Electricity, Cognition and the Next Productivity Regime

Abstract

This paper develops a thesis about the economic role of electricity
and artificial intelligence between 2026 and 2050. Its starting
proposition remains that reliable electricity is becoming the physical
substrate of industrial cognition. The production system, however, is
broader than electricity alone: capital must first be converted into
metals, generation, grids, powered sites, data centres and networks;
compute then converts electricity into inference; connectivity
distributes machine cognition to the point of use; and organisational
capability converts cognition into repeatable economic outcomes.

New industrial evidence from BHP and Telstra sharpens this production
chain. BHP’s August 2026 outlook puts hard numbers around the capital,
power, grid and copper requirements of the AI build-out, while Telstra’s
FY2026 results show the complementary importance of fibre, subsea
capacity, mobile-network latency and resilience, together with data
architecture, governance and workforce capability. These corporate
disclosures do not validate the paper’s 2050 gross-domestic-product
scenarios. They do, however, provide useful evidence that the physical
installation phase is occurring at macroeconomic scale and that
organisational conversion is moving from experimentation toward
operating practice. ^{[19] [20] [21]}

The analysis therefore still falls downstream to productivity.
Cognition is not itself productivity. Economic value appears when
cognition is reliably delivered and converted into repeatable
organisational and societal capability, capability expands functional
output, and functional output raises productivity. Historical evidence
from steam, electrification and information technology shows that major
enabling technologies tend to produce their largest productivity effects
only after complementary capital formation, organisational redesign and
diffusion.

The paper then develops a second-order argument. Artificial
intelligence may increase not only the productivity of current
production, but the productivity of invention itself. This matters
because research productivity has declined sharply across several fields
even as research effort has risen. Artificial intelligence is
structurally well matched to two causes of that decline – the growing
burden of accumulated knowledge and the increasing complexity of search
– although physical experimentation and institutional constraints remain
less tractable.

The thesis is linked to Ian Morris’s Social Development Index, which
treats energy capture, organisation, information technology and
war-making capacity as complementary dimensions of what societies are
capable of doing. The resulting 2050 global gross domestic product
scenarios are therefore framed as productivity and capability-conversion
regimes, not as mechanical electricity-to-output forecasts. The
numerical scenarios are unchanged in this edition; the BHP and Telstra
evidence instead improves the specification of the enabling system and
the constraints that must be monitored.

Executive Thesis

Capital -> Materials & Infrastructure -> Reliable
Electricity -> Compute & Inference -> Connectivity ->
Usable Cognition -> Capability -> Productivity -> Innovation
Productivity -> Economic Growth

Artificial intelligence is becoming an industrial production system.
BHP’s 2026 outlook suggests that the installation phase is already large
enough to be visible in national capital formation, electricity demand,
grid queues and copper requirements. Telstra’s 2026 disclosures show
that the production system does not end at the data centre: cognition
must be carried through resilient fibre, subsea and mobile networks and
then embedded in enterprise processes, data and governance.
^{[19] [21]}

If reliable electricity, compute and connectivity can be expanded
sufficiently, their direct unit costs are unlikely to be the dominant
determinants of the economic value created by artificial intelligence.
The central uncertainty migrates toward productivity: whether firms and
institutions can convert abundant cognition into repeatable capability.
This migration is not cleanly sequential. Physical scarcity in power,
grids, copper, powered sites and connectivity can coexist with
capability scarcity for a prolonged period, particularly through the
2030s.

Capital formation is not the same thing as productivity. A technology
can attract enormous investment and generate large task-level gains
while producing limited aggregate productivity if utilisation,
organisational redesign and diffusion disappoint. Conversely, a modest
persistent increase in economy-wide productivity growth can create an
enormous difference in real and nominal output when compounded over two
decades. The 2050 question therefore remains a capability and
productivity question enabled by an increasingly large physical
system.

1. The
Physical Production System of Industrial Cognition

Every industrial production system begins with physical inputs.
Artificial intelligence is no exception. Semiconductor fabrication, data
centres, networking, cooling and model execution require materials,
capital and electricity. The distinctive feature is what this physical
system ultimately produces: computation performs inference, and
inference produces scalable cognitive functions such as reasoning,
prediction, language, optimisation and planning.

Capital -> Materials & Infrastructure -> Reliable
Electricity -> Compute -> Inference -> Connectivity ->
Usable Cognition

The International Energy Agency projects global data-centre
electricity consumption to approximately double to around 945
terawatt-hours by 2030 in its base case, with accelerated servers
responsible for a large share of the increase. Yet even at that level
data centres remain under 3 per cent of projected global electricity
consumption. Artificial intelligence creates a rapidly growing new load,
but it does not consume the whole electricity system. ^[1]

BHP’s 18 August 2026 Economic and Commodity Outlook independently
reinforces the scale of the build-out. BHP cites AI data-centre power
demand rising 50 per cent in calendar 2025 and projects overall demand
from the segment to double by 2030, adding roughly 500 TWh over the
period. Its high case would add more than 1,100 TWh. BHP also expects
the largest US hyperscalers to spend approximately US$700-800 billion on
AI-related capital expenditure in 2026, equivalent to around 2.4 per
cent of US GDP. ^[19]

Measure	Celerity / existing paper	BHP FY2026 outlook	Interpretation
2030 data-centre electricity	~945 TWh total consumption in IEA base case	~500 TWh incremental demand to 2030	Compatible measures: one is the 2030 level, the other the increment.
High electricity case	No separate high case previously modelled	>1,100 TWh incremental demand in BHP high case	Adds a useful upper physical-demand sensitivity.
1 GW continuous load	8.76 TWh per year; ~US$0.7-0.9bn at US$75-100/MWh	Not modelled directly	Shows why electricity can be indispensable yet small relative to downstream value.
Grid connection	Treated as a feasibility constraint	>2,500 GW of renewables, storage and large-load projects awaiting connection globally	Supports a longer period of location-specific power scarcity.
Materials	Previously implicit	Every additional US$200bn of annual data-centre investment estimated to require copper equal to a new 150 ktpa mine	Adds metals and physical infrastructure explicitly before electricity and compute.

