NetworkingSubsea

Impact of AI on subsea cable demand: capacity forecast

Analyze the impact of AI on subsea cable demand, including private network expansion, lit versus potential capacity, and bandwidth growth forecasts.

, AI Researcher5 min read
In this article6
  1. What AI traffic asks of subsea capacity
  2. Announced cables, routes and owners
  3. How capacity demand is actually forecast
  4. Why cable capacity data is scarce
  5. Reading a capacity claim critically
  6. Sources
A cross-section of a fiber-optic subsea cable resting on a stylized ocean floor, with glowing data streams connecting international data center locations.
A cross-section of a fiber-optic subsea cable resting on a stylized ocean floor, with glowing data streams connecting international data center locations.

What AI traffic asks of subsea capacity

The mechanism linking AI to subsea demand is synchronization-intensive communication. A July 2024 survey of distributed LLM training systems identifies communication between distributed workers, alongside computation and memory, as a first-order constraint on training efficiency[1]. In one emulated geo-distributed training setup, published on arXiv in June 2026, with a 22 ms round-trip time, effective throughput over the spine links fell to roughly 800 Mbit/s, turning communication into the binding constraint[2]. Because AI inference is largely cloud-based, TeleGeography treats major cloud-region locations in the US, Europe and Asia as the key nodes for future subsea and terrestrial capacity growth[3].

The measurement basis for the networking side is published specification revisions, lane rates and switch radix, not vendor performance claims. The Ultra Ethernet Consortium publishes its 1.0.3 specification for download from its site, and the UALink Consortium publishes its specifications, including the 200G variant, on its public specification page, both retrieved 23 September 2026. A specification is a published capability, not a deployed product, and the recurring error on this topic is roadmap silicon quoted as if it were shipping. Within the datacenter, how GPUs communicate in an NVIDIA training cluster sets the scale-up fabric, and the corresponding limits are covered in GPU communication limits and in cluster networking requirements.

  • Distributed training across datacenters, where synchronization traffic must traverse long-haul links
  • Federated learning, which exchanges model updates rather than raw data
  • Spatial-temporal load shifting, as operators move compute between sites to follow power availability
  • Legal constraints, including export controls and data sovereignty requirements, which split workloads across jurisdictions

That factor list comes from a TeleGeography and Ciena webinar on AI and submarine networks, published in January 2025[4]. It is a demand-side framing, not a measured traffic split: TeleGeography states plainly that AI-specific demand cannot currently be disaggregated from overall bandwidth growth[3].

Announced cables, routes and owners

The ownership shift is measurable. Content and cloud networks accounted for 75% of all used international capacity globally in 2025, and TeleGeography forecasts a nine-fold increase in content-provider bandwidth between 2025 and 2035[3]. On TeleGeography's year-end 2024 data, these networks already accounted for at least 80% of used bandwidth on the trans-Atlantic, trans-Pacific and intra-Asian routes[5]. The newest private route is AWS Fastnet, a dedicated transatlantic system between Maryland and County Cork, Ireland, with a design capacity exceeding 320 Tbps, announced in November 2025 and planned operational in 2028[6]. For scale, TeleGeography tracks more than 600 active and planned submarine cables and over 1.5 million kilometres in service as of early 2026, and its submarine cable FAQ gives the MAREA cable as capable of carrying 224 Tbps[7].

ProviderUltra Ethernet Consortium roleUALink role
Alcatel Submarine Networks––
Arista NetworksSteering member–
Ayar Labs–Contributor member
BroadcomSteering memberContributor member
CienaContributor member–
CiscoSteering memberPromoter member
Coherent––
Credo TechnologyContributor memberContributor member
Juniper Networks–Contributor member
Marvell TechnologyGeneral memberContributor member

Membership tiers are as listed on the Ultra Ethernet Consortium member page[8] and the UALink Consortium member page[9], both read on 30 September 2026; a cell reading "–" means that consortium's list does not name the provider. Alcatel Submarine Networks appears in this set as the subsea supplier and installer side of the market rather than a fabric-silicon vendor; the other named providers sit in switching, optics and interconnect supply.

How capacity demand is actually forecast

Why cable capacity data is scarce

Cable capacity data is scarce by construction. TeleGeography publishes its capacity data and forecasts through its Transport Networks Research Service[3]. Public connectivity statistics, such as the ITU's reporting on connectivity in the least developed countries, measure access and usage rather than per-cable capacity[10]. The two measures differ materially:

  • Potential capacity: the total a cable could carry if its owner installed all available transmission equipment at both ends. This is the metric most cited in the press.
  • Lit capacity: the bandwidth actually running over installed equipment. Owners rarely buy day-one equipment for full potential because it is expensive, and instead upgrade gradually as customer demand dictates.

Both definitions are TeleGeography's own, from its submarine cable FAQ[7]. The practical consequence for forecasting is that a cable's headline design capacity says little about near-term supply; realistic forecasts follow the upgrade path, which is tied to observed demand rather than day-one potential.

Reading a capacity claim critically

  1. Ask whether a quoted figure is potential or lit capacity, and which it is.
  2. Check the date on the figure; capacity statements age as equipment is added.
  3. Separate forecast from measurement. TeleGeography's forecast of 24% compound annual growth in used international bandwidth between 2025 and 2035, implying a doubling roughly every three years, is a projection, and its publisher states that AI-specific demand cannot be disaggregated[3].
  4. Treat a published specification as capability, not deployment, and check whether shipping products exist before quoting lane rates or switch radix.

For a reader sizing the market, the defensible position is therefore a dated, sourced pair of numbers: the 75% content-provider share of used international capacity in 2025 and the 24% growth forecast to 2035, both attributed to TeleGeography, with any AI-specific claim kept qualitative until a publisher separates it from general bandwidth growth.

Sources

  1. arxiv.org
  2. arxiv.org
  3. resources.telegeography.com
  4. resources.telegeography.com
  5. resources.telegeography.com
  6. aboutamazon.com
  7. www2.telegeography.com
  8. ultraethernet.org
  9. ualinkconsortium.org
  10. itu.int

Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score

Related reading

All Networking →