Geoff Huston 0:00 Oh, I wish you wouldn't ask these hard questions, autonomy, I suppose, is actually a synonym, in this case, for independent you see, what you need in routing, as in any large system, is some form of abstraction. You need to be able to clump together things that are the same and treat them as one that way you can scale because if you disregarded the Internet as a collection of a few trillion atoms, you know, everyone's brain had melt and the system wouldn't work. So we group together groups of connected hosts and their associated IP addresses and call them an autonomous network. Now, what do we mean by autonomous, or, as I prefer to say, it independent? Well, every network has the ability to determine how it connects to the other networks that make up the Internet, and that decision as to how it connects is its own decision. It's not determined by anyone else. It has its own free will, in some ways, its own autonomy. So I might run a network here and say, Well, I choose to connect to service provider Fred. And you might be running a network in your town, you might say, well, oh, Fred's a bit expensive up here. I'm going to connect to service provider Carol. And that independent decision is yours to make. George Michaelson 1:34 You're listening to ping, a podcast by APNIC, discussing all things related to measuring the Internet. I'm your host, George Michaelson, this time I'm talking to Geoff Huston from APNIC labs again in his regular monthly spot on ping. This is Geoff's first ping episode for 2026 and he's kicking off the year with his annual review of what's been happening in BGP. It's a look back over 2025 in routing. Geoff has been doing this analysis for over a decade, and normally it's a series of up and to the right charts across almost any measure you care to mention, although we should note that there's always been a level of noise in the BGP system, the more specifics from traffic engineering and de aggregation, and to some extent that's constant under growth. Geoff has detected some changes in the dynamics of the BGP trend lines. This time around, we may be entering a new phase of routing, something we haven't seen before. It's going to be interesting to see how things emerge across 2026 and into the future. Geoff, welcome to 2026 and welcome back to ping. What should we talk about this time? Geoff Huston 2:46 Well, thank you, George. And yes, welcome to 2026 look. It's the start of another year, and I've often used to at the start of each year, sort of look back over the previous year and sort of see how the Internet fared. George Michaelson 2:59 I've been looking in the archives, and I can see you doing this for the last 10 years. You've probably been doing it since before then, but that's 10 years of a look at BGP in the APNIC blog. Geoff Huston 3:10 Well, that's it. And the reason why I do it is actually, really, I suppose, quite odd in some ways. You see, when you try and look at something as large as the Internet, you typically only see what's close to you. You know your neck of the woods. Now, it's like looking out at the weather. I know precisely the weather, where I live within a radius of a few kilometers. You know, I can tell you what it's like, but where you are, George and you don't live close to me, it's kind of like they have a clue. And how would I directly find out, well, I have to go there, and it's kind of the same the Internet in so many ways. You know, I could look at the Internet from where I sit, but that's not the same as the Internet from where you sit and so on. But there's one thing, and it's kind of magic that brings all the Internet to you, all of it, and oddly enough, that's the routing protocol, because when I announce a route into the routing system, hi everyone, you can reach 10.0.0.1, from here. The job of the routing system is to actually carry that little announcement everywhere. So wherever you might be, if you look at the routing system, you'll see me. You'll see my routes. And it kind of works the other way too, that if I turn on a BGP router and look at it, I get to see everybody's route all at once. George Michaelson 4:32 Now you get to see them. But if I were sitting in Paris, wish I was. I'm not, and I turned on BGP and I looked at routes, I'd expect to see the same thing you see. But surely there's some difference between how I see it, some quality in it that we see slightly different versions of the same thing, like we're looking at the same elephant, but from different sides. Geoff Huston 4:56 Well, to some extent you see. In terms of saying, I advertise an address prefix, that address prefix is pretty constant, no matter where we are. So the job of BGP is to carry that routing advertisement across the entire Internet fabric, and do it quickly. But the other job about BGP, which is kind of obvious if you think about it is to avoid the formation of loops. You see, if Alice tells Bob tells Carol tells Jim tells Alice, then that stuff is going to start looping. And effectively, BGP needs to know, because BGP is just a rumor protocol. Whenever you find something, you just tell your neighbors you need to have a really neat way of figuring out. Hey, hang on a second. I've been told that I don't need to get told it again, [George: right] And you need loop detection. Now, the way this works in BGP is the Internet is not just one vast, humongous thing. It's actually a connection of we call them autonomous systems, discrete sub networks, and they're typically, typically geographically limited. George Michaelson 6:09 That word autonomous is it's kind of interesting, isn't it, because it doesn't really crop up very often in normal conversation, and it has different meanings depending on what you're talking about like an autonomous robot would be really very specifically about the robot moving freely and not under the control of any external system. Well, BGP systems don't move like that. They're not machines. In that sense. The autonomy word here is, it's interesting. What is it trying to say? Geoff Huston 6:38 Oh, I wish you wouldn't ask these hard questions. Autonomy, I suppose, is actually a synonym, in this case, for independent. You see, what you need in routing, as in any large system, is some form of abstraction. You need to be able to clump together things that are the same and treat them as one that way you can scale because if you disregarded the Internet as a collection of a few trillion atoms, you know, everyone's brain had melt and the system wouldn't work. So we group together groups of connected hosts and their associated IP addresses and call them an autonomous network. Now, what do we mean by autonomous, or, as I prefer to say, it independent? Well, every network has the