A continuously operating one-gigawatt load consumes 8.76
terawatt-hours annually. At the working whole-system electricity
expenditure range developed in the underlying research – approximately
US$75-100 per megawatt-hour – that represents roughly US$0.7-0.9 billion
of annual electricity-system expenditure. BHP’s additional 500 TWh
base-case increment is equivalent to roughly 57 GW of continuous average
load and, at the same expenditure range, about US$37.5-50 billion a year
of electricity-system expenditure. Its greater-than-1,100 TWh high case
is equivalent to more than 125 GW and more than US$82.5-110 billion a
year. These calculations are Celerity conversions of BHP’s electricity
scenarios, not BHP forecasts of electricity-system cost.
^[19]

Electricity can be a small share of economic value and still be indispensable to producing that value.

BHP also makes the material layer explicit. It estimates that every
additional US$200 billion of annual data-centre investment requires
copper equivalent to the output of a new 150,000-tonne-per-year mine,
and it continues to expect global copper demand to exceed 50 million
tonnes by 2050. More than 2,500 GW of renewable, storage and large-load
projects are waiting for grid connection globally. These figures are not
proof of an AI productivity dividend; they are evidence that the
physical installation system is already creating scarcity in metals,
generation, networks and connection capacity. ^[19]

This qualifies an earlier tendency to describe constraint migration
as if electricity scarcity would simply give way to capability scarcity.
The more plausible path is overlapping scarcity. Generation may be
adequate globally while power remains unavailable at the right location;
copper may be available in aggregate while mine development lags; data
centres may be built while grid connections are delayed. Physical and
organisational bottlenecks can therefore coexist well into the diffusion
phase.

The original optimisation insight still holds. If an interruption
prevents high-value computation from operating, the value destroyed by
lost production can dwarf the saving achieved by purchasing marginally
cheaper electricity. The relevant metric is therefore not simply the
cheapest generated megawatt-hour, but the cost and availability of
reliably delivered energy at the point where compute requires it.

Traditional question	Cognition-production question
What is the lowest cost per MWh?	What is the cost per reliably delivered MWh?
How cheap is generation?	Can power be delivered continuously at the required location?
What is average supply?	What is dependable supply when compute requires it?
How much can procurement save?	How much downstream output is at risk from interruption or delay?

As economic output per unit of electricity rises, willingness to pay
for electricity reliability should rise with it.

2.
From Compute to Productivity: Connectivity and Capability

Once adequate reliable electricity exists, additional electricity
does not mechanically create gross domestic product. It enables a
production system. The value of that system depends on whether compute
can be converted into usable cognition, whether cognition can reach the
point of application, and whether firms, workers and institutions can
convert it into repeatable outcomes.

Connectivity is therefore a distinct production layer rather than a
background utility. Telstra’s FY2026 results report strong demand from
major cloud and AI companies across its Aura Network, subsea cable and
long-haul fibre assets. Telstra had more than 8,500 kilometres of Aura
fibre in the ground and six routes ready for service at year end, while
its FY2027 plan accelerates 5G Standalone to provide greater capacity
and lower latency for AI applications. ^[21]

The distribution requirement changes the meaning of reliability.
Telstra reports around 165,000 mains-power interruptions each year
across its fixed and mobile sites. After upgrades to backup power at
more than 1,800 sites, around 97 per cent of those interruptions had no
impact on customer services in FY2026. The economically relevant concept
is therefore broader than reliable electricity: it is reliable
end-to-end delivery of cognition. ^[21]

For high-value AI systems, reliably delivered cognition can matter more than marginally cheaper cognition.

Artificial intelligence can generate a recommendation, forecast,
design or piece of software without producing an economic outcome.
Capability remains the necessary intermediate variable: the repeatable
ability of a system to achieve intended outcomes under varying
conditions.

Cognition creates potential. Capability creates repeatable
outcomes.

Capability includes governance, decision rights, process design,
reliable data, human skills, incentives, integration and learning. The
FY2026 disclosures from BHP and Telstra provide useful corporate
evidence for this complementarity. BHP states that pairing its BHP
Operating System with faster adoption of technologies such as artificial
intelligence can create a compounding effect on safety and productivity
improvement; it also emphasises reliable, resilient and secure
technology foundations and people able to use the tools effectively.
^[20]

Telstra provides a more explicit view of the organisational
machinery. More than 15,000 employees completed at least one Data &
AI Academy course during FY2026; 86 per cent of employees with a Copilot
licence used it weekly or more in June; the company added the ability
for employees to build agents; and it implemented a company-wide AI
Control Plane to monitor adoption, cost, performance, risk, safety and
compliance and to switch applications between models according to cost,
speed or reliability. Telstra also continued to simplify its data
estate, reducing the number of data platforms to 17 with a stated target
of three. ^[21]

These disclosures are evidence of capability formation, not a causal
estimate of AI’s contribution to Telstra or BHP productivity. The
distinction matters. Training, architecture, governance, process
redesign and network resilience are complementary capital. They are the
conversion machinery through which cognition may become economic
output.

Layer	FY2026 corporate evidence	Implication for the thesis
Physical installation	BHP: hyperscaler capex, electricity growth, copper intensity and grid queues	The AI installation phase is already macroeconomically material.
Distribution	Telstra: fibre, subsea, 5G Standalone, network APIs and cloud/AI customers	Connectivity belongs explicitly between compute and usable cognition.
Reliability	BHP: grid connection; Telstra: backup power, resilience and redundancy	Availability must be measured end-to-end, not only at the generator.
Capability conversion	BHP Operating System + AI; Telstra training, data simplification and AI Control Plane	Cognition requires organisational complements before it becomes repeatable productivity.

Reliable Electricity -> Compute & Inference ->
Connectivity -> Usable Cognition -> Capability -> Functional
Output -> Productivity -> Real Economic Output

Electricity determines whether industrial cognition can be produced at scale. Connectivity determines whether it can be delivered. Productivity determines what it is worth.

3. Demography
Makes Productivity More Important

The United Nations estimates that the world population was about 8.2
billion in 2024 and will continue growing for several decades at a
slowing rate, peaking around 10.3 billion in the mid-2080s. Growth is
increasingly uneven, with ageing and stagnation in many advanced and
East Asian economies and faster population growth in parts of Africa.
^[2]

As demographic growth slows, maintaining historical rates of global
output growth requires a larger contribution from productivity.
Artificial intelligence therefore arrives at an economically important
moment. In ageing economies it can offset labour scarcity; in
faster-growing emerging economies it can amplify a larger labour force
by making cognition and expertise more widely available.

