At a glance
Australia needs at least two more international cable systems in each of its three directions, east, north and west, on Hamid Maani's estimate, over and above the projects already in the pipeline.
Maani puts a 100MW site in Australia or New Zealand at around US$1 billion before the computing platform, and potentially several times that again for the platform itself.
Cables without a Google or a Meta behind them are struggling to raise money, and Maani names Inligo's ACC-1 and BW Digital's Hawaiki Nui as projects that have sat in development for years.
Firmus underwrote the planned Bernacchi-1 cable to Tasmania, which Maani offers as a good example of how the gap gets closed.
Getting dedicated fibre on the exact route a campus needs may matter more than how much traffic a cable can carry in total, on Maani's reading.
Hamid Maani is managing director of A1 Infrastructure Partners in Auckland. He spent the previous nine years with the Hawaiki trans-Pacific cable and, after BW Digital bought it in 2022, with BW Digital, latterly as chief sales officer. He speaks on the subsea, fibre and interconnect panel at the Sydney Cloud & Datacenter Convention at ICC Sydney on 17 September 2026, and writes here as a guest contributor.
Maani will tell that panel that the next cables stall at the funding stage. Gigawatts of AI capacity are now being planned for sites chosen for power, away from the fibre routes that serve the established data centres in Australia. He says the thing that runs out first is a fibre pair, the two strands of glass a single customer lights for itself. One has to be free on the route a given campus needs.
Maani says neoclouds can turn forecast AI demand into the long contracts that finance new cables, and names Firmus's Bernacchi-1 to Tasmania as a good example. His assessment follows, in his own words.
Cloud and AI demand has drawn new cable investment to Australia
Australia's connectivity landscape has benefited from the growth of large cloud and AI deployments over the past few years, as well as from the geopolitical challenges associated with deploying submarine cables through some of the world's more troublesome regions.
From a demand perspective, the "Great Southern Route" concept has attracted sizeable investment in submarine cables, including the Oman Australia Cable, Tabua, Honomoana and Project Waterworth. It has also contributed to the development of domestic subsea systems such as SMAP and Google's planned Australia Connect interlink, which address missing submarine segments between Australia's east and west coasts.
At the same time, established systems such as Southern Cross, Hawaiki, Indigo and Australia Singapore Cable continue to carry a large share of the region's critical traffic. Southern Cross NEXT has added significant new capacity, but available resources on particular routes, and especially access to independent fibre pairs, are becoming increasingly scarce.
Cable projects are struggling to clear an investor's return test
The truth is that non-OTT cable systems are becoming more challenging to fund because they often fail to offer an attractive risk-adjusted return for investors. Inligo's ACC-1 and Hawaiki Nui are examples of projects that have remained in development for several years. Some projects have secured vendor arrangements and made significant market announcements, with limited subsequent progress made public. The challenge is unusual: there is simply too much perceived risk in investing in submarine cables.
Data centres are moving to power-rich regional sites
The flip side of the coin is the ever-increasing demand being fuelled by gigawatts of proposed AI-factory capacity. The million-dollar question is how much bandwidth will be required if these developments are successfully translated into operational data-centre capacity. Having spoken with multiple hyperscalers and neoclouds, I believe this remains one of the most challenging demand-planning tasks facing the industry.
The industry has, quite rightly, begun shifting new developments closer to available power sources. These locations, however, are not necessarily established connectivity hubs. Major metropolitan markets such as Sydney, Melbourne and Perth are already facing constraints around electricity, water, land and planning. Communities are also becoming increasingly vocal about the impact of large data centres. Future developments will therefore be pushed further into regional, power-rich but often less connected locations.
Voice and video traffic will grow as AI moves past training
Besides, most large AI deployments to date have focused on training large language models. NVIDIA has also dominated the accelerated-computing market, with its DGX systems and HGX platforms becoming something close to the global reference architecture for large GPU clusters. Together with NVIDIA's NVLink and high-speed networking technologies, these architectures have influenced how AI facilities are designed, from rack density, power and liquid cooling to the interconnection of computing clusters.
