Global Internet infrastructure at NANOG 98

The most recent North American Network Operators Group (NANOG) meeting was held in mid-October 2026 in Miami. The opening keynote presentation was by Telegeography's Paul Brodsky on the topic of global Internet infrastructure.

The big picture story of the Internet's infrastructure since its inception has been one of unrelenting demand growth. Growth rates were exceptionally high in the early 1990s as the core NSFNET in the United States transitioned into a mixed infrastructure with commercial providers set up in parallel to the existing academic and research infrastructure. 100% annual growth rates were common in all parts of the network.

Such massive demand growth rates are exceptionally challenging to manage over extended periods, as they call for a constant refresh of the underlying infrastructure, and from time to time a comprehensive change in the internal architecture of the platform. The overall growth rates have slackened off from this unsustainable peak, but the growth has remained positive to the present day, and the demands we are placing on the network infrastructure have persisted.

We've just completed a major architectural shift with the large-scale adoption of replicated service points and the overlay of content distribution networks. This change allows pre-provisioning content and services closer to where the users are located and matches their profile of on-demand access to content and services. The motivation for this architectural change was to push the network's growth pressures, and associated infrastructure investment costs, out to the network's edge, relieving various growth pressures that were being experienced by the set of operators of the inner core of the network's infrastructure.

This redesign of the network platform has been somewhat effective in managing growth. When we look at the growth in demand for international bandwidth over the past 15 years, the annual growth levels appear to be falling year-by-year. However, today's 30% annual growth level still represents a significant motivation for increases in the total capacity of international bandwidth, as a figure in Telegeography's presentation illustrates (Figure 1).

Each annual increment in demand represents increasing demand for additional in-service capacity, which is now two Petabits per second (Pbps) per year, and will rise to more than ten Pbps per year of additional global capacity in 2035.

We also need to factor in the service life of subsea cable infrastructure of about 25 years. Much of the international cable capacity that was installed at the turn of the century is now demanding replacement. Approximately a quarter of the world's current subsea cable kilometres are expected to be decommissioned and retired by 2030.

Capacity of these submarine cables has been constantly improving, with more fibre pairs in a single physical cable, expanded optical transmission capacity by using L and C band optics, and now using two-core fibre strands instead of single-core fibre. The outcome has been a continuing improvement in the unit cost of capacity, and that is reflected in the market price of leased capacity on the major transatlantic and transpacific routes (Figure 2). Falling unit costs expose more demand, which drives further increases in capacity, and these higher capacity systems can operate at higher levels of efficiency, which results in further falls in the unit cost of new cable systems.

The combination of ageing international submarine cable inventory, increased efficiency of new cable systems, growing demand, and a desire to improve resilience through route diversity is driving ongoing investment in new cable systems. Current announcements of new cable projects amount to annual investment in new cable systems of USD 5B to 6B over the 2026 to 2029 period.

So, where's the money coming from? Google, Meta, Microsoft, and Amazon have lifted their level of involvement in cable projects significantly. These days, 38% of cable investments are made by these four hyperscalers to operate as a private service for the exclusive use of the funding entity. They are also investors in consortium arrangements for a further 25% of cable projects. The remaining 37% of funding has no hyperscaler involvement. This investment picture varies by region. The transatlantic route is now 85% private hyperscaler-funded, which is the most extreme case of the rise of privately owned and operated subsea infrastructure. Google and Meta are the most active in this space, and their portfolio of owned cables is shown in Figures 3 and 4.

The reason for this trend toward private cable infrastructure lies in the high level of concentration in the Internet market. This market concentration has allowed the hyperscalers to accumulate large capital reserves, which allows them to own and operate a private cable system with complete control over the delivery of their service without third-party dependencies. At the same time, a consortium model makes less sense when the major capacity leaseholder controls a massive proportion of the cable's leases. If this single dominant lessee left the consortium, the cable would become insolvent, and the entities that funded the cable would be left with a stranded asset. The rise of the private cable model was a natural outcome of these considerations.

However, these private cable systems are not appreciated by everyone. Some governments are concerned about the centralization of investment in the major international communications conduits by foreign private entities. This situation also raises the question of sovereign control over the associated submarine repair vessel fleet and the existing maintenance pooling arrangements through submarine cable maintenance zones. Many would regard these international cable systems as critical infrastructure, with severe consequences resulting from their disruption. To have these essential communication platforms owned and operated by a foreign private entity can lead to a position of vulnerability.

Today's fibre optic cable system has replaced a prior reliance on satellite communications. The satellite systems in geostationary orbit were used to support all forms of international communication from the 1960s. However, this geostationary orbital plane is a limited resource, with 569 active spacecraft parked in orbit at present. Bouncing data packets off of a spacecraft that is 36,000km away results in extended delays. When undersea fibre systems were introduced, the major users shifted many of their international communications to use them as quickly as they could in those parts of the world they reached.

Communications that remained on geostationary spacecraft were services to otherwise unreachable destinations, and broadcast. Continual refinements in signal processing capabilities, and the development of reusable rocketry, supported a new generation of physically small spacecraft placed into Low Earth Orbit (LEO). Because these satellites have a small footprint, LEO systems require many thousands of satellites, with a user's communication switching from satellite to satellite every few seconds. The result is a system that can sustain data throughput at hundreds of Mbps.

Laser links between spacecraft allow packets to be relayed across the spacecraft constellation, covering oceans with a sparse set of ground stations. To a user with an antenna, the capacity and delay of LEO service is comparable to that delivered via terrestrial broadband infrastructure.

SpaceX has spent over USD 20B on its Starlink platform, and Amazon's LEO platform will cost the company USD 20B. In comparison, in the three years 2026 - 2029, all new submarine cable systems will cost a total of USD 22B.

In Q3 of 2026, Starlink reported quarterly revenue of USD 5.25B and a net profit of USD 2.4B. That's an annual profit of USD 9.6B on an outlay of USD 20B, or a rate of return on capital of 48%. Institutional investors in consortium submarine cables look for a 12% to 14% rate of return, and terrestrial broadband infrastructure has a typical rate of return of between 5.5% and 9%.

The major financial difference between cable and these LEO satellite platforms lies in their market position. Starlink is positioned as a last-mile access system, servicing through a direct retail model, a class of users with specialized requirements in remote or otherwise underserved locations. As a retail provider of last resort with an effective monopoly in this sector, Starlink can currently charge a monopoly premium for its services. As a retail provider, it does not have to share this revenue with a collection of retailers. The capital barrier to competitive entry is around USD 20B, which eliminates many forms of competition. A submarine cable system has a lower barrier to entry, where the cost of a cable is USD 100 to 400M, and the lower rates of return reflect this higher likelihood of competitive service.

The Amazon LEO service has the potential to disrupt Starlink's effective monopoly, but it remains to be seen what Amazon will do in terms of positioning its service into markets.

LEO systems and submarine cables service distinct markets, and are unlikely to encounter points of direct competition. LEOs fill a niche of last-mile service provision in demanding locations and scenarios, while submarine cable systems offer a high-volume, low-risk platform for bulk data movement in the core of distributed data platforms. If anything, they are complementary rather than competitive offerings.

Paul's NANOG 98 presentation on Global Infrastructure can be found online.

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.