The International Monetary Fund’s July 2026 update projects global
real growth of 3.0 per cent in 2026 and 3.4 per cent in 2027. These are
near-term projections, not a long-run baseline, but they show why a
sustained productivity contribution of even half or one percentage point
would be macroeconomically large. ^[3]

4.
Historical Test: Steam and the First Industrial Revolution

The history of steam is a warning against expecting an enabling
technology to create an immediate macroeconomic discontinuity. Nicholas
Crafts’ growth-accounting work finds that productivity growth during the
British Industrial Revolution was slower than older narratives implied
and that the contribution of steam was relatively small and delayed. Yet
technological change, including embodied innovation, ultimately
accounted for the acceleration in labour productivity that allowed
Britain to achieve modern economic growth. ^[4]

The lesson is not that steam was economically unimportant. It is that
invention, installation, complementary capital and economy-wide
diffusion are different stages. A technology can be revolutionary in
eventual consequence while appearing disappointing in early aggregate
statistics.

This distinction matters for artificial intelligence. Model
capability can improve much faster than organisational capital turns
over. Skills, processes, governance and institutions can therefore
become the slow-moving variables even when the underlying technology
advances rapidly.

5. Historical Test:
Electrification

Electrification is the closest physical analogue because electricity
was both infrastructure and a general-purpose production input. It
required generation, networks, motors and large amounts of capital, yet
its largest effects depended on how firms reorganised production around
its properties.

Fiszbein, Lafortune, Lewis and Tessada use U.S. manufacturing data
from 1890 to 1940 and find that electricity produced rapid and
long-lasting labour-productivity gains. Electrification was accompanied
by capital deepening and organisational changes, and the effects varied
with industrial structure. ^[5]

The critical lesson is that replacing a steam-driven system with an
electric motor was not the endpoint. Distributed electric power allowed
machinery to be placed differently, reduced dependence on line shafts,
changed factory layouts, altered material handling and enabled new
operating methods.

Electrification plus factory redesign then is structurally similar to abundant cognition plus enterprise redesign now.

The historical lag should not be interpreted as zero early benefit
followed by sudden productivity. Direct gains could appear relatively
quickly; the larger transformation came as complementary capital and
organisational change widened. This suggests a similar sequence for
artificial intelligence: task augmentation first, organisational
redesign second, economy-wide productivity later.

6. Historical Test:
Information Technology

Information technology provides the most relevant modern calibration
of aggregate productivity magnitude. Jovanovic and Rousseau identify
electricity and information technology as two major general-purpose
technologies and emphasise broad adoption, continuing technological
improvement and the spawning of complementary innovation.
^[6]

The information-technology era reinforces the organisational
argument. The technology became more economically important as firms
accumulated complementary intangible capital—software, processes,
skills, databases, business models and organisational redesign.

Brynjolfsson, Rock and Syverson formalise this as the Productivity
J-Curve. New general-purpose technologies require substantial
complementary investment in processes, products, business models and
human capital. Because much of this investment is intangible and poorly
measured, productivity can appear weak early in the diffusion period and
stronger later when the benefits are harvested. ^[7]

The early productivity paradox can be the accounting shadow of
capability formation.

The contemporary capital build-out is now large enough to resemble
the installation phase described by the historical literature. BHP
estimates 2026 AI-related capital expenditure by the largest US
hyperscalers at US$700-800 billion, around 2.4 per cent of US GDP. This
is evidence of extraordinary capital formation, not of realised
productivity. It therefore strengthens rather than removes the J-curve
distinction between installation and harvesting. ^[19]

7.
Current Evidence: Large Task Effects, Limited Aggregate Penetration

The strongest current evidence is microeconomic. Noy and Zhang’s
controlled experiment with 453 college-educated professionals found that
access to generative artificial intelligence reduced time spent on
professional writing tasks by about 40 per cent while increasing output
quality by about 18 per cent. ^[8]

Brynjolfsson, Li and Raymond studied 5,172 customer-support agents
and found that artificial-intelligence assistance increased
productivity, measured by issues resolved per hour, by about 15 per cent
on average, with larger gains among less experienced and lower-skilled
workers. Their results also suggest that artificial intelligence can
transmit elements of the practices of higher-performing workers to
others. ^[9]

These effects are far larger than any plausible near-term aggregate
productivity uplift. That is expected. Task-level gains must pass
through several filters: task exposure, adoption, verification, the
share of a job affected, organisational integration, and the share of
the economy in which the technology is useful.

The International Labour Organization’s 2025 refined global index
estimates that one in four workers worldwide is in an occupation with
some degree of exposure to generative artificial intelligence, while
emphasising that transformation rather than complete replacement is the
more likely outcome for most jobs. ^[10]

The 2026 BHP and Telstra disclosures add a different kind of
evidence. They show large-scale investment in physical AI infrastructure
and observable organisational investment in adoption, governance and
process conversion. They do not yet establish an economy-wide
productivity effect, but they indicate that the two preconditions
identified by the historical literature – complementary physical capital
and complementary organisational capital – are being built
simultaneously. ^{[19] [20] [21]}

8. From
Task Productivity to Economy-Wide Productivity

The correct aggregation is not to apply a 15 or 40 per cent task
improvement to the entire economy. A useful conceptual equation is:
sector productivity gain equals cognitive exposure multiplied by task
productivity gain, adoption and capability conversion. Aggregate
productivity is then the output-weighted sum of sector gains.

Organisation for Economic Co-operation and Development estimates
suggest that artificial intelligence could contribute roughly 0.5 to 1.0
percentage point to annual labour-productivity growth across Group of
Seven economies over the next decade under a central scenario. Slower
adoption produces gains around 0.2 to 0.4 percentage points, while rapid
adoption and broader capabilities can lift the estimate as high as about
1.3 percentage points. ^[11]

A separate 2026 micro-to-macro analysis presented through the
American Economic Association estimates an artificial-intelligence
contribution of roughly 0.3 to 0.9 percentage points to annual
total-factor-productivity growth over the next decade, with much larger
effects in knowledge-intensive services than in manual-intensive
sectors. ^[12]

A mature artificial-intelligence productivity contribution around 0.5
to 1.0 percentage point is therefore a serious evidence-based range, not
a number chosen for convenience.

