By Tim Toohey, Head of Macro and Strategy

There is genuine excitement over the size of the looming data centre boom in Australia and the tax windfall it could bring for governments. We believe most of the enthusiasm is vastly overstated as most bank and broker research fail to recognise how much of the total projected build out of data centres over the coming decade is imported, and there is a failure to recognise that the economic rent will largely be captured by offshore owners and investors, leaving the broader externalities to be absorbed by locals. There are strong echoes here of Australia’s failure to capture a suitable tax share from the LNG boom. We recommend policy makers take a clearer eyed view of the likely economic, social and tax implications from welcoming data centre investment beyond the city walls without first reflecting on past errors and taking a more thorough inspection.

The data centre investment boom is not an economic saviour

If there is a bias that exists in financial markets, it is to convince itself that we are perpetually in the upswing phase of the economic cycle. This is the phase where the future looks brighter, employment is abundant, incomes are rising and earnings (and tax receipts) surprise on the upside helping to sustain asset price gains. It helps of course if there are new shiny things that can be pointed to that will create a new regime for economic prosperity.

Even if you choose to believe Elon Musk that an AI-era of abundance awaits us all, precisely no one should heed his advice and stop saving and investing sensibly for retirement. It is, however, surprising just how many normally sensible economists appear to be conflating nominal capex announcements for data centres with the future prospects of economic growth in Australia.

The likelihood that the current data centre build out will result in a material uplift in near term Australian economic growth is negligible in our view. The most obvious reason is the misconception that data centres are ‘homemade’. They are not. It is reasonable to look at a large refrigerated shed and think that a lot of local economic value-add was required to build it. However, even if we assume that all the steel, roofing, site works, land acquisition, grid connection and most of the engineering design and specialist labour is domestic in nature (which frankly is a stretch), then we are still left with well over 80% of the cost of each data centre as being imported. The cost of data centre building – the thing we see with our own eyes – averages out to just 2.5% of the total capex. The cost of adding the servers and GPUs is closer to 70% of the total capex bill.

As a note of caution, it is worth noting some of the initial economic analysis that has been published on the impact of the data centre build out by some of Australia’s banks and brokers. It appears that analysts are confusing total data centre spend announcements with the breakdown of data centre construction costs provided by the widely cited Cushman & Wakefield reports, which excludes the cost of the servers and GPUs. This is problematic as not only do the servers and GPUs dominate the cost schedule. They are 100% imported.

Table 1. Typical Australian AI Data Centre Build (Including Servers/GPUs)

Note: Figures are an estimated composite, not a single published Australian benchmark. Basis: Epoch AI 1GW AI data centre TCO model (GPU compute share) rescaled against generic facility cost-category splits (Dgtl Infra/Turner & Townsend), calibrated to Australian $/MW benchmarks (Cushman & Wakefield ANZ Data Centre Cost Guide, ~US$7.9-12.1M/MW).

The consequence is that when using the Clean Energy Finance Corporation’s (CEFC) forecast scenarios of data centre GW installed capacity and capital cost over the next decade as a guide to what the total spend might look like, and calculating the local economic impact on the basis of a construction cost breakdown that excludes the GPUs, the output results in a dramatic overstatement of the economic resources required domestically to facilitate the build.

For instance, the CEFC estimate of the installed capacity for data centres is forecast to grow from 1.5GW at the end of 2025 to 4.7 GW (central case) and 7.4 GW (high case) by 2034–35. According to the CEFC this represents $85-135bn in investment.

Assuming the demand for compute keeps escalating, it is not unreasonable for different banks and brokers to be talking about an investment boom of circa $150bn over the next decade. However, when looking at the headline grabbing announcements on data centre capex, that figure needs to be discounted by around 80% when calculating the impact on actual domestic economic growth once imports are subtracted, and in particular the cost of the GPUs.

Moreover, economic growth is always a rate of change concept and ever larger increases in data centre spend will need to be announced in coming quarters if the contributions to economic growth are to remain. That may occur, but there are already rising concerns that the hyperscalers have overstepped the market’s capacity to fund their ambitions. And even if capital markets do commit to fund the global build out, there is no guarantee they will all survive. Will OpenAI or Meta still be around in five or ten years’ time? Perhaps, but then again under some plausible assumptions, maybe not. There is also an implicit assumption that community and regulatory pushback against the build out is minimal, which increasingly is not the lived experience in many developed world nations.

A figure of $150bn is of course a very large amount of capital investment. If this is spread evenly over the coming decade, then an incremental $15bn p.a. in investment should be seen in the context of total private sector investment of $345bn over the past 12 months (i.e. 4% of total private investment) and an average rise in annual private investment of ~$35bn p.a. Note that $15bn p.a. over the next 10 years would actually be a slowdown in the growth in capex of the information communication and media capex recorded over the past 12 months. In other words, even if the $150bn figure is accurate the acceleration in capex for data centres might already be tapering. Ultimately, this is what matters most for economic growth.

Declining State government funded engineering construction will likely exceed the domestic economic benefits from the data centre build

Fears that the build-out will result in a surge in demand for skilled construction labour and associated construction materials need also to recognise that labour and construction costs represent less than 5% of data centre capex announcements. They should also account for the pipeline of work yet to be done in engineering construction which is already declining as a share of the economy, with this trend likely to continue given that the major States (ex-Qld) will all see their large and lumpy road, rail and infrastructure programmes peak this financial year.

In the battle between declining State infrastructure spend (which has exceptionally high domestic content), and rising data centre spend (which has exceptionally high imported content), the net impact on economic growth is likely to be weaker – not stronger – economic growth over the coming 18 months. A reversal in the prospects for housing construction sector over the next 18 months is also likely post the Budget changes and the RBA’s ever-present threat to further throttle private demand growth.

