THE GREAT DATA CENTRE DEBATE - EV Curve Futurist

6 questions about AI, energy and our future

This essay is told largely from an Australian perspective, but the questions are global. Countries around the world are grappling with the same dilemma: how do we capture the enormous potential of AI while managing the energy, water, infrastructure, environmental and social pressures that come with it? The policies will differ, but many of the lessons won’t.

AI is starting to collide with the physical world.

Behind ChatGPT, robotics, autonomous vehicles and increasingly powerful AI models sit enormous data centres packed with GPUs, cooling systems, batteries, transformers and networking equipment. They need electricity. They generate heat. Some use a lot of water. They need land, transmission and billions of dollars of infrastructure.

And they’re growing fast.

The International Energy Agency expects global data-centre electricity use to more than double to around 945 TWh by 2030, slightly more than Japan uses today. Yet even then, data centres would consume less than 3% of global electricity.

Australia could feel it much more strongly. AEMO forecasts data-centre electricity use in the National Electricity Market rising from about 5 TWh in 2025–26 to 34 TWh by 2035–36, from roughly 3% to 13% of grid-supplied electricity in the NEM, which covers eastern and southern Australia.

That’s a huge change in a decade.

THE SIX QUESTIONS

Will AI raise your power bill, and who pays?
Can renewables really power AI 24/7?
Will data centres drain our water?
Will AI break the grid?
What about noise, land use and emissions?
Do we actually need all this AI?

→ Conclusion: The Choice Isn’t AI or the Environment

The problems are real. The question is whether they’re reasons to stop building, or problems we can solve by building better.

Energy Minister Chris Bowen recently called data centres electricity “whales.”

He’s right.

A 500 MW data centre running continuously could consume around 4.4 TWh a year. A 1 GW campus could theoretically consume 8.76 TWh. These aren’t ordinary electricity customers. They’re major industrial loads.

Ireland shows how quickly this can matter. Data centres went from 5% of Irish metered electricity consumption in 2015 to 23% in 2025.

So who pays when they arrive?

If a new data centre needs generation, transmission, substations and other grid upgrades, those costs shouldn’t simply be spread across households and existing businesses.

Big Tech shouldn’t get cheap infrastructure while everyone else gets the bill.

But there is another side. A data centre buying hundreds of megawatts for 10 or 20 years can help finance new solar farms, wind farms, batteries and transmission. Huge new demand can help create huge new supply.

And really, the principle isn’t that complicated:

Data centre? Sure. BYO energy.

If you’re bringing a massive new load onto the grid, help bring the new generation and firming needed to supply it, and pay your fair share of the network infrastructure required to connect it.

Australia is now moving toward exactly this kind of model: new, additional clean generation, backed by firm capacity, with large data centres expected to operate flexibly and pay their way.

There’s one more important question: what happens if the promised data centre never arrives?

If infrastructure is built for a 1 GW campus that is later delayed, downsized or cancelled, households shouldn’t inherit the stranded cost. Connection agreements can use performance bonds, phased charges and underwriting commitments so developers carry an appropriate share of that risk.

So yes, data centres could raise electricity costs if we get the rules wrong. But the answer isn’t particularly mysterious:

Bring the demand. Bring the energy. Pay for what you need. Carry the risk if your plans change.

It’s really not that difficult.

A data centre doesn’t stop when the sun goes down. So claims of “100% renewable energy” deserve scrutiny.

Buying enough renewable electricity or certificates over a year doesn’t mean renewable generation physically powered the servers every hour of that year.

At 2am on a still night, a certificate doesn’t power a server.

The electricity system does.

For Australia, the practical question is what new electricity we can build cheaply, reliably and quickly enough to match rapidly growing AI demand.

That points overwhelmingly toward solar, wind, batteries, transmission, flexible demand and firming.

And this is where the story gets interesting.

AI’s enormous electricity demand isn’t just a challenge for the energy transition. It could become one of its biggest accelerants.

Every hyperscale data centre creates an enormous new customer that needs enormous amounts of electricity, quickly and competitively. Globally, hyperscalers are already responding by contracting huge amounts of new renewable generation, storage and grid infrastructure.

They’re not doing this because they’ve suddenly become environmental charities.

They’re doing it because they need the electricity.

And scale has consequences. More solar and wind strengthens supply chains. More batteries make variable generation easier to firm. More transmission unlocks renewable resources. Flexible computing can help balance the grid. And enormous new demand can push these technologies further down their cost curves.

Australia has exceptional solar, strong wind resources and rapidly growing battery deployment. Batteries can soak up cheap daytime solar, move it into the evening and respond to grid disturbances almost instantly.