9. A Bottom-Up Cross-Check

The illustrative sector model below is a consistency test rather than
a forecast. It asks whether an aggregate result around one percentage
point can emerge from moderate sector assumptions rather than heroic
economy-wide gains.

Economic block	Model share	Mature uplift	Aggregate contribution
Agriculture, mining and primary production	6%	+0.3%	+0.02ppt
Manufacturing	16%	+1.2%	+0.19ppt
Construction, utilities and infrastructure	10%	+0.6%	+0.06ppt
Retail, wholesale, logistics and hospitality	16%	+0.8%	+0.13ppt
ICT, finance and professional/business services	22%	+2.0%	+0.44ppt
Healthcare, education and public administration	17%	+0.8%	+0.14ppt
Real estate, personal and other services	13%	+0.4%	+0.05ppt
Total	100%		~+1.03ppt

The model is intentionally conservative relative to the largest
task-level experiments. Its implication is that knowledge-intensive
sectors can do most of the work. Artificial intelligence does not need
to transform every sector equally to produce a macroeconomically
significant result.

10.
Diffusion Paths Matter More Than a Flat Productivity Shock

History argues against imposing a constant one-percentage-point
uplift immediately. A more plausible structure is a diffusion curve in
which direct task effects appear first, complementary investment
follows, and deeper organisational gains emerge later.

Path	2026-30	2031-35	2036-40	2041-50	Interpretation
Slow capability	+0.1ppt	+0.4ppt	+0.6ppt	+0.5ppt	AI remains useful but organisational conversion disappoints.
Electrification-like	+0.2ppt	+0.6ppt	+0.9ppt	+1.0ppt	Firms gradually redesign around abundant cognition.
Fast artificial intelligence	+0.4ppt	+1.0ppt	+1.2ppt	+1.2ppt	Software distribution compresses the historical diffusion lag.
Innovation acceleration	+0.4ppt	+1.2ppt	+1.6ppt	+2.0ppt	AI raises production productivity and the productivity of invention.

These paths are hypotheses, not forecasts. Their purpose is to make
timing explicit. A mature productivity gain reached only in 2040 creates
much less 2050 output than the same mature gain reached in 2030. The
shape of organisational diffusion therefore matters almost as much as
the eventual level of productivity improvement.

The 2026 industrial evidence does not justify changing these
diffusion paths. It does, however, alter what should be watched. BHP
suggests the physical installation phase may be faster and more
capital-intensive than a conservative diffusion narrative would imply,
while its grid and copper evidence also shows why physical bottlenecks
may delay usable capacity. Telstra provides early evidence that a large
incumbent can move from tool access toward governed enterprise-scale
adoption. Together they raise confidence in the mechanism without
resolving the timing of the aggregate productivity response.
^{[19] [21]}

11. Global Economic Output to
2050

The gross-domestic-product scenarios are consequences of productivity
assumptions rather than independent forecasts. Using the modelling
baseline developed in the underlying research—a
non-artificial-intelligence structural real-growth assumption of
approximately 2.75 per cent and long-run inflation around 2.25 per
cent—the no-artificial-intelligence path reaches roughly US$410 trillion
of nominal global gross domestic product by 2050.

Regime	Core mechanism	Indicative 2050 nominal GDP
No artificial-intelligence uplift	Baseline productivity	~US$410tn
Slow capability	Limited diffusion and conversion	~US$455-465tn
Electrification-like	Mature uplift approaches +1ppt	~US$490-500tn
Fast artificial intelligence	Earlier and broader capability diffusion	~US$520-530tn
Innovation acceleration	Production plus research-productivity gains	~US$575-600tn+

The precise dollar values are highly sensitive to inflation, exchange
rates, demographic outcomes and the assumed baseline. The economically
important result is the compounding. A sustained one-percentage-point
difference in productivity growth over 24 years makes real output
roughly 27 per cent larger than the counterfactual. At 1.5 percentage
points the difference is roughly 43 per cent; at two percentage points
it is roughly 61 per cent.

The 2050 economic question is a growth-rate question, not a one-time
artificial-intelligence revenue question.

The BHP and Telstra evidence is deliberately not used to raise these
numbers. BHP provides stronger evidence that the physical preconditions
are being financed and built; Telstra provides stronger evidence that
connectivity and organisational conversion are becoming operational
priorities. Neither provides a defensible estimate of the long-run
global productivity uplift. The mature 0.5-1.0 percentage-point central
range therefore remains the more important uncertainty than electricity
consumption itself. ^{[19] [21]}

12. The
Second-Order Channel: Productivity of Invention

The largest long-run effect of artificial intelligence may not be
automation of current tasks. General-purpose technologies can spawn
complementary invention. Information technology was especially important
because it improved the ability to manipulate information and supported
innovation in other technological domains.

Bontadini, Corrado, Haskel and Jona-Lasinio explicitly analyse
artificial intelligence as both a general-purpose technology and an
innovation in the method of innovation. Their framework separates
upstream innovation from downstream production and argues that
artificial intelligence can raise productivity in both. They also find
that software products and software research and development accounted
for a substantial share of recent U.S. nonfarm business
labour-productivity growth and its acceleration. ^[13]

This creates two distinct growth channels. The first is ordinary
production productivity: existing goods and services are produced more
efficiently or with greater functionality. The second is innovation
productivity: scientists, engineers and firms become more productive at
creating the next generation of technologies.

Artificial intelligence can improve today’s production and the
process that creates tomorrow’s productivity.

If the second channel becomes material, artificial intelligence can
affect not only the level of output but the rate at which the production
frontier itself advances. This is the economic mechanism behind the
upper 2050 scenarios.

13.
Declining Research Productivity as the Opportunity

Bloom, Jones, Van Reenen and Webb document a long-run decline in
research productivity across semiconductors, agriculture, medical
research and the aggregate economy. Research effort has risen
substantially while research productivity has fallen sharply. Their
Moore’s Law example is particularly striking: sustaining the historic
pace of semiconductor improvement required more than 18 times as many
researchers as in the early 1970s. ^[14]

This evidence changes the upper artificial-intelligence scenario.
Artificial intelligence does not need to create an unprecedented
scientific miracle to matter. It may create enormous value by weakening
an existing headwind: the increasing quantity of research effort
required to sustain technological progress.

