From General-Purpose AI to Industrial AI: Turning AI Capability to Productive Use

John H. Howard [1], 14 July 2026

This working paper is one of a series prepared by the Acton Institute for Policy Research and Innovation to inform the UTS research project Turning AI into Productivity: The Role of Innovation Ecosystems, supported by the Google Foundation. The project's final report will be launched in mid-September 2026.

Introduction

The working paper offers policymakers and advisers a vocabulary that separates AI technology from the work required to make it productive. General-purpose AI names the raw capability, built from largely ordinary computing methods running on very large quantities of hardware. Industrial AI names that capability once an economy has built the complements around it, with four applied segments: production, cognitive, scientific and civic.

The framing arrives at a moment of active policy interest. Industrial AI, and production AI in particular, sits close to current ministerial agendas: the Future Made in Australia agenda, national sovereignty, and the National Reconstruction Fund (Australian Government, 2024). Each turns on the question this paper addresses: how raw capability becomes sovereign industrial strength.

A general-purpose technology appears to have limited productive value on its own. Value emerges when organisations rebuild themselves around it, through redesigned work, new skills, better data and supporting institutions; electricity took roughly forty years to travel this distance. The same logic suggests why productivity statistics have barely moved: most organisations are still developing the complements, a pattern consistent with the trough of a J-curve rather than evidence that the capability is hollow.

The distinction is applied to a live decision: how much weight to place on data centre investment in Australia's AI strategy. Compute is worth having but may not be the transformation often envisaged. Because compute sits upstream of value creation, a country can host a massive amount of it and still build very little Industrial AI.

The complements are assembled in places, through innovation ecosystems where firms, universities, public research organisations, and governments interact, and ecosystem quality, more than access to the technology, may set the limit on how far and how evenly Industrial AI diffuses towards adoption, application and effective use. The policy task that emerges is to convert general-purpose AI into Industrial AI across the economy, more evenly than market forces alone would manage.

An ordinary mechanism, an unusual capability

Discerning audiences remain unsure, or at least curious, whether AI is a break with the past or a continuation of forty years of software development. This paper suggests that the reality is in fact both: an ordinary-seeming mechanism has crossed an unusual threshold of generality.

What we are seeing is largely a continuation of what has come before. A large language model (LLM) predicts the next element in what it sees as a logical sequence, adjusting billions of weights until its output matches patterns in a large body of data. What has changed is the quantity: more parameters, more data, more compute, and more patterns, applied with unusual persistence.

The AI capability, by contrast, is discontinuous. A single artefact now drafts, summarises, codes, translates and reasons over documents, where earlier tools each did one narrow job. That breadth marks AI out as a general-purpose technology in the sense of Bresnahan and Trajtenberg (1995). It also extends to non-routine cognitive work that earlier automation left untouched (Autor, Levy & Murnane, 2003), and may become a method of invention, changing how discovery itself is done (Cockburn, Henderson & Stern, 2019).

Two properties seem to shape everything downstream. A system trained to predict likely continuations produces plausible rather than certified-correct output. This means that checks, review and human oversight are essential components of the work required in putting it to use. The gains, such as they are, have so far been bought with an enormous hardware spend, a coupling to compute, data and energy that bears directly on the data centre question taken up below.

Why the technology appears to have no value on its own

A central premise of this paper is that a general-purpose technology has a limited productive identity of its own. This is the proposition that frames the debate: AI’s productive character appears to be conferred on it by the complements built around it: the organisational redesign, the skills, the data, and the institutional arrangements that surround the bare capability.

The lesson of electricity

Electricity illustrates the point. As a physical phenomenon, it changed nothing. It became transformative when factories were rebuilt around it, when the single steam shaft gave way to small motors driving individual machines, so the production line could follow the logic of the work rather than the geometry of the drivetrain.

Early electric factories simply swapped the steam engine for one large motor, kept the old layout, and saw little gain. The gain came later, once managers understood that many small motors allowed a wholly different organisation of work. David (1990) showed this took roughly forty years after the dynamo. The technology was latent. The transformation was constructed.

The capability is evident. The transformation is built, which may explain why two firms with the same model can sit far apart in what they gain from it.

