The Actually Existing Internet: Into the Cloud (2010 – today)

Liam Mullally

27th February 2024


Introduction


This is the fourth and final entry in a series of blogs on the changing shape of the actually existing internet. Previous entries have looked at the early development of the internet and its rapid privatisation from the 1990s onwards. From the moment of its privatisation onwards, I argued, the internet seemed to disappear: despite the fact that at the same time its material infrastructures were growing rapidly, maps and descriptions of core infrastructures very quickly fell out of use, while technical imaginaries and metaphors related to digital media’s immateriality spread rapidly. This blog tries to grapple with a difficult question: what, if anything, has changed in the material constitution of the internet since this disappearance, and how might it be rendered visible once again?

One methodology for “mapping” the internet is to infer its AS Level topology. I’ve discussed these methods in earlier blogs – in essence, they involve hijacking internet routing information to record connections between autonomous systems, the networks which collectively form the internet. AS level topology inference can teach us a great deal about the organisation of the internet, but relatively little about its use: we can see which networks are the most connected, but not how much traffic flows through them or what that traffic relates to.

The organisations that predominate backbone infrastructure today are not those you might expect. A location like the Telehouse London complex near East India Dock is especially revealing. Telehouse London is home to a range of digital infrastructure including several Internet eXchange Points (IXPs), the point where different networks meet. Among these is the largest IXP in London, LINX LON1, distributed between a number of buildings in the complex. Over 800 autonomous systems meet here, among them well-known consumer tech companies like Apple, media providers like Netflix, large telecommunications companies like Vodafone, and digital payment providers like Visa. But ranked in terms of their network significance a different set of names come to prominence: Level 3, GTT Communications, Tata Communications, Zayo, Arelion, Telecom Italia, and others. Household names, including ISPs, are comparatively marginal in the unglamorous business of building and maintaining the roads of the internet.

Figure 1: Telehouse North in East India Dock, screenshot taken from Google Maps.

Globally, this picture is similar: at the highest level AS topology data from 2010 to the present indicates a coming to prominence of organisations which won out in the fight to divide up US internet infrastructures in the 1990s and early 2000s (this can be seen through Figure 2 below, which charts the top 100 network organisations by CAIDA’s AS cone size, a measure of network significance, from 2012 to the present). Level 3 in particular – with a history originating with ARPANET contractor BBN – embodies this trajectory. Among the largest AS organisations are also several privatised former national-telecommunications companies, for example Telecom Italia or Arelion, as well as large international telecommunications companies like Vodafone and newer consolidations of smaller private providers such as Zayo. A few large infrastructure projects, in particular submarine cables (e.g. SEACOM’s coastal cable which surrounds the African continent), also appear prominently among the top 100 ASes shortly after their completion. But alongside some jostling among competing firms, what is noticeable even from a cursory glance at the data is a growing stratification between the largest three providers and the rest. Growth in the network overall is reflected not in rising cone numbers across the dataset, but an inflation of the cone numbers of these few organisations, suggesting a general centralisation of infrastructure around their backbones. As of late 2023, these top three network carriers are Level 3, Cogent Communications and Arelion. In the period from 2012 to 2023 the cone size of these three companies doubled, tripled and quadrupled respectively. Concretely, this centralisation marks a departure from the privatised but distributed organisation of the internet which predominated in the 1990s and early 2000s.

Figure 2: Autonomy analysis of top 100 network organisations by CAIDA’s AS cone size (a measure of network significance) from 2012 to present day. Hover for further details, and drag to zoom in further.

It is surprising to see data giants of our age – Microsoft, Amazon, Google, Huawei, Tencent, Dell, IBM, Alibaba – be so insignificant in such an analysis. This is because they do not generally operate infrastructures of the kind AS topology renders visible: network routes and backbones. That Google has recently branched into submarine cables might indicate future changes (perhaps the vertical integration of the internet is underway…), but as I’ve already suggested, AS topology only renders so much visible; it can “map” the internet’s “roads” but tells us very little about the infrastructures they run between. Recent changes in AS topology and the organisations which predominate within it are an important part of the picture, but it is true that they do not match the colossal transformations following the privatisation of the US backbone in the 1990s. For more radical changes we may have to look elsewhere, to changes not rendered visible through AS topology alone.

‘The Hyperscale Datacenter’


In the 1990s, when the internet’s network infrastructure was being rapidly marketised, the same could not be said for computation itself which remained a largely personal endeavour. Indeed, several decades would pass with computation mostly distributed between PCs, specific identifiable servers or massive specialised supercomputers. But what started with networking has since spread to computation: today, the internet’s data processing power is being increasingly centralised into private venues, owned and operated by a few massive players. The key form of this centralisation is the so-called “hyperscale datacenter”; these vast facilities consolidate data-power further by replacing an older, relatively distributed model of datacenter construction with singular campus locations dedicated solely to the maintenance and delivery of data infrastructure. While datacenters often operate services on a local, city or regional level, such “hyperscale” facilities act as hubs above even the national level – Google has just three such datacenters serving East and Southeast Asia, for instance. Many operations carried out on individual devices a decade ago have now moved into these facilities under the guise of “cloud computing.”