ability to determine how it connects to the other networks that make up the Internet, and that decision as to how it connects is its own decision. It's not determined by anyone else. It has its own free will, in some ways, its own autonomy. So I might run a network here and say, Well, I choose to connect to service provider Fred. And you might be running a network in your town. You might say, well, oh, Fred's a bit expensive up here. I'm going to connect to service provider Carol, and that independent decision is yours to make, as is mine George Michaelson 8:07 right And I could also choose to say to Carol, this bit of my network, I want you to carry, but I could have some other relationship with a geezer called Bob, and I could choose to tell him different things. Again, it's my decision what I choose to tell the people that I get to carry my bits. Geoff Huston 8:25 Really, really bad idea. Glad you fell into that rabbit trap. Don't do it. George Michaelson 8:30 Don't do it. Geoff Huston 8:31 Don't do it. You see, the assumption in the routing system, in BGP, is that all the networks in an autonomous system are treated the same way. [George: Oh, wow]. You can't hive some of them off and treat them differently. And the reason why is this whole thing around, how do we do loop detection? So let's just work through that. I'm a bunch of networks where I'm sitting and I'm going to be I think I called myself Carol, didn't I why George Michaelson 9:01 it's a good name? Geoff Huston 9:03 Yeah, it's just, in actual fact, technologically, it's just a number 131072, whatever. It's just a unique label that no one else has. It's mine. And when I announce a route to my neighbors, because that's what BGP does. You just talk to your neighbors. You say, Hi, I'm network Carol. And here are a bunch of networks that you can reach in me George Michaelson 9:25 prefixes here a bunch Geoff Huston 9:26 of, sorry, a bunch of address prefixes, 10.1/24 10.2/24 blah, blah, blah. And my neighbors will go, Okay, well, I will take those prefixes and I will add the label Carol. They tell their neighbors, but let's take one particular neighbor, Bob. So Bob learns this route to an address prefix, 10.0.0.1 label Carol, and when he tells his neighbors, Bob says, Well, you can reach them through me, but I don't actually have these addresses. Carol does. Is. So the new label is Bob, then Carol. George Michaelson 10:04 It's an ordered list. You stick your name on the front, and whatever you heard, you tell people with your autonomy tagged on the front. Geoff Huston 10:13 Right. So as each network learns a prefix, it sticks its identifier, what we call an AS number, but you know, it's just a label onto the front of this list and sends it onward. So as networks move through the interconnection fabric, as prefixes move through that interconnection fabric, that list of networks you need to go through to get to it gets larger and larger. Think of it as a snail trail, a history path. [George: Yeah] you know, this is the way the reachability information propagated, so that if you follow the path backwards, you will get to that prefix. George Michaelson 10:50 And the loop part, I'm guessing, is that sooner or later you Carol, you're going to hear somebody giving you this rumor pssst. I heard a road. I heard that Bob can reach this network, and the network Bob can reach is Carol's network, and at that point, you go, Aha, Geoff Huston 11:07 that's me. I'm Carol. I know that even if someone said to Bob, Bob, Bob looks at the part and says, But I'm already on that list. I told someone else, who told someone else, who told someone else, who's telling me I may not have originated it, but it's a loop, right? And so you can't get told stuff you already know by attribute of the AS path. So let's now go back to your fractured example. George Michaelson 11:36 I don't want to tell everyone the same thing. Geoff Huston 11:39 I'm going to tell Bob half my prefixes, and I'm going to tell Jim the other half. And that's okay, but the issue is, in essence, the prefix path, and let's say someone then tries to tell Bob about stuff I told Jim, then you have the potential of causing a loop. But what if someone tells me, you know, loop formation becomes a whole lot more difficult when you start splitting your routes up. Not the best of ideas, [George: yeah] not the best. Typically, you want a single routing policy with a single autonomous network. That's kind of what makes it work best. George Michaelson 12:16 So the difference between I see everything in this world, because the entry point for this conversation is everybody gets to see the same information. The difference is, how does that loop detection, snail trail look to me compared to how it looks to you? The actual thing, the announcement of the prefix, that part, should look the same everywhere, but the trail of ASs that identify how I learned it that might look a bit different. Is that it Geoff Huston 12:44 Well, that's exactly right. So we might look at the same map of a city and it's the same map, but where I sit, if I was then to draw on this map, how do I get to destination? A, destination B, all the lines I would draw on that map are relative to me. You will draw different lines relative to your position on that map, and that's what's going on with BGP. It's basically it's meant to carry the same information everywhere, but as it carries it, it carries the path of how that information was propagated, and that path information varies with each BGP speaker, with each listener, because the path they obtain that information from naturally differs. And so the issue is, what's useful to understand in these aggregated views? Well, a number of things. How big is the Internet? Really big. Okay, is it growing? Well, I guess so. George Michaelson 13:47 You'd think so. People are selling more and more devices that connect to it. So there's more people on it than there used to be. Geoff Huston 13:54 Well, I don't know the number of devices I have in my house hasn't changed. For a few years. I'm replacing, not adding. So it's kind of an interesting question, isn't it, because when you think about it, there is a difference between throwing one out and adding another and adding a new one, or throwing more out than I actually had, you know, shrinking and so networks do, you know, grow, shrink or stay the same. And it's kind of useful information to know why? Aha. Now the next thing to understand is, why do we do routing? We just do no because the issue with the Internet is that it is a Stateless hop by hop forwarding system. Well, that's easy to say, what does it mean? George Michaelson 14:41 Stateless. Geoff Huston 14:42 Stateless is the big key word, and I'm going to contrast