However, a one-time increase in researcher productivity is not
sufficient if the underlying decline then resumes. For artificial
intelligence to alter long-run growth materially, it must change the
trend or the research production function itself.

14. Why Research
Productivity Has Declined

Benjamin Jones’s burden-of-knowledge hypothesis provides one
mechanism. As knowledge accumulates, innovators face an increasing
educational burden. They respond through greater specialisation and
teamwork, but this raises coordination requirements and can reduce
individual breadth. ^[15]

Another mechanism is search complexity. As the frontier expands and
simpler opportunities are exhausted, the space of possible hypotheses,
combinations and experiments becomes harder to navigate. Artificial
intelligence is structurally well matched to both problems because
search, synthesis, retrieval, simulation, optimisation and cross-domain
pattern recognition are cognitive tasks.

Source of declining research productivity	AI attackability	Mechanism
Knowledge burden	Very high	Search, synthesis, memory and cross-domain cognition.
Combinatorial and search complexity	Very high	Simulation, optimisation, candidate generation and active learning.
Physical experimentation	Moderate and rising	AI plus robotics and automated laboratories.
Institutional and regulatory friction	Low to moderate	Decision support can help, but institutions must still adapt.
Irreducible physical limits	Low	Scientific and physical constraints remain real.

The first two mechanisms are fundamentally cognitive. This makes
artificial intelligence structurally matched to a meaningful portion of
the research-productivity problem. Physical experimentation and
institutional constraints, however, place a natural limit on purely
computational acceleration.

15. The Stronger Innovation
Hypothesis

The strongest long-run hypothesis is not that artificial intelligence
makes researchers a fixed percentage faster. It is that artificial
intelligence weakens the historical relationship between accumulated
knowledge and the human cognitive burden required to push the frontier
forward.

If the stock of scientific literature doubles but machine cognition
allows researchers to search, synthesise and reason across that larger
stock without doubling their own cognitive burden, the research
production function changes. The old trend of falling research
productivity need not continue unchanged.

This is a structural-break hypothesis. It is more demanding than
demonstrating that a writing task is completed faster, but it is also
more economically consequential. It would mean that artificial
intelligence acts partly as externalised cognitive infrastructure for
the research system.

Automation of experimentation can extend the mechanism further. The
relevant system becomes human researchers plus machine cognition plus
simulation plus robotics plus automated measurement. In such a system
the effective research input is no longer limited to the number of human
researchers or the hours they can personally devote to search and
experimentation.

16.
Constraint Migration: Scarcity Overlaps Rather Than Disappears

The history of information technologies can be interpreted as a
sequence of constraints. Writing reduced memory scarcity. Printing
reduced copying scarcity. Telegraphy and telecommunications reduced
transmission scarcity. Computing reduced calculation scarcity. The
internet reduced access and distribution scarcity. Artificial
intelligence may reduce the scarcity of cognition over abundant
information.

The new industrial evidence suggests that this sequence should not be
interpreted as one bottleneck disappearing before the next begins. BHP’s
grid queue, copper intensity and high electricity-demand case point to
durable physical constraints. Telstra’s fibre, subsea, mobile and
resilience investment points to distribution constraints. At the same
time, its training, data simplification and governance investment points
to capability constraints inside the enterprise. ^[19]
[21]

Constraint migration is therefore overlapping. Economic rents can
remain in copper, transmission, powered land, data-centre capacity,
fibre routes or network resilience even while cognition itself becomes
cheaper. Later, as physical capacity expands, rents may migrate toward
scarce organisational capability, proprietary workflows, trusted data,
integration and the ability to convert machine cognition into decisions
and action.

Modern economies already produce more scientific, technical and
commercial information than any individual can process. Information
abundance has created its own bottleneck: the ability to search,
synthesise, interpret and act on the accumulated stock. Artificial
intelligence potentially converts a much larger share of stored
information into usable cognition, but that cognition still requires
reliable physical delivery and institutional conversion.

Electricity sits underneath this process as the physical energy
input; connectivity links cognition to users and machines; capability
converts that cognition into repeatable outcomes. This is why the
energy, infrastructure and productivity arguments are not separate. They
are different layers of one production system.

The dominant constraint can migrate without the previous constraint becoming economically irrelevant.

17. Ian Morris
and the Social Development Evidence

Ian Morris’s Social Development Index is valuable because it does not
reduce development to income. Morris defines social development in terms
of societies’ capacity to get things done and decomposes it into energy
capture per capita, organisation, information technology and war-making
capacity. ^[16]

Morris dimension	Capability interpretation
Energy capture	Physical capacity to do work
Organisation	Institutional and coordination capability
Information technology	Information-processing and cognitive capability
War-making capacity	Large-scale technology, logistics and mobilisation capability

The relevance to the present thesis is not that Morris predicts
artificial-intelligence growth. It is that his long historical framework
treats energy, information and organisation as complementary
determinants of what societies can do. His account of the
late-eighteenth-century Western take-off gives particular weight to
northwest Europe’s ability to exploit fossil energy at scale.
^[16]

The artificial-intelligence era may be unusual because it potentially
advances several dimensions simultaneously. Electricity expands usable
energy. Artificial intelligence expands information processing and
cognition. Digital networks and software can expand coordination.
Robotics and automated systems can expand the ability to mobilise
technology in the physical world.

Energy abundance becomes economically transformative when combined
with information-processing and organisational capability.

This interpretation also explains why gross domestic product alone
may initially understate the transformation. Morris is interested in
functional social capability rather than market transactions alone. The
present thesis similarly treats capability and functional output as
mechanisms that precede or underlie measured productivity and gross
domestic product.

18.
Advanced and Emerging Economies May Experience Different Mechanisms

18.1 Advanced economies

In ageing advanced economies artificial intelligence can offset
labour scarcity, augment expensive cognitive labour and improve the
productivity of existing capital. The main constraints are likely to be
a combination of location-specific physical infrastructure and
organisational conversion: grid connections, powered data-centre
capacity and resilient networks can bind at the same time as established
processes, regulation, legacy technology and institutional risk
tolerance slow diffusion even when models are available.

