment. This didn’t take full advantage of Moore’s law.

As more and more personal computers were added to networks, they no longer filled the role of both client and server. They were positioned as simple clients. There was an emerging segmentation of connected computers into dedicated client and server roles. The large mainframe computers didn’t disappear. They were used to provide services to the millions, then tens of millions of users of these client personal computers.

Rather than a computer network built in the image of a telephone network, it began to look more like broadcast television, where every client can perform their own programming. Users at home were not providers, did not actually want to host services, and certainly did want these service delivery platforms to be located in their homes. They just wanted the computing equivalent of the television set.

Clients and servers

The emerging Internet network of the late 1990s was constructed around the model of clients and servers. This was convenient because the Internet was being built on the underlay of the telephone network. This asymmetric behaviour of clients and servers was built into the network architecture of the Internet itself. If you were a client, you couldn’t do much at all, but you could make contact with servers. The connected edge client computer was the thing that simply said, ‘Let’s look at the data out there’.

The dial-up world of the 1990s and the DSL/Cable modem world of the 2000s in the last mile access network were a good fit to the demands of client/server networking at the time. The Internet rapidly expanded by repurposing the existing last-mile infrastructure, avoiding the need for hefty upfront capital investment by Internet Service Providers.

Millions of clients were eager to own personal computers and use the Internet to access digital services, and the industry continued to grow. Around the year 2000, specialized data centres emerged to combine all these servers into a common curated environment with power, cooling and a couple of human attendants to scurry about and replace dead or dying server units. This server world also saw role specialization. There were web services, mail services, data services, archival services and so on.

Compared to the current AI-scale data centre designs, this was a small-scale activity. These Internet service data centres were only a room or two with a few hundred megawatts of power. Online services were now crammed into these dedicated service delivery points.

As the Internet continued to expand into the consumer space, large-scale ‘network exchanges’ and ‘network peering points’ were used to service the concentration points where data intersected with service delivery capability. Clients never exchanged data with other clients. Every transaction was a transaction between a client and one or more servers.

At the same time that the silicon industry was working hard to expand the network’s client base, traffic engineering was being used to ensure that the limited resource, network capacity, was used fairly and efficiently.

No longer ‘just a wire’

This network was now a long way away from being ‘just a wire’. Instead, it depended on hardware and software to perform all kinds of functions relating to resource control, variable service response, defensive measures, and, of course, data collection.

The progress of Moore’s Law led to increasing functionality, as the processing capability of these service delivery platforms increased. The progress of silicon chip capability extended further than data centre services and in-network middleware.

When you shine a laser light through a fibre optic cable and modulate the light to carry a digital signal with on/off keying, you can work with a simple light detector. If you want additional data capacity from a fibre circuit, then the logical approach is to use the same signal modulation techniques that were previously used for analogue modems, namely using phase and amplitude modulation of the light signal.

While this can radically improve the carrying capacity of a fibre system, the practical limits of this approach are strongly dependent on the signal processing capabilities used at either end. We can push up to almost a terabit per second of signal through fibre optic cable because of better Digital Signal Processors (DSPs).

Using three-nanometre chip technology, it appears that you can achieve 800Gbps from a single polarization of a light signal through fibre. It’s not easy, but it’s achievable. At the same time, we’ve lifted the common network bearer capacity to 100 to 200G with the latest generation of DSPs.

A network of abundance

These developments have caused a huge shift in the role of the network. When transmission capacity was a scarce resource, then rationing access to the resource across competing demands was the role of the network. Rationing scarcity is a high-value task, as in economic terms, the controller of the resource can demand a ‘scarcity premium’ from competing users.

But if capacity is abundant to the extent that it overwhelms all demands, then the network operator’s role shifts into a basic, undistinguished commodity broking role. We’ve seen this shift in the commercial world, where the share value of network carriage operators has plummeted in line with this shift from scarcity to abundance.

The response from the network operations sector has been an effort to add further functionality to the network, in the form of variable service responses, network segmentation and customizable forwarding functions. So far, this has increased costs in the network but has failed to attract sufficient interest from the network’s client base.

The same basic improvement in technology capacity also means that connected devices are increasingly capable of managing their own requirements for service responses from an amply dimensioned, abundant network. In this environment of network abundance, the network service itself quickly becomes an undistinguished commodity good and is priced accordingly.