This leads into a discussion of the Migration of Value Principle, which is addressed in a forthcoming working paper.

From general-purpose AI to Industrial AI

The public conversation has a rich vocabulary for the raw AI capability and almost none for the applied form, which is the use of AI that policymakers and corporate decision-makers really need to know about most. This paper proposes Industrial AI as a description of that capability which manifests when the economy has built the complements that give it a productive character.

The contrast term is general-purpose AI rather than general AI, keeping a clear distance from artificial general intelligence, a more speculative idea about autonomous human-level systems. The word Industrial is chosen deliberately. It connects the term to industrial strategy, and so to the policy machinery now in motion, and the sovereignty concerns that animate both.

Figure 1. The conversion from raw capability to productive use runs through the complements, which are assembled in place-based innovation ecosystems.

The electricity parallel reads more crisply through this lens. The dynamo was general-purpose electricity, a phenomenon that transformed nothing on its own. The rewired factory was electricity made industrial. The forty-year gap David measured was the time it took to move from one to the other. Australia is living in the equivalent gap for AI, which makes the policy question one of acceleration and direction rather than discovery. Table 1 draws out the contrast.

Table 1. Distinguishing general-purpose AI from Industrial AI.

	General-purpose AI	Industrial AI
What it is	The raw capability: models, compute, data centres, energy	The capability put to work through reconfigured workflows
How it travels	Effortlessly across borders	Built locally; accumulates in places through proximity and shared institutions
How it is obtained	Bought from a small number of global suppliers	Constructed through organisational redesign, skills, data and institutions
Time to acquire	Available now, growing cheaper and more ambient by the month	Slow to assemble; electricity took roughly forty years
Where value sits	No productive value on its own	Where the productivity dividend appears
What gets measured	Megawatts, headline investment figures	Adoption beyond pilots; breadth and evenness of diffusion

Four segments of applied AI

Four applied segments follow, and a national approach has to treat them as plural rather than as a single undifferentiated program. Each acts on a different object, requires a different set of complements, and answers to a different part of the policy machinery. Figure 2 sets out the four segments and the diffusion layer beneath them.

Figure 2. The four segments of Industrial AI, each with its own complements, and consumer AI as the diffusion channel that carries capability into daily use.

Production AI is the segment manufacturers already recognise (Lee, 2020), and the one where ministerial interest is currently most active, through the Future Made in Australia agenda and the National Reconstruction Fund's priority areas (National Reconstruction Fund Corporation, 2025).

Cognitive AI may deliver large gains that are harder to measure than on a factory floor or an office. Scientific AI could count for most over the long run, because it changes the rate at which the other three improve. Civic AI is constrained as much by legitimacy and consent as by technology.

A fifth category, consumer AI, identifies the mass-market AI assistant layer. It would appear to work more as a diffusion channel than as a source of value, carrying the capability into daily use and seeding familiarity across the workforce. Table 2 summarises the segments and their principal policy connections.

Table 2. The four segments of Industrial AI and their policy connections.

Segment	Acts on	Key complements	Policy connection
Production	Matter and physical systems: factories, supply chains, energy, logistics	Sensors, operational data, robotics, digital twins, process redesign	Industrial strategy: Future Made in Australia, National Reconstruction Fund
Cognitive	Symbolic and knowledge work: analysis, drafting, advice, design	Workflow redesign, curated domain data, professional judgement	Workforce, skills and professional standards
Scientific	The research process itself	Instruments, datasets, absorptive research institutions	Research funding and institutional capability
Civic	Public administration, health, education, service delivery	Trust, data governance, service redesign	Service delivery, legitimacy and consent
Consumer (diffusion channel)	Daily use across the workforce	Familiarity and habit rather than redesign	Digital inclusion and AI literacy

Readers familiar with the German literature may ask how Industrial AI relates to Industrie 4.0. The original conception (Kagermann et al., 2013) described a broad industrial paradigm centred on the smart factory and cyber-physical systems, with sensors, connectivity and digital twins linking physical operations to computing infrastructure. That program built the connected digital factory well before the current AI wave.