18.2 Emerging economies

In emerging economies the mechanism can be different. Reliable
electricity, digital connectivity and machine cognition can combine with
a growing labour force and capital deepening. Artificial intelligence
can provide access to forms of cognitive assistance that historically
required long periods of human-capital accumulation.

This does not eliminate the importance of education or institutions.
It changes the production function. A worker, entrepreneur, clinician or
engineer with access to high-quality machine cognition may be able to
perform tasks that previously required a much deeper local stock of
specialised expertise.

The largest global gross-domestic-product surprise may therefore come
not only from replacing expensive cognitive tasks in advanced economies
but from capability leapfrogging in economies where energy and
specialised cognition have historically both been scarce.

19. Overall Thesis

Capital -> Materials & Infrastructure -> Reliable
Electricity -> Compute & Inference -> Connectivity ->
Industrial Cognition -> Organisational Capability -> Productivity
-> Innovation Productivity -> Economic Growth

Industrial cognition is physically produced. Electricity remains
fundamental, but the production chain also requires materials, grids,
powered sites, data centres and networks. The direct production cost of
electricity is unlikely to be the principal long-run economic constraint
if reliable supply can be expanded at costs small relative to downstream
economic value. The important qualification is that physical bottlenecks
can remain locally scarce and rent-bearing for much longer than
aggregate electricity-supply statistics imply.

The principal uncertainty still migrates toward capability
conversion: whether organisations and societies can redesign themselves
around abundant cognition. BHP’s emphasis on combining its operating
system with AI and Telstra’s investment in skills, data architecture,
governance and a company-wide Control Plane are contemporary examples of
the complementary capital described by the historical
general-purpose-technology literature. ^[20]
[21]

The largest upside then comes from a second migration. If artificial
intelligence reduces the cognitive burden of operating at the frontier
of knowledge, it may slow or reverse declining research productivity. In
that case artificial intelligence affects not merely the level of output
but the rate at which future productive capability is created.

The thesis therefore connects materials and energy economics, digital
infrastructure, artificial-intelligence economics, productivity theory
and the history of social development through one production chain.
Materials and electricity make industrial cognition physically possible.
Connectivity makes it distributable. Cognition is the scalable new
input. Capability is the conversion mechanism. Productivity is the
observable economic result. Innovation productivity determines whether
the result compounds into a higher long-run growth regime.

20. Research Propositions

Proposition 1. Artificial intelligence is an industrial production
system requiring capital, materials, infrastructure, electricity,
compute and connectivity before cognition can be delivered at scale.

Proposition 2. Reliable electricity is becoming the physical energy
substrate of industrial cognition.

Proposition 3. Materials, grid connection, powered sites and digital
networks can remain binding constraints even when aggregate electricity
generation is sufficient.

Proposition 4. Electricity can be fundamental to
artificial-intelligence production without being the principal
determinant of economic value because its direct cost can be small
relative to downstream productive value.

Proposition 5. For high-value cognition infrastructure, end-to-end
reliability and availability can be economically more important than
marginal electricity, compute or network price.

Proposition 6. Physical-infrastructure scarcity and
organisational-capability scarcity can coexist; constraint migration is
overlapping rather than strictly sequential.

Proposition 7. Connectivity is a distinct production layer between
compute and usable cognition because cognition must be delivered
reliably to workers, enterprises, machines and agents.

Proposition 8. Cognition is not productivity; capability is the
conversion mechanism between cognition and repeatable economic
output.

Proposition 9. Contemporary corporate evidence from BHP and Telstra
is consistent with the view that AI productivity requires complementary
operating systems, data architecture, governance, skills, resilience and
process redesign. ^{[20] [21]}

Proposition 10. Historical general-purpose technologies generated
their largest gains after complementary capital investment and
organisational redesign.

Proposition 11. A mature global artificial-intelligence productivity
contribution of roughly 0.5 to 1.0 percentage point is compatible with
both historical experience and current bottom-up evidence.

Proposition 12. Artificial-intelligence diffusion should be modelled
as a J-curve or S-curve rather than a flat productivity shock.

Proposition 13. A sustained one-percentage-point productivity regime
can make the 2050 real economy roughly 27 per cent larger than the
no-artificial-intelligence counterfactual if sustained for 24 years.

Proposition 14. Artificial intelligence can affect growth through
both ordinary production productivity and the productivity of
invention.

Proposition 15. The strongest long-run artificial-intelligence
mechanism may be its ability to weaken the historical decline in
research productivity caused partly by knowledge burden and search
complexity.

Proposition 16. Artificial intelligence combined with automated
experimentation can extend cognition into a closed-loop research
capability, but physical and institutional constraints remain.

Proposition 17. Ian Morris’s long-run evidence supports a
complementary capability model of development in which energy,
information processing, organisation and technology mobilisation
reinforce one another.

Proposition 18. The 2050 gross-domestic-product outcome is best
understood as a productivity and capability-conversion problem enabled
by a large physical and digital infrastructure system, rather than an
electricity-to-gross-domestic-product conversion problem.

21. What Would Falsify
or Weaken the Thesis?

A serious thesis must identify evidence that would weaken it. The
framework would require material revision if large-scale investment in
power, compute and connectivity produces abundant capacity but
utilisation remains persistently weak, or if broad
artificial-intelligence adoption fails to produce measurable enterprise
productivity after sufficient time for organisational adjustment.

It would also weaken if current task-level gains prove largely
non-transferable to firm-level output; if verification, coordination,
security and network costs consume most apparent time savings; if the
physical system proves materially more capital-intensive than the
downstream value it enables; if artificial intelligence fails to improve
research search, synthesis and experimentation sufficiently to alter
declining research-productivity trends; or if regulatory and
institutional bottlenecks dominate cognitive gains for prolonged
periods.

Conversely, the thesis would strengthen if data-centre and network
utilisation rises alongside falling inference cost; grid queues and
connection times improve; AI-intensive organisations show persistent
output advantages after controlling for capital intensity; productivity
gains broaden from tasks to firms and sectors; research-cycle times
fall; idea output per unit of research effort stabilises or rises; and
the productivity uplift diffuses beyond information technology and
professional services into physical industries.

