TMT in Australia: An Introduction

Three trends are shaping Australia’s digital landscape more than any others:

• the adoption and emerging regulation of AI;
• the continued expansion of large-scale digital infrastructure investments; and
• industry shifts from the entry of LEOSat networks.

AI adoption and regulation

Australians are enthusiastic early adopters of technology, and AI is no exception. Businesses and governments are racing to deploy and experiment with the latest AI tools (primarily generative AI). At the same time, executives have been wary of making decisions that embed path dependencies that might cause problems under future regulation. So the Australian government’s AI regulatory announcement in January 2024 brought some welcome clarity. It not only provides a degree of certainty, but it also seems measured, helping decision makers to push forward with their AI ambitions.

The January announcement was an interim response to the Australian Department of Industry’s discussion paper six months earlier, on Safe and Responsible AI in Australia.

In contrast to the forthcoming EU AI Act, the Australian government is not proposing sweeping legislation that addresses AI as a singular topic across many domains. It is also not seeking to set up a comprehensive scheme to categorise every use of AI into a pyramid of risk levels from low-risk through unacceptable risk.

Instead, the Australian government is proposing more targeted legislation – building on existing laws that already regulate high-risk settings or introducing new domain-specific laws, to ensure they address the new dimensions of risk introduced by the deployment of AI. It is a proportionate approach that has several strengths to recommend it, as follows.

• By regulating only those areas where there is an identified risk, it limits regulation to domains where there is a proven need, allowing a great degree of freedoms to explore in other areas. With AI at a nascent stage and its upside potential not yet apparent, putting the brakes on innovation would be a terrible missed opportunity. Google, Amazon, PayPal, Uber and Canva might not have succeeded against powerful industry incumbents if their opportunities to innovate had been harshly curtailed by well-meaning but too-early legislation.
• Regulating high-risk areas without banning the use of AI outright will allow responsible actors to innovate in a socially sanctioned and transparent manner. The alternative in a global market would be Australians vacating the field or becoming ever-less competitive against their international peers. This has already happened in the media space, with increasingly convoluted regulatory frameworks being introduced to prop up flailing Australian companies.
• By adopting a domain-specific approach, risks can be regulated in a technology-neutral way. For example, the EU AI Act proposes to ban AI practices used by public authorities for social scoring. While there may be widespread consensus that social scoring is unacceptable, it is unclear why the practice would only be banned when AI is involved.
• Domain-specific regulation should also help minimise or avoid red tape from regulatory overlap. This has been a recent issue with overlaps between telco-specific network security legislation and economy-wide security of critical infrastructure legislation.
• Finally, embedding AI-driven rules in domain-specific regulation should help domain experts to quickly integrate the new rules into their existing risk and compliance frameworks, and to shape the new rules to be responsive to industry and process-specific considerations in their domain.

Of course, these benefits could be diluted or lost through poor implementation. Much remains unknown.

An initial list of high-risk settings in which AI might be regulated includes:

• recidivism prediction in law enforcement;
• employment applications;
• self-driving vehicles; and
• robotic surgery.

This list is from the government’s announcement, which notes that other areas might also be considered high-risk. The final list of areas for regulatory action is as yet unknown; even when more information is available, it is reasonable to expect that new areas will be added over time, as new risks are identified.

When specific AI-driven regulation is ultimately introduced, the government has flagged that it will likely take the following forms.

• Frontier model-specific regulation, entailing specific obligations on cutting-edge AI systems development, deployment and use. This is likely to affect only large hi-tech companies and a small number of educational and government agencies specialising in the field.
• Testing obligations, requiring testing before and after deployment, to ensure AI systems in high-risk settings operate as intended.
• Transparency obligations, requiring a range of measures, such as notices to advise users that they are engaging with AI systems in high-risk settings and labelling of AI-generated content in high-risk settings.
• Accountability obligations, requiring safety officers (much like privacy officers) and training for employees of organisations that develop or deploy AI systems in high-risk settings.

The government has flagged further consultation and expert agency work to develop various guidance tools to assist businesses and government agencies with safe deployment. It has also noted that it continues to consider the impacts of AI in its ongoing “business as usual” legislative reform agenda, citing ongoing updates to the Online Safety Act and privacy laws.

For now, most decision makers can continue to experiment with AI and adopt it within their businesses, applying usual best practice risk management frameworks and adopting a few key principles, as follows.