While Industrie 4.0 delivered connectivity, monitoring, and data capture, Industrial AI (sometimes referred to as applied AI) extends the digital factory concept to prediction, autonomy, and closed-loop optimisation (Barua et al., 2025). The two are complements rather than rivals. Regions that invested early in Industrie 4.0 infrastructure, as observed in Dortmund, Kaiserslautern and Heilbronn, could layer AI capabilities onto two decades of accumulated digital plant data.

The nesting should not, however, be read the other way. Industrie 4.0 is a production paradigm; Industrial AI, as defined here, extends more broadly across cognitive, scientific, civic, and community applications

Why the productivity gains have not appeared

This framing explains a puzzle that troubles many observers. The capability is evident, and aggregate productivity has barely moved. There is no contradiction. Brynjolfsson, Rock and Syverson (2021) describe a productivity J-curve, in which output first dips or flattens because firms pour effort into intangible complements that take years to show up as measured output.

Figure 3. The productivity J-curve. Measured output dips while firms invest in intangible complements, then turns upward for those who complete the building.

Australia, like most economies, appears to sit in the trough of that curve. The complements are expensive, slow, and often organisationally painful, and most organisations have barely begun building them. Retraining, redesigned processes and new data foundations all consume scarce resources, including attention, while the measured return arrives later, and only then if the work is done well.

An observer who reports no productivity gain is often describing the present accurately while perhaps drawing the wrong conclusion. A caution also runs the other way: the J-curve may turn upward only for those who do the building, and not automatically for everyone who buys the tools.

The data centre question

A national debate has formed around a simple proposition: that Australia should invest heavily in data centres and aspire to lead in data centre capability. The National AI Plan (Australian Government, 2025) named sovereign digital infrastructure a strategic priority, and in March 2026 the Government issued its expectations of data centres and AI infrastructure developers to shape how that capacity is built (Department of Industry, Science and Resources, 2026).

Capacity is forecast to roughly double, from about 1,350 megawatts in 2024 toward 3,100 megawatts by 2030, on investment estimated near A$26 billion (Department of Industry, Science and Resources, 2026). In late March 2026, fifteen projects worth A$51.9 billion were endorsed for prioritised approvals (United States Studies Centre, 2026), and Microsoft alone committed A$25 billion to Australian capacity to 2029 (Microsoft, 2026).

Figure 4. The scale of Australia's data centre commitment: forecast capacity growth to 2030 and headline investment announcements.

The core of the debate

The case for investment

The case for investment runs something like this: A middle power holding sensitive public, health and defence data has an interest in keeping that data, and the services that depend on it, within national jurisdiction rather than relying wholly on offshore capacity. Local capacity can strengthen control over data, support critical services and improve resilience during disruption (United States Studies Centre, 2026).

The economic arguments also carry weight. Hosting capacity anchors high-value digital investment onshore, generates construction and operational activity, and positions Australia as a regional hub. The instinct to build is understandable.

What the investment does and does not deliver

The framework in this paper suggests the limits of that instinct. A data centre is infrastructure at the substrate layer of general-purpose AI. It supplies the compute on which models train and run. On the account given here, compute belongs to the raw capability, not to the complements that turn capability into productive use.

Building or hosting compute is necessary, but it sits upstream of value creation. The productivity dividend comes from organisational redesign, skills, data and institutions, none of which a data centre supplies. A country can host a great deal of compute and still build very little Industrial AI. Table 3 separates what the investment delivers from what it leaves untouched

Table 3. What data centre investment delivers, and what it does not.

What the investment delivers	What it does not deliver
Sovereign hosting of sensitive public, health and defence data	Organisational redesign of workflows around the capability
Compute substrate on which models train and run	Skills and absorptive capacity to direct, check and integrate the technology
Construction and operational activity; regional hub positioning	Curated data foundations within firms and agencies
Resilience and continuity of critical services during disruption	Institutional arrangements: trust, standards, procurement norms, translation capacity
A headline investment figure	The productivity dividend itself

Analysts note that data sovereignty does not equal AI sovereignty, and that nations can localise their data while running their models on infrastructure they do not control.

The international evidence sharpens the point. In comparable economies, hyperscaler infrastructure is foreign-controlled even when physically hosted onshore, with effective control over compute, platforms and orchestration remaining external (Business Standard, 2026).