22. Investment and
Monitoring Implications

The thesis implies that investment monitoring should distinguish
physical enabling conditions, cognition production and distribution,
organisational conversion, and innovation productivity. BHP and Telstra
add several observable measures to the monitoring system: grid queues,
copper requirements, data-centre capital formation, fibre and subsea
contracts, network latency and resilience, enterprise AI adoption,
governance and model-cost control. ^{[19] [21]}

Layer	What to monitor	Why it matters
Physical inputs & installation	AI/data-centre capex; copper mine approvals and supply gap; generation additions; powered-land availability; grid queue and connection times; transformer and transmission lead times	Tests whether the physical production system can scale and identifies where scarcity rents sit.
Cognition production & distribution	Accelerator utilisation; inference cost; data-centre load; fibre/subsea capacity; 5G/edge capacity; latency; network availability and outage resilience	Tests whether compute can be produced and reliably delivered to the point of use.
Capability conversion	Enterprise adoption; workflow depth; data-platform simplification; governance/control-plane coverage; agent deployment; functional output; unit cost; margins and operating leverage	Tests whether cognition is becoming repeatable organisational performance rather than isolated tool use.
Innovation productivity	Research-cycle times; cost per validated experiment; high-quality outputs per unit of research input; software R&D productivity; drug/material discovery throughput; autonomous experimentation	Tests whether AI is changing the production function for future ideas, which drives the upper 2050 scenarios.

The long-run investment question is therefore not simply which
company owns the most electricity, accelerators or model capacity. It is
which owners of scarce physical infrastructure can earn durable rents
during the build-out, and which companies and economies can subsequently
convert reliable energy, connectivity and abundant cognition into
repeatable productivity and reinvest the surplus into further
capability.

The innovation-acceleration scenario should receive greater
probability only if research-productivity indicators begin to improve
persistently. Likewise, the faster diffusion scenarios should gain
probability only when firm-level productivity evidence broadens beyond
pilots and task studies into durable sector-level outcomes.

Conclusion

The economic significance of the artificial-intelligence era may be
misidentified if analysis begins and ends with models, chips or
data-centre revenue. The deeper production system begins earlier – with
capital, materials and infrastructure – and ends later, with productive
capability.

BHP’s August 2026 outlook provides unusually clear evidence for the
upstream system: historically large AI capital expenditure, rapidly
rising data-centre electricity demand, grid-connection queues and copper
requirements. Telstra’s FY2026 disclosures provide evidence for the next
layers: fibre, subsea and mobile connectivity, network resilience,
workforce skills, data simplification, governance and enterprise-scale
AI control. Together they refine the chain without changing its economic
endpoint. ^{[19] [20] [21]}

Electricity makes industrial cognition physically possible. Compute
makes inference scalable. Connectivity makes cognition distributable.
Capability determines whether cognition becomes productive. Productivity
determines the size of the economy. And if cognition also raises the
productivity of invention, artificial intelligence can alter the process
that creates future productivity itself.

The central question to 2050 is therefore not simply how much
electricity artificial intelligence consumes. It is how effectively
capital, materials, electricity, compute and networks can be assembled
into reliably delivered cognition, and how effectively organisations and
societies can convert that cognition into capability and a higher rate
of productive progress.

History does not guarantee that conversion. Steam, electrification
and information technology all required complementary capital,
organisational redesign and time. The 2026 BHP and Telstra evidence
indicates that both physical installation and organisational capability
formation are now underway at significant scale, but it is still too
early to infer the long-run aggregate productivity result. The 2050 GDP
scenarios therefore remain unchanged: they continue to be conditional
tests of alternative productivity regimes rather than forecasts derived
from electricity consumption or AI capital expenditure.

If reliable physical and digital infrastructure expands, cognition
becomes abundant and capability conversion succeeds, a sustained global
productivity uplift approaching the historical scale of prior
general-purpose technologies is plausible. If artificial intelligence
also weakens the long-run decline in research productivity, the upper
growth scenarios become economically credible rather than merely
speculative.

References

^[1] International Energy Agency. Energy and AI: Energy
demand from AI. Paris: IEA, 2025. Data-centre electricity consumption
base case reaches approximately 945 TWh by 2030.
https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai

^[2] United Nations, Department of Economic and Social
Affairs, Population Division. World Population Prospects 2024: Summary
of Results. New York: United Nations, 2024.

^[3] International Monetary Fund. World Economic Outlook
Update, July 2026. Washington, DC: IMF, 8 July 2026. Global real growth
projected at 3.0% in 2026 and 3.4% in 2027.
https://www.imf.org/en/videos/view/6400479002112

^[4] Crafts, Nicholas. ‘Productivity Growth in the
Industrial Revolution: A New Growth Accounting Perspective.’ Journal of
Economic History 64, no. 2 (2004): 521–535.
https://doi.org/10.1017/S0022050704002746

^[5] Fiszbein, Martin, Jeanne Lafortune, Ethan G. Lewis,
and José Tessada. ‘Powering Up Productivity: The Effects of
Electrification on U.S. Manufacturing.’ NBER Working Paper 28076,
revised April 2024. https://doi.org/10.3386/w28076

^[6] Jovanovic, Boyan, and Peter L. Rousseau. ‘General
Purpose Technologies.’ NBER Working Paper 11093 (2005); subsequently in
Handbook of Economic Growth, vol. 1B. https://doi.org/10.3386/w11093

^[7] Brynjolfsson, Erik, Daniel Rock, and Chad Syverson.
‘The Productivity J-Curve: How Intangibles Complement General Purpose
Technologies.’ American Economic Journal: Macroeconomics 13, no. 1
(2021): 333–372. Earlier NBER Working Paper 25148.
https://doi.org/10.3386/w25148

^[8] Noy, Shakked, and Whitney Zhang. ‘Experimental
Evidence on the Productivity Effects of Generative Artificial
Intelligence.’ Science 381, no. 6654 (2023): 187–192.
https://doi.org/10.1126/science.adh2586

^[9] Brynjolfsson, Erik, Danielle Li, and Lindsey R.
Raymond. ‘Generative AI at Work.’ Field study of 5,172 customer-support
agents; working-paper version 2023 and subsequent publication.
https://arxiv.org/abs/2304.11771