The silicon industry then turned its attention to handheld and embedded devices, and once more, it overachieved! Billions of devices were added to the Internet in a couple of years, resulting in a digital world that is defined largely by abundance and scale. Abundant processing, abundant storage, and abundant carriage capacity,

What happens in an environment of abundance? Pricing is used to ration the distribution of a limited resource — a condition that no longer applies. This represents a particular challenge to network carriers. Their historical role has been to share a single, limited resource between people who want to use it. The underlying value of the network infrastructure is much reduced. In 2023, T-Mobile sold the Sprint wireline work to Cogent for one dollar.

The role of the network as a ‘just-in-time’ service that connected users to service on demand has been largely replaced by Content Delivery Networks (CDNs). The delivery model of a CDN is to position replicas of the service content across a large set of service points. This could be characterized as a ‘just-in-case’ model, where content is pushed out to the edges, close to potential consumers, just in case they ask for it.

This has had its own ripple effect. As content and service move closer to users, the network becomes smaller. Distance is eliminated as the dominating characteristic of the network. We don’t grab content from far away any more.

The packet-miles, or the distance a packet travels from the server to the client, are now tiny. When you shrink the network from the circumference of the globe or across a continent down to the dimensions of traversing a single city, then everything is easier, faster and cheaper.

If you can build an Internet that looks and performs like a LAN, then where’s the scarcity? Where are the performance issues? They’re largely gone.

Bigger, faster and cheaper

Today’s distance-shrunken Internet is, oddly enough, bigger. There are terabits of capacity in the core systems, and gigabits out to the edges — in last-mile fibre and last-mile radio. Content is published everywhere, all at once. The network is no longer ‘distance big’, but rather ‘capacity big’. As distance shrinks, protocol performance improves and cost decreases.

This network is now fast. Really fast. Not only do I get a gigabit per second to my house, but I can actually pull down content from a cooperating server or two at that same line speed. Shorter distances improve network performance because network transport protocols are feedback-controlled. When you bring the two ends closer together, you increase the number of potential feedback cycles per unit of time. This increased information flow provides better guidance to the transport protocol to optimize the efficiency of delivery.

And, yes, it’s cheaper. If I took a gigabit per second Internet service and pushed it back to the currency unit of 1975, fifty years ago, then the USD 100 per month that users pay on average for such a residential service would cost USD 17 of these 1975 dollars. Not only is this cheap by historic standards, but Low Earth Orbit (LEO) satellite-based services have extended the Internet to the most remote corners of the surface of Earth.

Continuous refinement in the fabrication of silicon chips has resulted in a better, faster, and cheaper network.

The end of Moore’s Law?

So far, we’ve looked at the upbeat story of the past fifty years. But while Moore’s Law has carried the burden of the progressive improvement in this industry for the past sixty years, are we reaching the end of this journey?

Current silicon chips have feature sizes of three nanometers (nm), and it looks like 2nm is achievable. To put that into context, a silicon atom has a diameter of 0.23nm. This implication is that a chip feature of 2nm in size is only ten silicon atoms wide.

Are one-nanometer feature sizes on silicon chips possible? Well, you can try. Extreme ultraviolet rotation is already used to produce a wavelength of 13.5nm for chip lithography, and the process is extremely complex.

How much further can this continuous improvement in silicon chip manufacturing continue? When will we reach the end of Moore’s Law?

Even if a reliable and scalable lithography process is possible, there is also the problem of electrons. Let’s assume they’re particles, leaving aside the duality of quantum mechanics. In small enough spaces, electrons have a mind of their own. It’s hard to get this chip feature size even smaller using today’s planar integrated circuit technology.

The process that was used in 1957 to create an integrated circuit has been progressively refined year after year, and the features have become smaller and smaller as a result. But this process of continuous improvement may grind to a halt.

Nobody is announcing a four trillion gate silicon chip this year, and I’m not sure they ever will. This technology, as it stands, is close to that physical edge of impossibility, and it appears that the next innovation needs to be really off-the-wall clever and do something that none of us has thought possible.

That’s not the only limitation we’ve encountered in chip capability. The clock speed of processors topped out around the year 2000. Processors may have more gates, and CPUs may have more cores, but the processor’s clock speed has remained pretty constant since then (Figure 4).

Figure 4 — CPU clock rate. Source: Wikipedia.

It’s way too early to proclaim that Moore’s Law is over. But perhaps it is wise to contemplate what it means for us if Moore’s Law becomes the gift that has effectively stopped giving

IPv6 and IP addresses

How does all this relate to IPv6 deployment?

The effort of the early 1990s was intended to provide a path through the situation where the population of connected devices exceeded the capacity of the 32-bit address space. The response, a redefinition of the IP protocol using a 128-bit address field, preserved all the behaviours of the Internet.