But batteries don’t yet solve every multi-day renewable shortfall at the scale we’re discussing. Today, firming can come from batteries, pumped hydro, demand response, stronger transmission and flexible use of existing dispatchable generation when needed. Longer-duration batteries and other storage technologies should increasingly take on that role.

Existing coal and gas will remain part of Australia’s grid during the transition. That’s simply the reality of today’s system, not an argument for building more of them.

The important question is:

What can deliver huge amounts of reliable new electricity at the lowest cost, at scale, in the timeframe we actually need it?

For Australia, increasingly that means renewables, storage, transmission and flexible demand.

This creates a fascinating possibility.

The industry accused of becoming one of the world’s great new energy problems could become a Trojan Horse for rebuilding the electricity system itself.

Can one solar farm run a data centre 24/7? No.

Can a rapidly expanding system of solar, wind, storage, transmission, flexible demand and firming do it? Yes.

And AI may help accelerate the buildout required to get us there.

This concern is real too.

Millions of processors produce enormous amounts of heat. Some data centres remove that heat using evaporative cooling, which can consume large amounts of water. At hyperscale, the numbers become significant. Google’s freshwater consumption across its operations is measured in billions of gallons annually.

But high water consumption isn’t unavoidable.

Microsoft’s newest AI data-centre design uses closed-loop direct-to-chip liquid cooling, recirculating coolant rather than continually evaporating municipal water. Microsoft estimates this can avoid more than 125 million litres of water per data centre each year compared with its previous average cooling-water intensity.

And what’s left? Microsoft says the remaining annual water use of an entire facility can be roughly equivalent to a single restaurant.

That’s a dramatic change.

Other options include recycled or non-potable water, air-cooled chillers, adiabatic cooling and hybrid systems. So water use is increasingly about technology and location, not an unavoidable consequence of AI.

But there are trade-offs. Reducing on-site water consumption can increase electricity use. Air-cooled and other water-saving systems may require more energy for fans, pumps or compressors, depending on the design and climate. If that extra electricity comes from water-intensive or high-emissions generation, some of the environmental burden has simply moved elsewhere.

Even a “zero-water” data centre therefore needs to be viewed as part of the wider energy system. The relevant measure isn’t just the water consumed inside the fence, but the water and emissions associated with supplying its electricity too.

This is another reason clean electricity matters. As the grid shifts toward solar and wind, which use very little water during operation, the trade-off between saving water at the data centre and consuming more water at the power plant becomes much smaller.

What works in Melbourne may not be optimal in Darwin or western Sydney. Australia is a dry country, so the basic rule should be straightforward:

Don’t use drinking water to cool computers when practical alternatives exist.

If a project wants large amounts of potable water in a water-stressed area, make the developer justify it.

And if it can’t?

Build it differently or build it somewhere else.

It could, if we get this wrong.

A giant data centre can be built faster than a major transmission line. AEMO expects NEM data-centre electricity consumption to increase roughly sevenfold over the next decade. That creates a serious engineering challenge.

This isn’t just about generating enough electricity. We also have to move enormous amounts of it to where it’s needed.

That means more transmission lines, substations, transformers, interconnectors and high-voltage infrastructure. Increasingly, it also means high-voltage direct current, or HVDC, which can move very large amounts of electricity efficiently over long distances and connect remote renewable resources with major population and computing centres.

China is pushing this harder and faster than anywhere. It has built the world’s largest network of ultra-high-voltage AC and DC transmission, moving enormous amounts of electricity across thousands of kilometres from energy-rich regions in the west and north toward major demand centres. Its national computing strategy is increasingly being developed alongside this enormous electricity system.

That’s an important part of the AI race that gets much less attention than chips and models.

Compute ultimately needs electricity, and electricity needs infrastructure.

Australia is moving too. Renewable Energy Zones and major transmission projects including EnergyConnect, HumeLink and Marinus Link are expanding the ability to move electricity between renewable resources, storage and demand centres. Other countries are confronting the same reality: falling behind on electricity infrastructure increasingly means falling behind on electrification, advanced industry and ultimately compute.

But building transmission is slower than building a data centre. Planning, environmental approvals, community consultation, land access, transformers and construction can take years.

So the grid can’t simply react to AI demand after it arrives.

It has to be built ahead of it.

And simply building more wires isn’t enough. The grid itself is becoming smarter.

A smart grid combines sensors, digital substations, advanced forecasting, automated controls, smart meters, batteries and flexible loads to continuously balance electricity across the system. AI can increasingly sit inside that system too, forecasting renewable generation and demand, detecting faults, managing congestion and deciding when storage or flexible loads should respond.