this a tiny bit with the old telephone system, which, you know, 20 years ago was an easy thing to do because everyone understood how it worked. Today, it's kind of telephone system. I'm sorry. What was that? Well, bring out your history books, children, and let's go through this. Nonsense. You see, the old telephone system was stateful. What do you mean by stateful? Well, when I tried to make a call, pick up the handset, remember them, dial a number. Remember that, then the network itself would establish a state within the relevant switches to go I'm going to make a set of connected forwarding decisions. Geoff's phone is connected to exchange A and the call that he's making is connected from exchange A to circuit, along this circuit to the next exchange, to the next exchange to reach George. And that state is almost as if it's hard wired, you can imagine little ladies on on big telephone patch, you know, systems plugging in the cords for for the call and for the duration of the call that's now hard wired into the phone system. When I hang up, the plugs are removed. George Michaelson 15:56 So that system kind of has an organic quality that when that system was literally your voice making the volts vary on a piece of copper between you and the exchange. They were literally sitting on a chair connecting a copper plug to connect it to the next link up the chain. It was a static circuit. If you put a light bulb across that, presumably you could make the light bulb turn on, because there was a wire circuit all the way Geoff Huston 16:22 in the early days when we were switching connectivity, yes George Michaelson 16:25 and when we cut it over so that there were radio links and this links and that links, and shifted up in frequency. We kept the model the same, that when you said, I want to make the connection, it was as if there was that full circuit loop, even if bits of it were kind of different. Geoff Huston 16:42 we started sharing the conductor the wire. We're sharing it, and there are a number of ways you could share it, frequency division multiplexing. You know that one carrier signal is a higher frequency and another, and they don't interfere, but the cleverest way of sharing is to actually share bits of time, so I divide the network up into, say, 10 slots, and the clocks are all synchronized, so at time one, it's slot one. Time two, slot two, slot three, slot four. And in essence, I've got slot one for my conversation, and all the packets in slot one get routed in a certain way. Now this is interesting, because the packets don't need to know where they're going. They don't. The network knows where the packets are going. The network routes the call and forwards the call, and in essence, if the network state is there, the call is working. When the network drops that state, nothing happens anymore. The network has lost the knowledge of how to get connection from one end to the other. George Michaelson 17:44 So even with having this sharing, the basic model of how do things know where to go is there's state in the system. It's stateful, it's structured, it's predictable, and it's based on aspects like the structure in that number, that telephone number you were dialing, told the system which bits of state to use and construct in order to make that call happen. That was the old model. But you've said the new model ain't like that. Geoff Huston 18:15 Well, the old model was based on expensive networks and cheap dumb peripherals, which when we talk about humans cheap and dumb, might well be true, and the artifact of the network was the pinnacle of shared technology. So it made sense economically as much as technically. But as we progress with the world of computing, what we found was that over the evolution of time. Computers got bigger, faster, cheaper. They didn't need to, if you will, outsource that problem to a central network. They could do something quite different. They could actually label each quantum of information, each item with, I'm going to George, and I'm from Geoff. So as well as just sending a little bit of the call, they sent a header that said where it's coming from and where it was going to. What did this mean? Well, if every packet is labeled with where it's going to, you actually don't need network state, as long as each switch in the middle knows all possible destinations. Then when a packet comes into George, it looks up its little table and says, I know where George is. Interface number three, out you go packet. And oddly enough, that makes really cheap networks, because you put a lot of the problem back again to the end points of having to label every single packet with you know who it's from and who it's intended to. The only trick is now in every switch. Instead of having state per call, which is expensive, you have to load up all possible destinations and their relative switching position, George Michaelson 19:54 and I'm going to go with BGP is how you trade that information about what exists out there that everyone has to know the path bit is kind of how you know that it's not in a loop, and what you have to do is work out what you want to record in your system about what you're going to do with this address. How are you going to deal with it. That's your local decision. But BGP is how you're told you have a decision to make. Geoff Huston 20:27 right. So BGP is kind of the provisioning protocol. Why is this so good? Well, let's say that there were 20 ways to get between me and you. Well, the routing system will actually make a call. It will select what it thinks is the best route. But what if that breaks? You know, bush fires, floods, some idiot pulling a plug out of the socket. You know, anything's possible, routing systems are able to self heal on the fly, so that, in essence, even in the middle of a conversation, the routing system can reconfigure the network in response to some incident, and the packets will continue to flow. So the networks become a whole lot more resilient and a whole lot more able to be coped with, you know, to cope with the various forms of things that would otherwise completely disrupt, you know, a stateful network, because there is no state and in essence, all you're doing in the routing protocol is to make sure that the network is constantly talking to itself those switches about who can reach reach what. So the implication of all of this is that in every major switch, it needs to carry a complete set of reachable address prefixes. Many years ago, the Internet had about 20,000 such prefixes. 