Evidence from production AI points in a similar direction. Sector analyses suggest that many industrial applications do not require frontier-scale models at all. Smaller, domain-specific language models, fine-tuned on plant and operational data and run on local or edge hardware, can match or outperform general-purpose systems on factory tasks while avoiding cloud dependence and per-token costs (Tech Monitor, 2026; IIoT World, 2026).

Australian practice supports this reading. ARM Hub, the Brisbane-based AI Adopt Centre led by Professor Cori Stewart, reports that the binding constraint for manufacturers is rarely access to models; it is data readiness, with fragmented data and legacy systems standing between firms and AI's value (Australian Manufacturing, 2025). If production AI can run on small models and local data, hyperscale compute may be less central to industrial capability than the debate assumes.

Reading the agendas

There are several considerations that require serious attention in the debate: 

• Hyperscaler value capture. The largest announced investments have been made by global cloud providers building their own capacity. As a host, Australia and the States/Territories/Local Governments, supply land, power, water and approvals, while the hosting margin and platform control generally sits with a provider.

• Hardware and frontier suppliers. A small number of chip and model suppliers capture revenue from almost every sovereign-compute push. A would-be sovereign buyer enters as a price-taker in a market tilted toward suppliers and the largest frontier buyers (International Center for Law & Economics, 2026), with the United States and China capturing most global AI investment (World Economic Forum & Bain & Company, 2026).

• Energy and water allocation. Australian data centres used about two per cent of National Electricity Market power in 2024-25, projected to roughly triple by 2030 (Climate Council, 2026). A megalitre of water has been valued near A$2.3 million in data centre use against about A$4,600 in agriculture (The Conversation, 2026), a contrast that frames real allocation choices.

• The conflation itself. The most consequential agenda may be rhetorical. Presenting compute as capability lets a hosting investment be counted as building national AI capacity, when on this analysis it builds the substrate and leaves the complements untouched. The conflation flatters the headline number and obscures the harder work.

Social licence has now entered the debate at the highest level. The Prime Minister is expected to use a speech in Sydney in mid-July 2026, titled “AI in Australia’s interests”, to set out guardrails intended to ensure AI earns its social licence, supporting growth without undercutting working conditions, fragmenting society or damaging the environment (Capital Brief, 2026).

The intervention follows union pressure for stronger worker protections and community resistance to the expansion of suburban data centres. On the analysis here, social licence is an institutional complement. Hosting capacity that communities resist, or that strains local energy and water, may convert into productive use only slowly, if at all.

The opportunity with Australian partners

These agendas are mostly global in orientation. The picture changes when attention turns to Australian participants. Australia has produced data centre operators of world standing. NEXTDC, CDC Data Centres and AirTrunk grew out of the broadband era into hyperscale platforms. Macquarie Data Centres, Goodman Group and DCI are building substantial AI-ready capacity (Data Centres Australia, 2025).

The capital is both domestic and foreign. NEXTDC raised A$1 billion in April 2026, backed by the Quebec pension investor CDPQ, against a contracted forward order book near 300 megawatts to 2029 (MinterEllison, 2026). In the same month, planning applications for new data centre campuses, together with equity raisings, signalled more than A$6 billion of intended investment in a single week (Global Data Center Hub, 2026).

Partnerships are already forming around sovereign uses. NEXTDC has signed a memorandum of understanding with OpenAI to develop a hyperscale campus at Eastern Creek in Sydney, designed as sovereign infrastructure for sensitive government workloads (Stockhead, 2026). CDC's Canberra facilities anchor defence and government demand, and the sector formed a peak body, Data Centres Australia, in late 2025.

Australian-owned operators may be better placed than offshore providers to offer sovereign access on fair terms, contracted research compute, and precinct-level anchoring, since their commercial interests align more closely with domestic demand.

The AirTrunk story carries both the promise and a caution. Founded in Sydney, it grew into the largest data centre platform in the Asia-Pacific before its sale to Blackstone for about A$24 billion in 2024 (MinterEllison, 2026). Australian capability built the asset; control and returns now sit with global capital. It follows that partnership design should keep that migration of value in view.