That becomes increasingly important as electricity decentralises. Instead of a relatively simple system built around a handful of large power stations sending electricity largely in one direction, the emerging grid has to coordinate millions of rooftop solar systems, batteries and EVs alongside wind and solar farms, transmission links and enormous new loads such as data centres.

Then there’s the operational challenge.

What happens if a 500 MW data centre suddenly disconnects? What happens to frequency and voltage? How do its batteries and UPS systems respond? Can its electricity consumption ramp up and down? Are the local transformers, substations and transmission lines designed for it?

When individual electricity customers approach the scale of power stations, these questions matter.

But data centres also have a major advantage: they’re programmable.

Not every computing task needs to happen immediately. AI training and other flexible workloads can potentially increase when electricity is abundant and reduce when the grid is stressed. Some workloads may eventually be shifted not only through time, but between locations.

Data centres already contain sophisticated power electronics, batteries and backup systems because even tiny power interruptions matter. Batteries can respond in milliseconds while non-urgent computing can potentially be delayed or shifted.

That means a data centre doesn’t have to behave like a giant appliance permanently switched on.

More computing when electricity is abundant. Less flexible computing when the grid is stressed. Batteries responding instantly when required.

Combine that with stronger interconnection, HVDC transmission, storage and a smarter grid, and the electricity system becomes far more capable of balancing geographically diverse solar, wind, hydro, storage and demand.

So the answer to rapidly growing AI demand isn’t simply build more generation.

It’s:

Build generation. Build storage. Build HVDC transmission. Build substations. Strengthen interconnection. Digitise the grid. Make the loads flexible.

China understands the scale of that infrastructure challenge and is moving extraordinarily quickly. Australia and other countries are moving too, because the alternative is increasingly obvious.

You can build the world’s most advanced data centre, fill it with the world’s most advanced chips and connect it to the world’s best AI models.

But without the electricity infrastructure underneath it, it’s just a very expensive building full of computers.

Done badly, data centres could strain the grid.

Done well, they could help finance the next generation of it.

And there’s another irony here: AI creates part of the electricity challenge, while AI itself could become one of the tools used to manage the smarter, more complex grid that supplies it.

Electricity and water aren’t the only concerns. Data centres are enormous industrial facilities, and where we put them matters.

They require land, substations, transmission connections, cooling equipment and backup power. Construction can mean years of heavy vehicles, noise and disruption. Once operating, cooling fans, chillers and transformers can run 24/7. A constant low-frequency hum that seems insignificant in an industrial precinct can be a very different proposition beside someone’s home at 2am.

Backup diesel generators are another legitimate concern. Hyperscale campuses can contain large numbers of them because power interruptions aren’t acceptable. They may operate infrequently, but testing and emergency use still create local noise and air pollution. Batteries, cleaner backup technologies and better grid connections can reduce that dependence over time, but today it remains a real impact that should be regulated.

Then there’s the footprint we don’t see: embodied carbon and materials. Before a data centre processes its first request, emissions have already been produced making its concrete, steel, transformers, cooling equipment, servers and GPUs. Semiconductor fabrication is particularly energy and resource intensive, and rapidly replacing generations of AI hardware also creates a growing stream of electronic waste. As the electricity powering data centres becomes cleaner, these construction, manufacturing and hardware impacts become a larger share of their lifetime footprint.

Land use needs perspective too. Data centres can occupy large sites, but the bigger question is which land. Building on appropriately zoned industrial land near transmission and renewable generation is very different from displacing housing, environmentally sensitive areas or higher-value uses. The same applies to visual impact and new transmission infrastructure: good siting can eliminate or greatly reduce many conflicts before they start.

Then there is the local economic bargain. Data centres can bring billions of dollars of investment, major construction activity, rates and infrastructure, but once operating they generally employ far fewer people than labour-intensive industries occupying a similar footprint. Communities are entitled to ask: what do we actually get in return? That means looking beyond headline investment figures to permanent jobs, local procurement, infrastructure contributions, skills development and whether the wider economic benefits remain in Australia.

None of this makes data centres uniquely destructive. Factories, warehouses, mines, roads, power stations, transmission lines, renewable projects and housing all consume land and materials and affect surrounding communities. The relevant question isn’t whether a data centre has a footprint. Everything we build has one. The question is whether that footprint is justified, minimised and properly paid for.

And many of these problems have straightforward solutions: locate facilities in suitable industrial areas, enforce strict noise limits, minimise diesel use, require responsible water and waste management, recycle hardware where practical, account for embodied emissions, protect sensitive land and make developers pay for the infrastructure and community impacts they create.

Some proposals still won’t stack up.

Too close to homes? Move it. Water constrained? Change the design. Grid can’t support it? Build the infrastructure first. Developer won’t pay its share? Don’t approve it.