However, the lack of any form of backward compatibility implied that all the network’s users and operators were exposed to the marginal cost of adopting a new network protocol, but in fact gained none of the marginal benefits that a completely different network protocol might offer. It was still just IP.

There was an alternative option available at the time: Deploy Network Address Translators (NATs) in the network’s infrastructure. NATs are a client-side response that uses greater processing capability to make up for this shortfall in IPv4 addressing. This approach is viable in the long term only if we can scale processing efficiency in line with demand growth.

The purpose of NATs was to bridge that gap between the capabilities of the technology we had and the size of the network that we wanted to deploy. As the network grew, the size of the gap increased, and the cost in terms of infrastructure complexity got higher and higher (Figure 5).

Figure 5 — Bridging the IPv4 address gap.

At some point, this embedded processing in the network becomes problematic. It’s a cost element, and costs in a commodity network are intolerable in the long run.

The NAT response to the IPv4 address depletion gives little leeway to reduce network complexity. Complexity must be reduced to continue to deploy digital services at scale, securely and affordably. Deploying IPv6 is one obvious response as a way of achieving this.

Who received this memo? India got the memo in 2017 and rolled out a V6 network across a billion users in 12 months, a scale and speed that is unprecedented and well-suited to the Internet of today. (Figure 6)

Figure 6 — IPv6 in India.

China is facing a similar set of problems, but is proceeding at a more leisurely pace with its program of IPv6 adoption (Figure 7).

Figure 7 — IPv6 in China.

There’s another tension going on here, and it’s the vertical tension caused by value moving from the protocol to the application layer. The application is now the service.

The client/server network needs accessible address points for servers — or to be more correct, for services — but not for clients. In the public Internet, clients do not need to be associated with a persistent IP address.

The QUIC transport protocol takes this a step further with the observation that the network can change the IP address of the client at will, and as long as the address assignment is stable for a couple of round-trip time intervals, all will work!

Even servers do not strictly require persistent IP addresses. As long as they remain usable across some DNS resolver cache lifetime, IP addresses for servers need not be stable for very long either. It’s the DNS namespace we use as the persistent service identity schema for the Internet. Our mechanisms for service authenticity and channel security are based on persistent service names, not on their IP addresses.

In the light of this, I can’t help but ask: Do IP addresses matter?

I have to observe that, yes, IP addresses do matter. But I must qualify that to add that they matter at the moment, but this is not necessarily an invariant property going forward. IP addresses are not persistent identity tokens, nor are they useful location tokens.

They are useful tokens to pass into the routing system to assist packets in being passed to their intended destination, but even this is not quite what it seems. For example, the Multiprotocol Label Switching (MPLS) approach places a wrapper around an incoming packet that identifies the desired network egress point and a path to reach that egress point. On its transit across that MPLS network, the inner IP protocol destination address is unused for forwarding the packet. So more accurately, addresses are used by individual networks to determine where to evict each packet from the network!

An extreme view is that IP addresses are temporary tokens that map the service layer of named services to the underlying layer of packet forwarding, and that’s about it. Yes, there is a whole set of ancillary roles about reputation, accountability, location and reporting attached to IP addresses. But perhaps these additional uses of addresses were making assumptions about the role and persistence of IP addresses, which are unwarranted in the longer term.

Networks are markets

The story is not done. The silicon chip industry is still operating at full speed with the expectation that all of these processors will end up in Internet-connected devices in one form or another. The network must be scaled further. It’s service level scaling that’s the challenge, orchestrating the behaviour of individual network components to generate coherent outcomes.

Do we know how we will achieve this? Not really.

But don’t forget that the Internet is not a centrally planned global enterprise. No single person, entity, or organization is in charge, despite some persistent suspicions about the role of the United States in the background. But such suspicions are more in the realm of conspiracy theory fodder than a clearly established role.

If there is no single economy in charge, then is it a group of economies? There is no specific international treaty to provide oversight for the Internet, no international convention or international protocol. There is no imposition of control over the Internet, or even a clear exposition of direction.

The Internet is a diverse collection of markets, large and small. Markets are a means of orchestrating the behaviour of actors to produce coherent outcomes in a manner that also attempts, often imperfectly, to maximize efficiencies along the way.

However, markets do not necessarily behave rationally or even predicably. When we gaze into the Internet’s crystal ball and attempt to make such predictions about its future, the vision is made as if through a glass darkly.

Rate this article

---

The views expressed by the authors of this blog are their own and do not necessarily reflect the views of APNIC. Please note a Code of Conduct applies to this blog.