^[10] Gmyrek, Paweł, Janine Berg, Karol Kamiński, Filip
Konopczyński, Agnieszka Ładna, Balint Nafradi, Konrad Rosłaniec, and
Marek Troszyński. Generative AI and Jobs: A Refined Global Index of
Occupational Exposure. ILO Working Paper 140. Geneva: International
Labour Organization, 2025.
https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure

^[11] Organisation for Economic Co-operation and
Development. Foundations for Growth and Competitiveness 2026, overview
and productivity analysis. Central estimates place
artificial-intelligence contributions to annual labour-productivity
growth across Group of Seven economies around 0.5–1.0 percentage point
over ten years, depending on adoption and capabilities.
https://www.oecd.org/en/publications/2026/04/foundations-for-growth-and-competitiveness-2026_f68a156b/full-report/overview_97442815.html

^[12] Filippucci, Francesco, Peter Gal, Katharina Laengle,
and Matthias Schief. Micro-to-macro estimates of artificial-intelligence
productivity effects, presented in the 2026 American Economic
Association programme. Estimated contribution: approximately 0.3–0.9
percentage points to annual total-factor-productivity growth over the
next decade. https://www.aeaweb.org/conference/2026/program/1160

^[13] Bontadini, Filippo, Carol Corrado, Jonathan Haskel,
and Cecilia Jona-Lasinio. ‘AI as an Innovation in the Method of
Innovation: Implications for Productivity Growth.’ AEA Papers and
Proceedings 116 (2026): 36–40.
https://doi.org/10.1257/pandp.20261036

^[14] Bloom, Nicholas, Charles I. Jones, John Van Reenen,
and Michael Webb. ‘Are Ideas Getting Harder to Find?’ American Economic
Review 110, no. 4 (2020): 1104–1144.
https://doi.org/10.1257/aer.20180338

^[15] Jones, Benjamin F. ‘The Burden of Knowledge and the
Death of the Renaissance Man: Is Innovation Getting Harder?’ Review of
Economic Studies 76, no. 1 (2009): 283–317; NBER Working Paper 11360.
https://doi.org/10.3386/w11360

^[16] Morris, Ian. The Measure of Civilization: How Social
Development Decides the Fate of Nations. Princeton University Press,
2013; and Morris, Ian. Why the West Rules—For Now: The Patterns of
History, and What They Reveal about the Future. Farrar, Straus and
Giroux, 2010. Stanford summary of the index:
https://classics.stanford.edu/publications/measure-civilization-how-social-development-decides-fate-nations

^[17] Crafts, Nicholas. ‘Artificial Intelligence as a
General-Purpose Technology: An Historical Perspective.’ Oxford Review of
Economic Policy 37, no. 3 (2021): 521–536. The paper reviews steam,
electricity and information and communication technology, emphasising
diffusion lags and the possibility that artificial intelligence raises
research-and-development productivity.

^[18] Organisation for Economic Co-operation and
Development. OECD Compendium of Productivity Indicators 2026. Paris:
OECD, 2026. Notes tentative recent productivity signals consistent with
artificial-intelligence adoption while emphasising diffusion and
complementary investment.

^[19] BHP. ‘Economic and Commodity Outlook.’ 18 August
2026. BHP Investor Hub.
https://www.bhp.com/investor-hub/reports-and-presentations/economic-and-commodity-outlook/2026/08/economic-and-commodity-outlook

^[20] BHP. Annual Report 2026. 18 August 2026. See sections
on leveraging the BHP Operating System and technology to drive
productivity.
https://www.bhp.com/investor-hub/reports-and-presentations/annual-report

^[21] Telstra Group Limited. FY26 Results – CEO and CFO
Analyst Briefing Presentation and Materials. 13 August 2026. Includes
CEO/CFO speeches and Full-Year Results and Operations Review.
https://www.telstra.com.au/content/dam/tcom/about-us/investors/pdf-i/financial-results-fy26-presentation-materials.pdf

Source and Assumption Notes

The electricity whole-system expenditure range, sector weights,
diffusion paths and 2050 nominal gross-domestic-product scenarios are
Celerity Research modelling assumptions developed in the underlying
research. They are not forecasts published by the cited institutions.
The external references support the historical mechanisms, current
electricity and artificial-intelligence evidence, demographic context,
productivity ranges, research-productivity arguments and the industrial
evidence discussed in this edition; they do not validate every numerical
scenario in the paper.

The BHP scenarios are BHP forward-looking estimates and include
third-party information that BHP states has not been independently
verified. The conversions of BHP’s 500 TWh and greater-than-1,100 TWh
electricity increments into average gigawatts and Celerity’s
US$75-100/MWh whole-system expenditure range are Celerity calculations
for analytical comparison. They are not BHP forecasts of power cost or
gross-domestic-product impact.

The Telstra evidence is used as a corporate case of infrastructure
demand and capability formation. Telstra’s disclosures demonstrate
investment, adoption, governance and operating priorities; they do not
isolate a causal AI contribution to group productivity. Accordingly,
neither BHP nor Telstra evidence is used to change the 2050
gross-domestic-product scenario values in this edition.

The modelling framework should continue to be maintained as a
controlled assumptions register with explicit source provenance,
sensitivity analysis and probability weights. The scenarios are intended
to test economic consequences of alternative productivity and diffusion
regimes, not to present a single point forecast.

Publication Note

This Publication Edition develops a Celerity Research thesis rather
than a point forecast. Historical and empirical claims are referenced to
the evidence base listed in the paper. The electricity whole-system
expenditure range, sector weights, diffusion paths and 2050 nominal
global gross-domestic-product outcomes are Celerity Research modelling
assumptions and scenarios. They are presented to test the economic
consequences of alternative productivity regimes and should not be read
as forecasts issued by the cited institutions.

General Research Disclaimer

This publication is general research and commentary only. It does not
constitute personal financial advice, investment advice, an offer,
solicitation or recommendation to buy or sell any security or financial
product. Celerity Research uses historical evidence, external research
and scenario analysis to develop and test investment and economic
hypotheses. Readers should make their own assessment of the assumptions
and obtain professional advice appropriate to their circumstances.

General information

Celerity publishes general research and commentary only. Nothing in this publication constitutes financial advice, investment advice, personal advice, an offer, solicitation or recommendation to buy or sell any financial product or security.