20,000 32 bit prefixes is 20,000 32 bit words of memory. And in today's world, you kind of go, oh, yeah, boring. Back in the world of the late 80s, that wasn't exactly boring. It was, you know, a reasonable load. George Michaelson 22:01 So this question, how big is the network? I mean, there's the academic exercise. We're sitting down the pub, and we're interested in having a Guinness Book of Records conversation. Is it bigger or smaller you've actually gone to? Turns out, when you're building one of these things, decisions like how big it is affects how you build it. How much memory do you have to put into every active element to hold the things that are going to represent this network? And back then, that size limit was something we might look at and say, That's peanuts. But at the time, that was actually a non trivial decision, that stuff costs money. Geoff Huston 22:40 Well, it costs money. But let's introduce another problem that I have to do that look up inside the time it takes to send a packet. So if I've got a network with a line that's got, I don't know, a million bits per second, then I actually don't need to shovel that many packets, 1000 packets a second. I need to do a lookup in a 1,000th of a second for computers, that's kind of easy, right? But who runs a megabit wire these days? Nobody. All of a sudden, these days, the lines are a whole lot faster. And inside a switching system, I might just not have one line. I might have 100 200 all of which are running at, I don't know, 200 gigabits per second, 200,000 megabits per second. So how many packets per second do I have to cope with in this switch? And the answer is, buddy, a lot, so much that we can't quite make memory fast enough. George Michaelson 23:39 So the thing here is, if you can't look up the forwarding decision you need to make as you're reading the packet in and seeing the bits of the address you're sending it to, you've got to put it somewhere. So you've got to both delay the packet and you've got to have somewhere to put it, and you've still got to have the space in your memory to have the giant table of all the different things you can look up, and that's time and space are not working in your favor. What you want is to get rid of that packet as quickly as humanly possible. Geoff Huston 24:14 So even if you give yourself one packets time, you don't do it bit by bit, but you assemble a packet, you look up the net with their your forwarding tables, you make a forwarding decision and get rid of the packet. You need to do so without taking longer than the time for the next packet to get assembled and be ready to be switched because if you start taking more time than that, you can't get rid of the packets as quickly as they're arriving, bad things happen. So oddly enough, that memory is one of the most expensive items in a router. It's more expensive than the processing system because you need brilliantly fast memory George Michaelson 24:54 special memory Geoff Huston 24:55 the best we can build, DDR4 or whatever. It's not particularly. Particularly special. It's just particularly fast, really fast, [George: right] And the real question is, then, how much do I need? And what it means is I've got to store all those reachable address prefixes in my router. But I'm not just buying a router for today. Hi, the network routing system has increased in size by 10,000 address prefixes. Can I get a new router? Please? Hi, another 10,000 Can I get a new route? It doesn't work like that. Does it? [George: No] you know, that's instant death. George Michaelson 25:33 So what we're getting to here is how big is the network is not just a pub conversation. It affects how you build it, and a second order function, how quickly is it growing? If it is growing is important, because you need to know how big it is and how big it's growing to plan forward for when you're next going to buy a unit of expensive equipment, [Geoff: right] And this is some of the most expensive purchasing you're going to make, Geoff Huston 26:00 yep, and you know, we've scared ourselves completely in years gone by. The first of the shocks was actually in around 1990 just as the Internet was taking off, and it was pointed out that the growth of the routing system was going to run out of available addresses within a couple of years, oops, but equally, was going to run out of viable routing in about the same period. We needed to do something pretty quickly, because at that point, the Internet would have outgrown available hardware. We just couldn't route it anymore. Now the whole thing about addressing is the V6 story, and so on. The thing about routing was a bit curious, and the way we solved it, oddly enough, was to get rid of the implicit structure in addressing. And got to a point what we call classless routing, that instead of having big networks, medium networks and teeny tiny networks, George Michaelson 27:02 the Goldilocks problem, Geoff Huston 27:03 Goldilocks problem, we eventually said, look any size you like, in terms of the number of bits, you decide how many or how big each network is, how big each prefix is. If you're a big network, one prefix will cover millions of individual hosts. If you're a teeny, tiny network, one prefix will cover a couple of 100 hosts. George Michaelson 27:27 So this actually is quite an interesting point, isn't it, because people think big size is about how many host objects you can address. Oh, I've got a bigger network than you. I've got a slash 16 with 65,000 I've only got a slash 24 with 256 but if each of those are announced as one thing in routing, the big one has a cost of one route, and the little one has a cost of one route, their cost is the same, which means the efficiency of making an announcement is different. The slash 16 is a more efficient announcement. Geoff Huston 28:04 Oh, yes, bigger networks. Yes, it's the same cost in your router. And you know, if the routing economics folk had their way, you know, the world would be a couple of 100 massive routes. But it's always a compromise between the elasticity of routing, what you want to achieve the functionality and its cost. George Michaelson 28:23 So you've been looking at this for quite a long time, at least 10 years in the APNIC blog and on potaroo, but probably before. And you've also said there's been a remarkable consistency across that time that it's been an up and to the right cost in terms of how much memory, how much speed, how much computing power is needed to make this machine fly, and consistently, we've been able to meet that market demand, even though we've had some moments like in the late 90s, where we panicked, there hasn't been a catastrophic failure to achieve the necessary growth. Are things carrying on the same way? Are we in the same story? Geoff Huston 29:04 Well, you know, it was sort of a refined question, because at the same time, dear old Moore's Law and the refinement of chips was actually making sure