Chart 1. Private Investment and Government demand as a % of GDP

Source: YCM, ABS, Aug 2026.

Who collects the economic rent?

It is also worth remembering that statistical agencies measure economic growth on the basis of the physical location of the economic activity, rather than who actually owns the asset and who receives the economic rent from that asset.

Australians became somewhat familiar with the concept of real net national disposable income (RNNDI) post the surge in LNG investment during the 2009-11 period. It progressively dawned on Australians that the combination of high imported share of the LNG capex and high foreign ownership meant that looking at the traditional gross domestic product (GDP) growth measure is misleading for what this meant for >99% of Australians who had nothing to do with the LNG sector. RNNDI accounts for the impact of profits accruing to offshore owners of the assets and the depreciation of those assets.

So, rather than focusing on how many construction workers will be required to build the data centre sheds, the real question is what proportion of the data centres are owned by foreign owned hyperscalers, what percentage is owned by operators who build and then lease capacity to the hyperscalers and enterprises (e.g. Equinix, Digital Realty, NextDC, AirTrunk etc.) and what share of the returns are diverted to an increasingly complex financing layer of special-purpose vehicles, joint ventures and private credit funds (e.g. Blue Owl, PIMCO, Apollo, Blackstone, and increasingly pension and superannuation funds).

Typically, the higher-margin economic value — the cloud services, AI models, and software running on the hardware — is in the intellectual property held in the parent company or in a low-tax jurisdiction. It is highly likely the Australian data centre subsidiary will be treated as a cost centre charged back to the parent at a thin cost-plus margin. The practical result is that the host country gets some construction jobs, electricity revenue, some very small operational employment but they also take on all the negative externalities. In contrast, the bulk of the profit generated by the data centres is booked — and taxed — wherever the multinational’s IP and service contracts sit. Hint: that won’t be in Australia.

Moreover, given the very rapid depreciation schedules applied to data centre capex, any ‘pre-tax profit’ that is declared in Australia will quickly disappear post depreciation. Australian residents are still waiting to share in the promised benefits of the LNG investment boom 15 years ago. It should be little wonder why Australia is seen as such an attractive destination for the Hyperscaler’s data centre ambitions: the LNG industry provided the blue-print.

Predictably, Australian politicians have fallen over themselves welcoming the data centre boom. The LNG investment boom may well have been a case of ‘fool me once’; the data centre boom is lining up as a strong case of ‘fool me twice’ and shame should sit with any politician who fails to recognise that the economic impacts are going to be a lot smaller than those cited by the data centre proponents. Hopes of a revenue windfall for governments will largely prove illusionary unless mechanisms are put in place to safeguard against profit shifting and potential labour market displacement.

At least the LNG sector pays the comically inadequate Petroleum Rent Resource Tax (PRRT). Despite LNG export revenues having increased by more than 5x since 2011, the quantum of PRRT paid has remained largely unchanged at around $1.5bn p.a. As a share of LNG exports, the PRRT has shrunk from 12% of exports in 2011 to just 2% in 2025. Compared to Treasury forecasts in 2011 that the LNG sector would generate an $11bn increase in the Petroleum Rent Resource Tax (PRRT) over the next four years, the tax office received less than half of that forecast revenue. The tax take continued to shrink in subsequent years. Today the PRRT represents just 0.17% of Australia’s taxation revenue. In a case of history rhyming, note that the projected installed capacity of data centres over the next decade is forecast by the CEFC as a 4x event.

Chart 2. LNG exports value and PRRT receipts

Source: Commonwealth Budget Papers, ATO, Yarra Capital. Aug 2026.

Policymakers, prepare for a battle of Odyssey-like proportions

There is no equivalent PRRT on data centres for the excess grid electricity and water that they will consume, the negative spillovers of cohabitating near data centres, let alone the potential for these AI factories to contribute to the disintermediation of the labour force and ultimately the tax base of Australia at some point in the future.

Thus far, Governments have moved to insist that future data centres will have to pay their way on electricity and water. That is sensible first step, however, economic theory is pretty clear as to what the optimal solution should look like. A Pigouvian tax (essentially a per-unit charge set to be equal to the marginal external cost of the externality) is theoretically best at aligning economic agents’ private incentives with the social cost of an investment. In the case of pollution or excess draws on electricity and water resources, the imposition of a Pigouvian tax is pretty clear and straightforward. However, in the case where the AI investment and automation is also shown to undermine worker income and system-wide demand, then each firm that substitutes labour with automated processes could also be taxed for demand destruction of its rivals. The idea here is that a downward spiral of labour substitution which undermines system demand growth is avoided while also ensuring negative externalities are appropriately compensated.

Australia major political parties may well shy away from such a solution on the grounds of difficulty in administration or fears of upsetting foreign interests. However, to not move down this path (or something similar) will result in an even worse revenue sharing outcome for Australian taxpayers – and ultimately households than – the aftermath of the LNG investment boom.

This is not something that can be designed in retrospect.

If all this discussion of RNNDI and Pigouvian taxes is too much to cut through, then perhaps a modern-day movie reference will help policy makers: if Australia is Troy, data centres are the Trojan horse and AI and automation could open the gates to a debasement of the tax base which, in the limit, compromise the structure of Australia’s labour market. Australian politicians are currently viewing the horse as a gift and dragging it towards the city walls. It would be better to not repeat the mistakes of the past, whether that be the LNG boom of the 2010s, let alone that of the 12th century BC.