An Australian Data Centre capability for national benefit

The conclusion drawn here is not that Australia should avoid building data centres. It is that compute is better treated as enabling infrastructure rather than as the achievement itself, and paired deliberately with the complements that convert it into productive use. Several conditions would tilt the investment toward capability.

• Sovereign access on fair terms, so domestic firms, researchers and public agencies can obtain compute at predictable cost rather than as price-takers.

• Coupling to the discovery and knowledge-work segments, so hosted capacity feeds national research institutions and the professional economy rather than only offshore inference demand.

• Energy and water discipline, with siting and offtake arrangements that support rather than strain the clean-energy transition (Department of Industry, Science and Resources, 2026).

• Skills and absorptive capacity built alongside the concrete and steel, since compute an economy cannot use well is a stranded asset.

Treated this way, data centre investment becomes one complement among several rather than a substitute for the rest.

The risk is a story in which Australia supplies the land, power, and water for hosting, while capability and value capture settle elsewhere. The infrastructure is worth having, but it may not, by itself, be the transformation.

What realising Australia's AI potential requires

To the extent that transformation is conferred rather than intrinsic, the work is to build the complements, and it is where policy and management have real leverage. The agenda is not about acquiring the technology, which grows cheaper and more ambient by the month; it is about purposefully building everything around it.

The four complements

• Organisational redesign. Rebuilding workflows around what the capability does well, rather than bolting it onto processes designed for human-only work. This is the factory-rewiring step and is often the slowest and most valuable. It is also the one most often skipped in the rush to adopt tools.

• Skills and absorptive capacity. An economy needs the human capability to direct, check and integrate the technology, since the binding constraint is rarely the model; it is more often the surrounding competence. This is also where the probabilistic nature of the output is managed, through review and judgement rather than blind trust.

• Data and infrastructure. This is about laying the foundations and covers compute access and connectivity within reach of firms and regions that cannot fund them on their own. The data centre debate belongs here, as enabling infrastructure must be paired with the other complements rather than mistaken for the whole task.

• Institutional arrangements. Trust, standards, procurement norms, and translation capacity enable diffusion to spread across an economy rather than settle within a handful of frontier firms.

The distributional risk

Building institutional complements involves risks where diffusion might be can be  uneven. That is, the capability is general, but the complements are local, and they accrue fastest where ecosystem quality is already high. Without deliberate effort the gains may concentrate in a few places and a few firms, leaving the dividend narrow. That distributional question sits at the centre of the policy task.

The task may be summarised as converting general-purpose AI into Industrial AI across the economy, and to do it more evenly than markets alone would manage.

Where this happens: place-based innovation ecosystems

We examined the proposition that complements tend to be assembled in places, where firms, research institutions, investors and government interact repeatedly, in innovation precincts, districts and regions.

The logic is in the character of the complements. That is, compute and models travel effortlessly across borders whereas skills, trust, tacit knowledge and translation capacity tend not to. They tend to accumulate through proximity and shared institutions, which is why innovation clusters in identifiable districts rather than spreading disparately across a map (Katz & Wagner, 2014).

What the fieldwork found

Fieldwork for the research project behind this paper, across fourteen innovation ecosystems in Europe and the Nordic region during 2026, found a consistent pattern. Where anchor firms, universities, applied research institutes and government converge around shared priorities, and absorptive capacity is high, AI adoption moves beyond pilots into reconfigured workflows. Where that convergence is absent, adoption stalls.

The pattern matches the Triple Helix account of innovation, addressed in the Innovation Insight of 7 July, The Triple Helix Deficit and Australia’s Business R&D Problem, in which it was argued that sustained interaction between university, industry and government produces capability none of the three could build alone (Etzkowitz & Leydesdorff, 2000). On the evidence gathered, ecosystem quality, rather than access to the technology, may be the binding constraint on how far and how evenly Industrial AI develops.

Place-based innovation ecosystems appear to be the engine room of capability. They are where the complements are assembled, and the level at which policy either connects or fails to.