that we could cope with some growth and keep our costs constant, so as long as the network wasn't doubling in size every say six months, that would be a nightmare, if it's doubling in size every say four years. Yawn. Nothing to see here. Why? Because as long as Moore's Law keeps on doubling the capacity of silicon about every 18 months to two years, then it's not a big deal that the network is growing, as long as the costs are not going up as well. If I can keep on acquiring equipment that will constrain that problem to a fixed cost point, I can accommodate growth in the routing system. So we've had our issues. The first one was the Internet boom of the year, 2000 you know what happened then? Well, everybody, everybody. It was the AI of the turn of the millennium. It was the the Bitcoin. It was everyone wanted a web page. Everyone was piling in. And the routing system basically doubled inside of six months. And a lot of us got worried. We were discussing, did we need new routing systems? Did we need to actually change the way routing worked? And there were a few proposals around that period, around 98 around the whole thing of tunneling and forwarding in different ways. [George: Yeah] What fixed the problem? The bubble burst, and a whole bunch of people stopped routing. And some sanity prevailed, and it's kind of well, back to our beds. Everything's fine, and the Internet kept on growing, but growing at a rate that was under Moore's law. George Michaelson 30:52 So there was an odd moment for me there. This is around the time that I finally had people say, We'd rather you didn't have privileges to control routers, and the reason was, we put a protection limit in our router config to say you haven't got memory, don't run if you see more than a certain number of things. And we'd forgotten we put that ceiling in and that growth trend you're talking about, well, it had kind of flattened out, and we'd got away with this, but things were ticking along at a slow rate. We kicked over that limit. And the way the limit worked was it said, you hit this limit, drop everything on the floor and start again. And we were basically getting about 72 hours of stable routing before it would creep up to receiving that last little message coming in, and then it would go too much and drop everything. And we were sitting there going, Why is it every three days our routing fails? When they found that it was a fixed limit that we hadn't documented and put in the config, they looked at me and they said, Yeah, you shouldn't be doing this configuration anymore, buddy. And they took my permits away. Geoff Huston 31:58 Wouldn't have been the first to have got caught like that. You certainly weren't. The last one of the major Australian providers got caught the same way a couple of years ago, insidious problems when the routing system gets bigger than, you know, your available memory, and basically it will shut down. So we track this stuff closely, and certainly one of the reasons is you look at that first order differential of growth, and the real issue is George Michaelson 32:21 rate of growth. Geoff Huston 32:22 The rate of growth, George Michaelson 32:23 rather than just the simple growth. What is the rate of change in this growth? Geoff Huston 32:28 Is it growing faster than our ability to make new silicon to cope with it? So in other words, how does the rate of growth compared to our expectations of next year's chips of the year afters? And, of course, the vendors who are trying to sell equipment to go she'll be great. Mate, she'll work for five years. No sweat. Well, they're trying to build a system that is actually configured not for today's network, but for a network with five years of inbuilt growth. So the real question is, well, what model of projection did you use for that five years of growth? I don't know, mate, well, you got to do something. And a lot of us started to look at this with a lot of seriousness, going, what is the shape of that growth, and how does it affect what we're doing? George Michaelson 33:12 So we've moved a conversation of how big is the Internet from being just a matter of winning to being no it actually affects how we build it to being I don't care how big it is. I want to know how fast it's growing, [Geoff: right] Because I've got to build the box it fits in. And if it's going to get too big for the box, it can be a big box or a little box, but if it doesn't fit, I'm in trouble. Geoff Huston 33:36 So if the rate of growth is linear, it's actually, oddly enough, really easy. Because, in essence, I need to make the box this much bigger each year, and I can project that forward in five years, and it's five times this much bigger. And in some ways, linear is not going to cause anyone to lose sleep. The nightmare is exponential growth, because if this thing doubles in size every year doubles, then all of a sudden I need double the capacity in my routers every 12 months. After five years, two to the fifth, dude, two to the fifth that's 16 times the size. Now, if that's 16 times the size, unit is more than 16 times the cost, then this is looking particularly unattractive. [George: yeah]. So exponential growth is really, really hard to cope with, and it only kind of works as long as the growth is under the ability of the technology to map it. And so the whole question in looking at the size of the routing table is, how are we doing with Moore's Law and chip technology? How much faster and capable are we getting as a trend line, and how fast is the network growing? And do the trend lines work? Because if it's growing faster than our technology, we've got some. Serious problems, and we need to rethink the way we do Stateless forwarding. If it's under that trend line, the answer is, press on. Dude, everything is just fine. We can cope George Michaelson 35:10 Now the marvelous thing about radio is that radio is not brilliant at showing people pictures. So at this point, I'm going to say if you actually find yourself interested in the well, what is really happening? No matter what gets said in the rest of this podcast, you are going to actually have to go and look at the charts, because it's pretty hard to communicate them over the medium of sound. And Geoff has been publishing this information every year, pretty consistently, and each one reflects back on prior history. So you don't need to go looking at all of them, although you can do because it's interesting. You've pretty much carried that time history, that time series of growth and change and rate of change into each version of this story as you publish it each year, haven't you? Geoff Huston 36:02 Yes, yes. So we started with the move from the rarefied atmosphere of research and, you know, academic