• Develop and document rigorous testing of AI deployments, and ensure testing is ongoing or regular, much like business continuity testing. This may require external validation.
• Label the use and outputs of AI, both within organisations and externally. This need not be onerous – the ChatGPT interface simply says “ChatGPT can make mistakes. Consider checking important information”. If external stakeholders like customers are likely to be unhappy with outputs directly from an AI system (eg, automated loan decisions), this is an indication that:
1. it may be an inappropriate or premature use of AI;
2. it should be intermediated by an employee who understands the limitations of the system, can explain outcomes where they are right and is empowered to overturn it if wrong; or
3. more detailed disclosures and paths for redress are needed where the external stakeholder is unhappy with the interactions or outputs.
• Take a nuanced and incremental approach to AI deployment. Conceptually, replacing a step in a process is easier and more likely to succeed than replacing the whole process. An incremental approach like this also maximises experimentation and minimises path dependencies. It is easier to swap out vendors or re-architect processes without business disruption.
• Develop in-house expertise to help understand the opportunities and risks of AI tools in the context of the organisation’s specific goals, internal responsibilities for achieving those goals, and internal responsibilities for assessing and managing risk for specific AI deployments in the context of surrounding business processes and interactions. Allocate a specific person on project and management teams to be responsible for the safe deployment of each AI system, and ensure they have proper training.

Risk management frameworks applied to AI deployment should take the following into account:

• design-making oversight (noting that directors and officers must ultimately be able to understand and agree with decisions made by, or on behalf of, an organisation);
• interaction and output oversight and training (noting that embarrassing or illegal acts or statements are no less damaging whether made by an employee or an AI tool);
• privacy and data security;
• data sovereignty, control and availability (especially where compute loads and storage are offshore or in cloud environments controlled by third parties); and
• the security of critical infrastructure (including “safety brakes” but also fall-back systems and processes that can be invoked if automation fails).

These approaches can be augmented, particularly in high-risk settings, as further regulation and best practice guidelines are developed over the coming year. The government has not announced specific consultation or legislation timelines.

Digital infrastructure investment expansion

The second of the three most defining trends in the Australian digital landscape is continued and expanded investment in digital infrastructure. Current focuses are submarine and terrestrial backbone cables and data centres, together with their attendant power infrastructure needs. These infrastructure types are seeing a dynamic landscape of investors including hyper-scalers, superannuation and pension funds, and investment banks.

In parallel, traditional network operators continue to do the hard work of driving fibre ever-deeper into fixed and mobile access networks, and a small number of global players are building out LEOSat networks (which are discussed below). This traditional telco activity continues to be dynamic, with continued competition, new technologies, kaleidoscopic network alliances and active government reviews of universal service funding models (discussed below, with LEOSats). However, at a macro level, investors and key players in the access network are reasonably stable.

By contrast, new investors and investment structures are continuing to evolve to fund submarine and terrestrial backbone cables and data centres. There are a few key reasons for this investment dynamism.

• Demand for this infrastructure continues to materially outstrip supply, offering investors good returns. AI is just one of the many technologies requiring ever-increasing data centre capacity. Others include the now mature but not-yet-finished migration of business computing off premises and into the cloud, and the increasing number and size of pipes required for the latest high definition gaming and video entertainment.
• The immense demand means that active participation from the largest incumbents does not preclude the entry or expansion of other sector participants. This creates opportunities for investors with different size investment appetites to participate in a variety of differently sized projects and in independent operators hungry for new sources of investment.
• New investment activity is required by geographic constraints. This includes the partial or complete saturation of existing connectivity hubs like Singapore and heightened geopolitical risks – both of which seem set to drive new digital infrastructure investments towards Australia (and New Zealand), amongst other relatively stable and safe investment jurisdictions.
• As these new opportunities drive the need for investment, large suppliers of data centre inputs, like original equipment manufacturers (OEM), are increasingly seeking to move up the stack and manage ever-higher layers of the compute and storage stack on behalf of their data centre customers (much as mature TowerCos are seeking to manage more of the actives and power on behalf of telco customers). The scale of these suppliers and their financing capabilities can help those data centre operators that do not have vertically integrated hardware manufacturing capabilities compete with vertically integrated players like hyper-scalers. Whether these OEMs can keep up with the demand for cutting-edge chips is yet to be seen. Bans adopted by the US, Japan and the Netherlands on exporting key chip and chip-making equipment to China may help cutting-edge manufacturers like TSMC dedicate demand to OEMs seeking to serve mid-tier markets like Australia faster than would otherwise be the case, but time will tell.

Investors in these infrastructure types are sophisticated. They do not approach their investments as standalone projects, but rather as elements in a rapidly maturing and integrated ecosystem. Submarine cable companies are fusing cable landing stations with data centre capabilities. Data centre operators are seeking to build in proximity to existing and planned backbones and inter-city networks, and to secure backbone network operators as anchor tenants for their builds.

It is not uncommon to see large segments of a data centre dedicated to terminating infrastructure for a major cable system. This, in turn, makes it easier to secure tenancies from other operators seeking to interconnect with the major cable system. And this then needs to be enabled by power and cooling, both of which are increasingly difficult problems to solve.

Power is difficult because existing energy generators and grids are already saturated. Expanding them is not easy – the shift to green generation, the incorporation of batteries and distributed generation are all sapping the focus and balance sheets of energy companies. Added to these supply-side challenges are increasing demand-side challenges. AI workloads can typically require six times as much energy as traditional workloads.