The growth of agentic AI and inference workloads will change the traffic profile. Humans will interface with AI models increasingly using voice, images and video as opposed to text. Connectivity will become increasingly important and, in some markets, could become an inhibiting factor for growth. The adoption of alternative chips and computing architectures may also affect how AI clusters are designed, distributed and interconnected.
An AI factory finances faster than the cable serving it
So, with exponential demand growth but limited appetite for investment in independent submarine cables, where will the required connectivity come from?
A 100MW facility in Australia or New Zealand could require approximately US$1 billion of investment in the underlying data-centre infrastructure, and potentially several times that amount for the computing platform. The total investment envelope could therefore reach several billion US dollars, depending on the technology and configuration.
From a pure financing perspective, the risk profile of an AI factory may be an order of magnitude greater than that of a submarine cable. Yet the industry appears more willing to finance the compute than the infrastructure needed to connect it.
Submarine cables can take four to six years, and sometimes considerably longer, to progress from concept to ready for service. This creates a potential mismatch between the delivery of AI capacity and the international and domestic networks required to support it.
Firmus underwrote Bernacchi-1 to bring Tasmania more than 60Tbps
The industry must address this gap now. Time is ticking, and greater awareness needs to be raised among developers, investors, customers and governments. Neoclouds will have a particularly important role to play. They can translate forecast AI demand into bankable, long-term capacity commitments, providing the commercial foundation required to finance new systems.
Firmus's investment in Bernacchi-1 is a good example. Underwritten by Firmus, the new submarine cable will connect Tasmania to SUBCO's SMAP system and provide diverse onward paths towards Melbourne and Sydney. It is expected to deliver more than 60Tbps of additional capacity to Tasmania on day one, supporting the development of the state's emerging AI infrastructure.
This is precisely the type of commitment the market needs. Rather than waiting for connectivity to arrive, a neocloud is helping underwrite the infrastructure required to support its future computing platform. Similar partnerships between neoclouds, cable developers and long-term infrastructure investors could help close the connectivity gap elsewhere in Australia and New Zealand.
It is important to remember that the real constraint may not be the headline design capacity of a submarine cable. It may be access to the right fibre pair, on the right route, connected through diverse terrestrial infrastructure to the right data-centre location.
Australia needs at least two more systems in each direction
In my view, Australia will require at least two additional submarine cable systems across each of its strategic directions: east towards New Zealand, the Pacific and the United States; north towards Japan, Korea, and ASEAN markets such as Thailand and the Philippines; and west towards Singapore, India, the Middle East and Europe. These systems would be required over and above the projects already in the pipeline, including APX East.
New Zealand faces a similar challenge. Its domestic fibre infrastructure is strong, but its international cable landings remain concentrated in the upper North Island. If large AI campuses are developed in the South Island, they will require genuinely diverse domestic backhaul or direct international submarine connectivity.
Australia and New Zealand have an opportunity to become an important AI infrastructure region. However, power, compute and connectivity can no longer be planned as separate workstreams.
The AI factories may arrive faster than the cables required to connect them.
Certified Strategic's analysis: is capacity the constraint, or the route?
Maani's contribution ends above. Everything below is Certified Strategic's.
A fibre pair is the two strands of glass one customer lights for itself, and it is the unit a cable is sold in. SUBCO founder Bevan Slattery puts the Australian AI factory pipeline at 3GW by 2028. He puts the international capacity it needs at 75 to 150Tbps. A single sixteen-pair system carries headline capacity well past that, so bandwidth is not what Maani is short of. Deputy Prime Minister Richard Marles put Australia's count at around fifteen subsea cables in May 2026, and Maani is asking for at least six more routes on top.
SUBCO said in January 2026 that only two of SMAP's 16 fibre pairs remained available for sale. Ten were sold and four reserved, on a domestic cable that had not yet entered service. That is the shortage Maani describes, in a published number.
Firmus has contracted up to 150Tbps on APX East. Spread across all sixteen of that cable's pairs the figure would be about 9.4Tbps each, below what a system entering service in the late 2020s carries per pair, so the ceiling buys a share of the cable rather than the whole of it. Our guess is that the next Australian system names a compute buyer before it reaches financial close, because a neocloud can turn forecast demand into the long contract a cable developer needs, and Firmus has now done it twice.