endeavor, into the consumer market, dial up modems, DSL. We then had the next big shift when the Internet moved into mobile phones, the release of the Apple iPhone, and it's ilk, and that was a massive impetus. All of a sudden we could number the users in billions, and the size of the routing system was 200,000 400,000 etc. It grew as the networks expanded to accommodate everyone wanted a mobile phone. Everyone wanted TWO damn them. I don't know why, but there are a lot of people, a lot of mobile phones, George Michaelson 36:43 and we also had the rise of middle class consumerism in economies that previously just hadn't had exposure to this technology. So not only were we seeing new classes of use in the developed West, in the G20 we were seeing the tiger economies in Asia with huge populations suddenly getting mass subscription to Internet, huge numbers of new entrants, new addresses, new prefixes, new providers coming into the system. Geoff Huston 37:11 So, you know, in the five years, 2015 to 2020 the size of the network in terms of the number of things that are in the routing system. The number I have to cram into these boxes lifted from half a million to 900,000 [George: Ooh] in other words, a router that was five years old was either extravagantly over provisioned when I bought it, and it was able to still cope, or it wasn't and it's now dead because it couldn't cope. George Michaelson 37:41 500,000 to 900,000 Geoff Huston 37:45 in five years, George Michaelson 37:46 five years. So that's pretty accelerated growth right. Geoff Huston 37:50 Now, in theory, Moore's Law was sucking that up, going, no problem, dude, but it's still a big number. Now, what happened around the time that covid happened? 2020, 2021, is that a lot of things stopped just in their tracks. And one of the things that stopped in its tracks was actually the expansion of this network just stopped. Now. It didn't go flat, but the growth rates were remarkably small. So in that period, 2021 2022 In fact, the five years from 2020 to 2025, it grew by under 200,000 so 500,000 in the previous George Michaelson 38:30 it's a useful reminder that we can talk about what we think is happening because of technology and our beliefs in behavior of individuals. Oh, I'm going to get the shiny new m4 Mac when you grow up to the scale of measuring what economies do, it's not the same as what you or I do. There was an economic component to this. It was just not happening. Because everyone's economy kind of went to sleep for a while, because goods and services were just not happening at scale Geoff Huston 39:01 not happening. So it kind of went to sleep, and the recovery into growth hasn't happened as rapidly in the last year or two. Why not? Well, George Michaelson 39:13 Wow, we've had a we've had a shock event like the GFC in prior times, but instead of a post slump, boom and a recovery to the old rate, you're saying, you're actually seeing a change in the trend line. As the economy comes back into play Geoff Huston 39:30 It came back more slowly, and has been going more slowly. It's been growing, but the growth rates are 30,000 40,000 new networks a year, dramatically lower. It's as if this market is now saturated and that I can fit no more mobile phones into my pockets. My car has so many SIMs that it doesn't need another SIM card connectivity is not, if you will, in great demand to increase. George Michaelson 39:58 So we've seen this. Over time, time and time again, the world lives with pianolas, and some guy called Marconi invents this thing called Radio, and there is a massive surge in uptake of homes having radios, and a drop off in homes having pianolas. But then every home has a radio, and sales in radios kind of tips off. And around the 1950s some smart dudes invent TV, and there is a massive surge in people buying and selling TVs, until things tip off, and then they invent color TV, and the story repeats that technology adoption curve. It's always had a quality starts kind of slow, amazing midlife crisis period when things are going gangbusters, tails off a bit as the market saturates, you're seeing that kind of a shape of behavior in the system. Geoff Huston 40:50 We are, you know, we've all got our devices, and these days I don't replace them. In fact, I might replace them, but I don't buy more. I'm not adding to the load as an individual consumer, and most of us aren't. And so the network is kind of tapering off in its growth trajectory Now, throughout all this conversation, and there is still growth, but I'd say it's sort of linear. We've had a conversation exclusively about IPV4, and it's reasonable to say right now that a linear projection of around 34,000 entries a year is a decent projection for the next five years for V4. But wind back to 1990 we thought we had a problem, and we convinced everyone, including ourselves, that we did. We needed to have a new protocol. It needed to have more addresses. It's called V6, and V6 has been slowly getting deployed for, you know, the last 30 years. George Michaelson 41:52 That's a long time and slow here means we have not stopped needing v4 we've just needed both, both v4 and V6 Geoff Huston 42:03 for a long time. And by long I mean up to around 2018 2015 2018 the size of the V6 network was under 50,000 routing entries compared to v4 it was tiny, insignificant. And in terms of its value, I think it was relatively low. It was still the domain of the enthusiast. It wasn't a threat. George Michaelson 42:27 I do want to say, as a slight contrarian on this, those prefixes were phenomenally efficient statements, because, to the extent you had six, you didn't need to announce lots of itty bitty little pieces, because you've been given lots of itty bitty little pieces of address, you had one giant block, and you were able to announce efficiently all of your six with one prefix announcement. So I agree the rate of growth was slow, and I agree it was a bunch of enthusiasts and non economically relevant, but it was very efficient. Geoff Huston 43:02 Human Nature makes fools of us all, George and including that last statement, [George: uh-oh] what we call fragmentation, and it's a typical technique in routing, because you're trying to achieve a number of things, including the quality of the service, defense against attacks, all kinds of things that you tend to announce a big covering aggregate. We call it. Here's all my addresses in one prefix. And then you also say, here's a little prefix. Here's a little one, here's the one. You take out the elements and separately announce it. It's as if you say, here's a car and here's wheel number one, wheel number two, wheel number three, shock absorbers spree. George Michaelson 43:42 Oh, you do the thing that 