Indeed, hyperscale datacenters owe much of their existence to cloud computing, prominently via Google Cloud, Amazon Web Services, Microsoft Azure, Alibaba Cloud and others. It is an irony (or maybe an attempt to mislead) that this centralisation of infrastructure and the so-called “cloud” are so intimately connected; on the one hand we are told that computing has become gaseous, while on the other it is being rounded up into a number of large warehouses. As it is sold to businesses and consumers, cloud computing promises a platform detached from specific hardware, distributed between machines and geographic locations. This architecture is attached to promises of seamless movement between devices, data security, on-demand access to both storage and computing power, and instant scalability. All of these help contribute towards an illusion of immateriality, which is of course only achieved via highly developed infrastructures (both servers and networks) which require massive bandwidth and intense energy consumption.

At the extreme end of this tendency is something like the Inner Mongolia Information Park, reportedly the largest datacenter in the world, constructed near the city of Hohhot in Northern China to build capacity for cloud computing; several media outlets have repeated a claim that the datacenter has an incredible 1,000,000m2 floor plan. While in reality this seems to be a stretch estimate based on the potential capacity of the site rather than existing infrastructure, both trade literature and satellite images suggest that a significant amount of work has been done to expand capacity. Despite this expansion, it is surprisingly hard to find photographs that are not either models or renderings, and these are often in exceptionally poor resolution – the most recent publicity photograph I’ve been able to identify dates to 2016. Locations like the Inner Mongolia Information Park reveal that, far from distributing computing, cloud computing has allowed for previously unseen accumulations of computing power, immense infrastructure projects away from the public eye. Like all good magic tricks it relies on distraction: the computing appears to have disappeared only to reappear somewhere else.

Figure 3: Satellite image of the Inner Mongolia Information Park (2022), taken from Google Earth. At least two additional datacenter buildings have been added to the site since 2016.

An imaginative occlusion


So, “the cloud” does constitute a real architecture, but it also entails both an imaginative occlusion (clouds – the central metaphor – are not solid) and an architectural occlusion, which seeks to deny us knowledge of where data is stored, who it ultimately belongs to, what hardware is being used, etc. That it is so hard to determine what actually exists at “the world’s largest datacenter” (whether or not it deserves the title) is indicative of this problem. The form of the datacenter existed in the 1980s and 90s, but the intensity of their use has increased radically since, especially in the last decade. Despite improvements in efficiency, energy demand from datacenters in Europe today is more than four times what it was in 2000, and continues to grow year-on–year (Avgerinou, et al. 2017); a similar trend can be identified in the number of hyperscale datacenters globally, which have almost tripled in number since 2015.

So, “the cloud” does constitute a real architecture, but it also entails both an imaginative occlusion (clouds – the central metaphor – are not solid) and an architectural occlusion, which seeks to deny us knowledge of where data is stored, who it ultimately belongs to, what hardware is being used, etc.

These massive accumulations of infrastructure and computing power don’t show up directly in AS topology, although they likely indirectly influence its shape. The biggest cloud computing providers, therefore, have played at most minor roles in AS topology over the last decade. But since they occupy significant amounts of space and operate as businesses, datacenters (and especially hyperscale datacenters) produce other kinds of visibility. There are several databases which track the locations of commercial datacenters worldwide (generally designed to draw in customers for related b2b services). They show, unsurprisingly, congregations of computational power in Western Europe, India, China, Japan, across the coasts of North America, in Brazil and in Australia. This information is topographical in a way AS topology cannot be, and (although it isn’t a simple or small task) brings geographical information that might be mapped onto other kinds of topological data.

Figure 4: A datacenter map as provided by www.datacentermap.com, showing the global distribution of commercial datacenters

Distribution and its discontents


An important question is raised by the rise of cloud computing today: is the internet still distributed or decentralised as it was originally conceived? The monopolising instincts of the data giants have been characterised as reintroducing critical points of failure into the web’s ecosystem. There is a security angle to this problem but also a political one; of course if the internet is less distributed it is more vulnerable to attack or disruption, but it is also less democratic, less emancipatory, more easily controlled, more powerful as a tool of surveillance, less vibrant, varied and creative. A number of initiatives from across the political spectrum might be read as reactions to the corporate consolidation of the internet – notably blockchain, which attempts to distribute record-keeping and (in the form of cryptocurrencies, especially Etherium) to escalate the marketisation of the network. We need to be aware of this phenomenon, even if our political response does not end up being the same as that offered by libertarians. Blockchain will not save the internet from capitalism; collective ownership and governance will – be that through public, municipal or community mechanisms.

… if the internet is less distributed it is more vulnerable to attack or disruption, but it is also less democratic, less emancipatory, more easily controlled, more powerful as a tool of surveillance, less vibrant, varied and creative.