This is leading data centre builders to increasingly build their own generation capacity, often co-located with the data centre itself. Where conditions permit, it is not unusual to have the same amount of land dedicated to energy generation as to the data centre to which the energy is being supplied. The green demands of the largest data centre customers often mean this new generation capability is usually renewable, including solar, wind and hydro energy. All of this, of course, increases the size of the investment and the opportunities for investors.

Many of the same trends are making cooling more difficult. Higher energy use for AI workloads means more heat, and this is exacerbated by AI systems benefitting from being closer together, leaving less room for passive cooling. Chilled water, water and air conditioning systems, captured heat transfer and co-location with water treatment plants are all solutions being adopted where possible.

Despite these challenges, the demand for more data centre capacity in Australia and submarine cables using Australia as both a hub and a waypoint seems set to continue. Continued innovation is expected in response to these demand factors. This points to vibrant opportunities for investors and suppliers alike.

LEOSats

As with AI and digital infrastructure investment, LEOSats are not an Australian phenomenon nor even being pioneered in Australia, but they do have a special relevance to the world’s only island continent. Australia has a similar population to Syria, but to get from one end of Australia to the other you could cross from Syria’s capital, Damascus, to London.

Communication across this sparsely populated place has always been a problem, made harder by a deeply rooted Australian insistence on striving for near-total equity between the city and the bush. In 2009, the Australian government established the National Broadband Network Company (NBN Co), with a mandate to serve 100% of homes and businesses in Australia with broadband over fibre, fixed wireless and satellite technologies and offer “postage stamp” pricing – at a wholesale level, a broadband connection would cost the same in the most far-flung islands off the Australian coast (including the Cocos Islands, over 2,000 km from the mainland) as in the centre of Australia’s biggest cities. This implied a massive cross-subsidy from the city to the bush. The average cost to serve a premises in the city was initially estimated to be ~AUD2,500 (with a full fibre connection capable of delivering 1 Gbps), whereas the average cost to serve a premises in remote areas was estimated to be closer to ~AUD45,000 (with a connection capable of closer to 100 Mbps).

The advent of LEOSat networks has almost inverted the cost calculations. The latest satellite connections still cannot compete with the capacity or speed of fixed line connections. Crucially, however, the networks are global by their nature, so they are built to serve the population of Europe (~600 million people) and, almost as a by-product of the technology, a large portion of that capacity is available over the Pacific (46 million people total, 26 million in Australia). This implies that the cost per Gbps for Australians to access latent capacity in the network is likely to plummet as competition between LEOSat providers develops in the latter half of this decade. NBN Co’s current satellite network is expected to come to the end of its useful life over the same time horizon.

This gives the government an unprecedented opportunity to achieve century-old universal service ambitions. In late 2023, it announced a sweeping review of universal service arrangements, with consultations closing in March 2024. There are parallel processes for refining the statutory infrastructure provider of last resort arrangements, a triennial regional telecommunications review and work being progressed by a First Nations Digital Inclusion Advisory Group. State and territory governments are also progressing trials of LEOSats and other technologies to fulfil public safety, emergency service and service delivery functions.

It seems all but inevitable that LEOSats will play a key role in digital inclusion for remote communities, particularly remote indigenous communities that continue to suffer shocking disparities in education, health and employment outcomes compared to the broader Australian community. They will also help in responding to the natural disasters that have become ever-more frequent. The police force in Western Australia has already incorporated LEOSats into its communications network, which covers a state that on its own is the size of all Western Europe.

The economics of these networks will mean that government funding for such services may change radically from the traditional models in Australia, which were developed to work with very different economic realities. Funding no longer has to assume a single loss-making network that needs to be a heavily regulated bottleneck that is built through public funding and subject to massive ongoing cross-subsidies. With the re-balancing of the relationship to give the provider a more equal voice, it would be natural to see LEOSat providers pushing for greater regulatory harmony with the main jurisdictions in which they operate – at least for operational matters like law enforcement and security agency engagement.

Beyond the realm of government-subsidised universal services and public engagement, the same revolutionary economics also create opportunities for LEOSats. Agriculture uses an incredible 55% of the Australian landmass and is highly mechanised, and mining, which often takes place in remote areas, accounts for 14–15% of Australia’s GDP. Both are likely be eager users of LEOSats, and any material gains may flow through to the wider Australian economy, leading to a generally more prosperous Australia. This too may create opportunities for regulatory streamlining and refinement in favour of LEOSat operators in areas like spectrum management and planning rules for the deployment of VSATs and other terrestrial support infrastructure.

Conclusion

The trends shaping Australia’s digital landscape are not unique: AI, digital infrastructure investment (particularly in submarine and terrestrial backbone networks, and in data centres) and the advent of LEOSat networks are amongst the most important technological trends worldwide. Harnessing these developments successfully in Australia’s unique geography and economy requires a keen understanding of the nuances and implications of the technology and familiarity with the rapidly evolving commercial and regulatory landscape.