30 minutes ago. You said, Don't do that. It's a bad idea, right? Geoff Huston 43:47 You announce all the components. And if I look at the v4 routing table, which is currently a little over 1,100,000 one half of those announcements are actually covered by others already. [George: Wow] In other words, 50% to be perfectly frank, 52.4% [George: Wow]. Of all announcements in v4 are more specific. They don't tell you how to reach new addresses. They just qualify previous reachability. George Michaelson 44:15 So hit me with the clue. Stick about V6 and how much that happens. Geoff Huston 44:20 You're saying, well, V6 makes this so much better, doesn't it? V6 has kind of got all this sussed but one big announcement, I can actually announce a whole bunch of things that don't need to do more specifics there. George Michaelson 44:34 I was believing it. So how many are there in V6? Geoff Huston 44:37 Well, 58.7% are more specific. George Michaelson 44:41 It's less efficient. Geoff Huston 44:44 And in actual fact, even though these addresses are remarkably big, a lot of people announce what we call a slash 32 which is the kind of bundle they typically get from their Regional Internet Registry. And then they get out the chopping board and go. Slash 48 slash 48 slash 48 [George: Wow] tiny little bits and two thirds of the routing system are actually slash 48 in V6, George Michaelson 45:08 so you slip through there history. We were sitting at around 900,000 as of today, v4 is now broken through the million mark, and roughly half of them are more specifics, because people are engineering a smarter outcome. You slipped in that V6 for its birth up to about 2018 there wasn't a hell of a lot of it. What if the numbers done since then in absolute size Geoff? What's 6 at if v4 is it over a million? Geoff Huston 45:37 Well, V6 is now kicking 200,000 entries. [George: Oh...] well, that's that's okay. It's less than a quarter of the size of the four well, from memory terms, think again, dude. Because, like it or not, each entry is 128 bits long, four times the size. So if I was mapping out my memory use to be a router, then 200,000 entries in V6 takes up much the same real estate, [George: oh, wow] as just under a million entries in v4 George Michaelson 46:11 so we've doubled the memory burden. If you needed a million memory costs to do four because of dual-stack and V6 inefficiency, you need 2 million. Oh, wow. Geoff Huston 46:23 Yeah. No, we need 2 million. [George: Wow]. And because there's enough v4 that no one's brave enough to do six only, everyone's running both. You can't just run a V6 only network. Maybe sometime in the future, that might be the case, but not today. So in actual fact, everyone who wants to sort of play this game, if you're playing the V6 game, you can't stop playing the v4 game. You can elect not to do V6, but you can't elect not to do v4 George Michaelson 46:51 so number of years ago, I wrote a rather silly piece with Emil Aben from ripe. We called it routing Pixies. We imagined every night, a routing pixie went around, counted up how many v4 you needed, counted up how many v4 AT & T needed. And instead of all these random prefixes that you were announcing, we just swap the addresses so you get a single block that's as big as you need to announce. And you can, in principle, handle the traffic in the global network with the same number of players announcing routes for the volume of four with a lot less than a million routes, if you had a routing pixie to swap your addresses, and I have, I have wondered if we shouldn't open up a new kind of trading market, not to get new addresses, but to swap them with the people who've got the other half of what you hold, so that you can have one prefix instead of two. Are we ever going to see any efficiencies come into this Geoff? Geoff Huston 47:50 Well, it's, it's weirder than that. George, it's always weirder. I'm a big router. I have, let's say, 10 exterior connections, which for a big core router, is actually a lot. Most folk have one or two. So I have this routing table that has 1 million v4 entries and 200,000 V6 entries, and each of them have a forwarding decision that says, interface one, interface two, interface three or interface four. That's your lot. So I've got this massive table, but there's only four answers, and you sit there and go, Why don't I compress it that if I find an adjacent bunch of addresses, why don't I smack them together? And keep on smacking them together. And in actual fact, you can do what we call fib compression. George Michaelson 48:39 You don't care about origin AS you just care about which port on your box it's coming and going on. Geoff Huston 48:45 How small can I make the decision process? Because small is fast, and there are a number of router vendors that do feature fib compression. And you know you can get that number right down. If you're prepared to go down that path, you're swapping space for processing [George: Yeah] And if space is your major problem, and you got enough processing grant, well, it's easy. Processing grant just fixes it. So yes, people have gone there and have tried to solve it, and there are solutions around. So it's not quite as dire as as we make out. There's always solutions. But I want to go somewhere slightly different. You see, for the last four years, V6 has been growing like crazy. Crazy growing. The growth models for V6, say, two or three years ago, were clearly exponential, super linear every year. It wasn't just 10,000 20,000 No, it was growing each year, over over and on the growth, you know. But then covid happened. And what covid did was it stopped the V6 rollout beyond where it had currently reached. And so what we found is that in the developed world, which in terms of V6, is a relatively small world, it. It's the US, bits of Canada, Western Europe, not Southern Europe, bits of Asia, but not all and nowhere else, not in Africa. George Michaelson 50:10 That's the V6 world. It's kind of fitted that hole [Geoff: spotty] and it's not occupying new territory. Geoff Huston 50:17 It's not the only country. The last big country to move was India in 2017 China is moving. But it's such a big problem. It's movement is incredibly slow, and none of the other big populations so George Michaelson 50:30 India, India was a hell of a move. India was not an insignificant move, was it? Geoff Huston 50:36 Oh, it was breathtaking. But there's a massive population of users in Nigeria, and every last one of them, everything is on v4 and there's no signs of movement. And so what it kind of points out, I suppose, is that almost a bit like saturation and stagnation, the V6 World has kind of got to this point of a meta stable equilibrium, where those that have it