For Australia, this gives the national task a geography. Converting general-purpose AI into Industrial AI proceeds precinct by precinct and region by region, through deliberate investment in the connective tissue that links firms to research institutions and both to government (Howard, 2025). A national program without a place dimension may leave the conversion to the few locations where ecosystem quality is already high.

Implications for policy and practice

Based on the analysis presented in this paper, AI has limited productive value in itself. Value emerges as organisations and institutions build the complements around it, transforming general-purpose AI into Industrial AI.  

Whether the present period in AI infrastructure investment indicates an early slope or an expensive plateau will be settled by whether those complements are built, and built widely enough to reach break the productivity frontier. The implications differ by level of government and by sector.

For the Australian Government

The Commonwealth has begun to invest in the policy architecture for the task. The Future Made in Australia agenda and the National Reconstruction Fund already direct capital and attention toward sovereign industrial capability. The implication of this paper is one of aim rather than design: point policy development toward complements and, in particular, toward production AI, rather than toward infrastructure alone.

In practice, this means treating data centre investment as enabling infrastructure within the emerging industrial strategy, subject to the conditions set out earlier: sovereign access on fair terms, coupling of hosted capacity to research and knowledge work, energy and water discipline, and skills built alongside the concrete and steel.

It also means measuring the right things. Megawatts and headline investment figures track the substrate. Adoption beyond pilots and the breadth and evenness of diffusion track capability. A focus on capability will keep the program aligned with the outcomes sought from the headline investment numbers.

Government is also an adopter in its own right. The civic segment of Industrial AI runs on complements that the Commonwealth controls directly: procurement norms, data governance, standards and service redesign. Applying the discipline of complements to its own operations would give the government both a demonstration effect and a source of practical learning.

For the States, Territories and local government

Many of the complements are assembled in places, and much of the decision-making responsibility belongs to the states: precinct authorities, skills and TAFE systems, planning and energy approvals, and the anchor institutions around which ecosystems form. Ecosystem quality, more than access to the technology, may set the limit on how far Industrial AI spreads, which makes it a proper object of state policy.

Two practical implications follow.

• Data centre siting and approvals could be aligned with precinct strategies, so that hosted capacity feeds local research institutions and firms rather than bypassing them.

• Translation capacity, the applied institutes, testbeds and intermediaries that carry capability into small and medium firms, warrants investment on the same footing as physical infrastructure.

For local government, the resource questions are immediate. Siting, water and energy allocations are decided locally, and the wide gap between the value of these inputs in data centre use and in other uses signals genuine trade-offs.

Approval processes that ask what a facility returns to its region, in access, skills or infrastructure, would work towards hosting into participation.

For manufacturers and the physical production economy

Production AI is the segment where the returns are nearest and ministerial interest most active. The evidence assembled in this and other papers suggests the binding constraint is rarely access to models; it is data readiness. Fragmented data and legacy systems, rather than the absence of frontier compute, stand between most firms and the value.

This ordering carries a practical sequence: data foundations first, then workflow redesign, then skills, with model choice last. Smaller, domain-specific models running on a firm's own operational data may be sufficient for many factory tasks, meaning manufacturers need not wait for the data centre build-out to begin.

The J-curve carries the warning. The dividend accrues to firms that do the organisational building, not to everyone who buys the tools. The early electric factories that swapped the steam engine for a motor and kept the old layout saw little gain; those that rebuilt the production line captured the transformation. The same choice now sits in front of Australian manufacturing.

The one line takeaway is this:

Convert general-purpose AI into Industrial AI across the economy, more evenly than market forces alone would manage. The technology has arrived. The transformation, this paper argues, is a work in progress.

---

[1] Dr John H. Howard is Executive Director of the Acton Institute for Policy Research and Innovation, Sydney, and Research Director of the UTS research project Turning AI into Productivity: The Role of Innovation Ecosystems, supported by the Google Foundation. The project is led by Emeritus Professor Roy Green AM, Special Innovation Adviser at the University of Technology Sydney, who reviewed drafts of this paper and contributed advice on its argument and policy framing.

The paper draws on research undertaken for the project, including visits to 14 innovation ecosystems across eight European countries in May and June 2026. Responsibility for the views expressed, and for any errors or omissions, rests with the author.