One thing on-demand cloud services require is capacity. In his book Gramophone, Film, Typewriter (published in 1986), media theorist Freidrich Kittler predicted an imminent rollout of fibre-optic networks; one contemporary logic which computing has brought about is the real-time, and Kittler is perhaps the key theorist of its implications for both technical media and human perception (see his essay “Real time analysis, time axis manipulation”). It follows easily from his perspective that real-time media will demand optical speed (the physical limit-point of communication), and Kittler has been vindicated in his view that fibre-optic connections would be a defining feature of network expansion, even if the widespread installation of fibre optic cables has taken several decades longer than he imagined (and is not only still in progress today, but incredibly uneven globally). So, faster, wider connections facilitate a greater intensity of communication, and makes it technically possible for data giants to reallocate computing power from the peripheries to the new centres of the internet. In doing so they achieve a greater degree of overreach onto our devices (and into our lives) without offensive depreciation in user experience.

Figure 5: Zayo’s AS topology network cone size 2016 to 2023, which grew by around there times over the period

This rise of fibre infrastructure has occurred in tandem with the arrival of “the cloud”. Looking back to the AS topology data, unlike the emergence of “the cloud”, the growth of fibre-optic infrastructure is very quickly identifiable. Zayo was founded in 2007 as an aggressive attempt to acquire new fibre-optic infrastructure, its continually growing network cone size is an indicator of both the coming to prominence of this infrastructure and its consolidation. In the UK, Virgin Media (originally Croydon Cable) shows a similar trajectory. It has steadily consolidated fibre optic infrastructure via acquisitions and mergers originally for its cable television service (see: CATV manhole covers all over the UK), and has since leveraged these to become the largest fibre-optic internet network provider in the UK.

Figure 6: Photograph of a cable television (CATV) manhole cover, today likely housing Virgin Media’s fibre-optic cable network, photographed in South London.

Even so, an organisation like Virgin Media does not operate a significant amount of infrastructure globally, and is a (lower) tier 2 provider, meaning it pays larger providers to carry its traffic. This is even though it holds a large minority of home access connections in the UK. Any attempt to nationalise home access today without addressing core infrastructure risks being beholden to exorbitant fees from tier 1 and 2 networks operating the backbone in the UK and abroad; even a total capture of British internet infrastructure would leave a new National Internet Service operating as a tier 2 provider internationally (there are no tier 1 British Internet Providers). A more sophisticated vision for a nationalised internet is needed across multiple strata of digital infrastructure – backbone carriage, home access, data infrastructure, software, etc.; any singular attempt to bring one of these things into public ownership may quickly run into issues without attention to the others.

So what hope is there for building the socialist internet? It can help to remember that the capitalist internet is itself a parasite on its predecessors. The first barrier to an socialist project is almost certainly the cloud of opacity which covers existing networked communications systems. In spite of the internet’s ongoing disappearance there remain a number of tools available today which offer partial visualisations of the actually existing internet: there is the CAIDA AS topology data which I’ve described at length in these blogs; I’ve also mentioned in passing, but not explored, commercial mappings of the global distribution of datacenters (e.g. https://www.datacentermap.com/), Internet eXchange Points (IXPS, e.g. https://www.internetexchangemap.com/)  and submarine cables (e.g. https://www.submarinecablemap.com/). Any fight against the disappearance of the internet needs to start by joining these sources together into a clearer analysis of where we are today. Even these blogs have begun to make the dispel some of this, but a more formulaic mapping project is undoubtedly necessary.

One thing this history makes clear is just how much of the public character of the early internet has been lost. Another, is how the rapid privatisation of internet infrastructures preceded the general erosion of its free, open character (at least as originally conceived). At the same time, by revealing some of the enormity of the task of freeing the internet (and computation more generally) from its corporate captors, such an reading also begins to reveal its possibility. There are, concretely, steps that might be made to re-establish municipal projects in the tradition of the Cleveland free-net, or, as has been achieved in Kerala, build state provision of internet access. Here a dialectic offers a useful heuristic: if imagination and action can shrink together, they can also expand together, exactly through their tension and contradiction. If we can identify and claim pressure points in the topography of the network – for its arrangement from the outside, in space – we might begin to form new horizons for its topography – its arrangement from inside, technically organised social relations.

The next question is this: if we could capture the network – what would we do with it? I have some more speculative writing on the way, beginning my thinking on what other kinds of network might be imaginable.


Selected bibliography

  • Paul Dourish, “Protocols, Packets and Proximity: The Materiality of Internet Routing”, Signal Traffic: Critical Studies in Media Infrastructures (Urbana: University of Illinois Press, 2015)
  • Freidrich Kittler, Gramophone, Film, Typewriter (Berlin : Brinkmann & Bose, 1986)
  • Freidrich Kittler, trans. Geoffrey Winthrop-Young, “Real Time Analysis, Time Axis Manipulation”, Cultural Politics 13.1 (2017). Originally published 1990.
  • Maria Avgerinou, “Trends in Data Centre Energy Consumption under the European Code of Conduct for Data Centre Energy Efficiency”, Energies 10 (2017)

Liam Mullally is a CHASE-funded PhD candidate in Cultural Studies at Goldsmiths, University of London. His research interests include digital culture, the history of computing, information theory, glitch studies, and the politics and philosophy of noise. Previously he has worked as a copywriter and tech journalist. He is working on several projects with Autonomy, from skills commissions to policy strategy.