have it, those that don't, don't feel they're missing out on anything. [George: yeah] and nothing. Much is changing, and that's reflected in the routing system, because now, when I put the projection models against the last four years, I don't see exponential growth. I don't see linear growth. I actually see a logistical saturation curve that has peaked, and the stats say that the V6 table will shrink over the next three years if it continues the current trajectory. Now that's nonsense. That's just maths lying to us. You know, you can't take that seriously. George Michaelson 51:39 but it does say that it is not super growth. It's not in any sense on it Geoff Huston 51:44 oh god, it's hardly growth at all. All of those growth factors have disappeared from the market, and just a few economies are doing more. SIG s, those that haven't are not doing more 6, those that haven't got it by and large aren't, apart from just a small number of players. And so what we actually find in the large scheme of things, in today's world is v4 you've got about 34,000 entries a year. You've got to cope with it and you and that looks pretty stable interesting, George Michaelson 52:14 and that's set against an existing state of around a million things to deal with. So we're talking a low ish increment in the one to 2% mark on your existing state. Is your burden of growth, Geoff Huston 52:28 right. The V6, I currently need about 250,000 routes in two years time, two whole years, it will peak at 260,000 in four years, it'll still be at 260,000 and the math say in 2032 it will actually be lower. At 2051 it will have peaked and gone down. Now that's maths that's just fitting a polynomial curve against recent data George Michaelson 52:55 but it bears thinking about, doesn't it? Because if we go back to the base economics, what would a telco do in a situation like this, when the logistics of being a telco demands you capitalize your routing every single year to expect doubling in five you have to think a certain way. If you somebody comes at you and says, No, you're probably going to be able to sweat the current technology for longer if you can afford to just do small increments in it with minor upgrades. That's very tempting to say, right, freeze on capital spend. Geoff Huston 53:27 Of course it is. Of course it is, you know, the telco makes money by sort of buying something and sweating the asset for as long as it possibly can. Because, in essence, I've got to buy up front, and then I sell its functionality to users. The longer it can be, sort of placed into an active life, the greater my revenue, the greater my return, the more I wanted to defray getting a new one. You know, the only reason why we used to refresh equipment every few years was to sort of keep the competitors at bay. Moore's Law kept on saying that, you know, the competitors were buying new equipment. It was costing them less. I needed to recapitalize. But again, the other part of the stasis argument is Moore's Law has stopped giving, and if it's stopped giving, next year's equipment's the same cost and the same functionality as this year's I don't need to buy. And what that means is that a lot of these growth trajectories in V6 are slackening off. We've eased off the accelerator. Those that have it fine, no need to go back. Those that don't, don't feel the need to change. And when I say those that don't, I am talking about the entire continent of Africa, large amounts of southern Europe, [George: yeah] all of Eastern Europe and western Asia, you know, yeah, and bits of South America. George Michaelson 54:44 It's going to be different pressures that drive change in their technology basis. It's not going to be the underlying economics or the technology needs to make a network work. Geoff Huston 54:54 Yeah, it's not the fear of missing out, the fear of new markets, you know, the need to move. What we're actually seeing right now in the routing system is an entirely different dynamic, that the infrastructure of the Internet has reached this sort of equilibrium point that says, until we get another technology shock, nothing's changing. Nothing is it kind of sitting there going, I'm okay and I look to be okay tomorrow. That's as good as I get, but that's okay. And the world's attention, all the venture capital, we're not building more Internet, [George: wow] we're building more AI data centers, all the trillions of dollars. And it is trillions of dollars is actually turned into the next global fascination and the Internet, so a problem, dude, we're cool. You don't need to worry. George Michaelson 55:43 We've gone back to maybe where we always should have been. We're plumbing, and plumbing isn't exciting right now compared to other stories out there. Geoff Huston 55:52 George, I always thought we were sewage technicians, even more unglamorous. George Michaelson 55:56 Geoff, that's fascinating. I think people definitely should be diving in to look at your charts and the numbers and get a sense of the dynamics of this story for themselves. This is a moment of change. We're not seeing what we used to see. We got to think differently into the future. I think that's really interesting. Geoff Huston 56:14 Well, it certainly kept me entertained, and hopefully, dear listeners, if you're still with us, you're being entertained too. I'm here. I'm happy to answer questions you might have, or look up the numerous articles on the history of routing that I've written over the years. George Michaelson 56:27 You're going to be presenting on this at the forthcoming APRICOT meeting in Jakarta. Geoff Huston 56:32 I will be presenting on that and presenting at the forthcoming NANOG meeting in North America in San Francisco at the start of February, on the same topic. Because it's not just about routing. It's about the Internet. It's about where it's heading, how quickly and why. And that's why, sort of this study grounds those predictions and some pretty, pretty firm truths about, you know, how you can make these predictions, and the degree of certainty that lies behind it, which I think is actually pretty high. George Michaelson 57:00 That's really great. Geoff, thanks. Geoff Huston 57:02 Well, thank you. George Michaelson 57:04 If you've got a story or research to share here on ping, why not get in contact by email to ping@apnic.net or via the APNIC social media channels. Also remember the measurement@apnic.net mailing lists on orbit is there to discuss and share relevant collaborative opportunities, grants and funding opportunities, jobs and graduate placings, or to seek feedback from the community on your own measurement projects. Be sure to check out the APNIC website for all your resource and community needs until next time you you.