References

Australian Government. (2024). Future Made in Australia. The Treasury. https://treasury.gov.au

Australian Government. (2025). National AI Plan. Department of Industry, Science and Resources.

Australian Manufacturing. (2025). ARM Hub contributes to AI initiatives supporting manufacturing under $73B plan. https://www.australianmanufacturing.com.au

Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. Quarterly Journal of Economics, 118(4), 1279–1333.

Bresnahan, T. F., & Trajtenberg, M. (1995). General purpose technologies: 'Engines of growth'? Journal of Econometrics, 65(1), 83–108.

Barua, D. A., Sami, S. A., & Barua, L. (2025). Leveraging artificial intelligence for smart production management in Industry 4.0. Scientific Reports, 15. https://doi.org/10.1038/s41598-025-25413-6

Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The productivity J-curve: How intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1), 333–372.

Business Standard. (2026). India's data centre sector isn't fully sovereign as AI debate deepens. https://www.business-standard.com

Capital Brief. (2026, July 13). [Newsletter item on the Prime Minister’s “AI in Australia’s interests” speech and social licence]. Capital Brief.

Climate Council. (2026). What does the data centre boom mean for Australia's switch to renewables? https://www.climatecouncil.org.au

Cockburn, I. M., Henderson, R., & Stern, S. (2019). The impact of artificial intelligence on innovation. In A. Agrawal, J. Gans, & A. Goldfarb (Eds.), The economics of artificial intelligence: An agenda (pp. 115–146). University of Chicago Press.

Data Centres Australia. (2025). Introducing Data Centres Australia. https://datacentres.org.au

David, P. A. (1990). The dynamo and the computer: An historical perspective on the modern productivity paradox. American Economic Review, 80(2), 355–361.

Department of Industry, Science and Resources. (2026). Expectations of data centres and AI infrastructure developers. Australian Government. https://www.industry.gov.au

Etzkowitz, H., & Leydesdorff, L. (2000). The dynamics of innovation: From National Systems and "Mode 2" to a Triple Helix of university–industry–government relations. Research Policy, 29(2), 109–123.

Global Data Center Hub. (2026). Australia's data centers just pulled A$6B in one week. https://www.globaldatacenterhub.com

Howard, J. H. (2025). The handbook of innovation ecosystems. Acton Institute Publishing.

Howard, J. H. (2026). Making Sense of AI in 2026: A framework for policy and practice. Acton Institute Publishing

IIoT World. (2026). Small language models: The factory AI infrastructure for the modern plant floor. https://www.iiot-world.com

International Center for Law & Economics. (2026). Build AI, don't block access: The European Union's digital-sovereignty trap. https://laweconcenter.org

Katz, B., & Wagner, J. (2014). The rise of innovation districts: A new geography of innovation in America. Brookings Institution.

Kagermann, H., Wahlster, W., & Helbig, J. (2013). Recommendations for implementing the strategic initiative INDUSTRIE 4.0. acatech.

Lee, J. (2020). Industrial AI: Applications with sustainable performance. Springer.

Microsoft. (2026). Microsoft deepens commitment to Australia with A$25 billion investment in AI infrastructure, security, and skills. https://news.microsoft.com

MinterEllison. (2026). Digital infrastructure M&A in Australia. https://www.minterellison.com

National Reconstruction Fund Corporation. (2025). Priority areas and investment mandate. https://www.nrf.gov.au

Stockhead. (2026). AI's power binge turns Australia into data centre hotspot. https://stockhead.com.au

Tech Monitor. (2026). Small language models will take centre stage in 2026. https://www.techmonitor.ai

The Conversation. (2026). Analysis of AI data centre water and energy use in Australia. https://theconversation.com

United States Studies Centre. (2026). Powering the cloud: Data centres and the future of Australia's grid. https://www.ussc.edu.au

World Economic Forum, & Bain & Company. (2026). Rethinking AI sovereignty: Pathways to competitiveness through strategic investments. https://www.weforum.org

AI declaration

This working paper was prepared with the assistance of an AI Digital Research Assistant, including web research into current Australian data centre policy and the preparation of figures and tables. The author directed the analysis, framing, terminology and conclusions, and is responsible for the final text.