Supporting AI doesn’t mean supporting every data centre.

Right technology, wrong design or wrong location can still be the wrong project.

This is ultimately the question that matters most.

If AI were only about generating pictures, writing emails and making chatbots, building gigawatts of new infrastructure for it would be much harder to justify.

But that’s not where this is heading.

AI is rapidly becoming a tool for medicine, biology, drug discovery, materials science, weather and climate modelling, energy systems, engineering, manufacturing, robotics and scientific research.

In medicine, AI can analyse medical images, proteins, genomes and enormous biological datasets, helping researchers identify patterns humans would struggle to find. In science, it can search vast numbers of possible molecules and materials, potentially accelerating discoveries in batteries, solar cells, catalysts and new medicines.

In weather and climate science, AI can improve forecasting and help researchers model increasingly complex systems. In energy, it can forecast electricity demand and renewable generation, optimise grids and industrial processes, and help us use resources more efficiently. In robotics, AI is beginning to give machines the ability to understand and interact with the physical world.

And across the economy, AI can increasingly perform cognitive tasks that previously required human labour.

That’s the opportunity.

Now comes the uncomfortable part.

There’s no particularly nice way to say this, and I don’t think we should sugar-coat it.

Data centres sit at the physical heart of what may become the largest technological disruption since industrialisation. That’s also one reason they attract so much of society’s anger. They aren’t just warehouses full of computers. They are physical infrastructure for a technological shift many people understandably find frightening.

Look at the coming wave of humanoid robots being developed across China and the West. Combine increasingly capable robots with increasingly capable AI and the direction becomes difficult to ignore.

We’ve seen technological disruption before. Cars replaced horses. Machines replaced enormous amounts of physical labour. Computers automated clerical work. The internet destroyed industries while creating entirely new ones.

Except this time there’s an uncomfortable difference:

We may be the horses.

Disruption doesn’t care whether we’re emotionally ready for it.

My own expectation is that AI and robotics will be capable of replacing the overwhelming majority of human labour somewhere around 2035–2040. That timeline is obviously debatable. Nobody knows exactly how quickly this will happen. But whether it takes 10 years, 15 years or longer doesn’t change the direction of travel.

As AI takes on more cognitive work and robots take on more physical work, societies will face questions much bigger than data-centre electricity consumption: what happens to employment, wages, taxation, inequality and ultimately the relationship between work and income?

Previous technological revolutions eventually created new jobs to replace many of those destroyed. We cannot simply assume that continues indefinitely. If machines can perform both cognitive and physical work more cheaply than humans, new jobs may increasingly fail to offset those displaced.

That’s why universal basic income, social dividends or other mechanisms for distributing the enormous productivity gains from automation may eventually move from political theory toward economic necessity.

That transition could be messy. It could be politically explosive. And it could be frightening.

But Australia opting out won’t stop it.

China won’t stop. The US won’t stop. Europe won’t stop. Companies won’t stop competing to build better AI and robots because another country finds the consequences uncomfortable.

Technological disruption doesn’t wait for consensus.

So Australia’s choice isn’t really whether this technological revolution happens. It’s whether we’re near the front of it or watching from behind.

Simply filling Australia with foreign-owned server farms isn’t enough. We should use data-centre investment to build Australian computing capability, research, startups, robotics, engineering, skills and advanced industries around it.

Because if this disruption is anywhere near as large as I expect, being 20 years behind the technological frontier isn’t the safer position.

It’s arguably the riskier one.

And Australia has one enormous advantage: energy.

We have extraordinary solar resources, strong wind resources, land, critical minerals and proximity to Asia. If AI increasingly turns electricity into economically useful intelligence, abundant cheap electricity becomes a computational advantage.

For generations Australia exported energy embodied in coal and gas.

This century we have an opportunity to turn abundant clean energy into something much more valuable: computation, medicine, science, automation and intelligence.

So yes, data centres have costs.

But we need to count both sides of the ledger.

The electricity, water, land and infrastructure are visible. A drug discovered faster, a better weather forecast, a new battery chemistry, a more efficient electricity grid, scientific breakthroughs and enormous productivity gains are harder to see when we’re standing outside a data centre looking at its cooling towers.

But they’re part of the equation too.

The disruption is coming whether we like it or not.

The real question is whether Australia helps shape it and captures some of its enormous value, or watches it happen from the sidelines.

AI has an environmental footprint. Pretending otherwise is pointless.

Data centres consume electricity and water. They need land, transmission, concrete, steel and semiconductors. They generate heat and noise. Poorly planned projects can leave communities carrying costs while technology companies collect the benefits.

Those concerns deserve answers. But they aren’t automatically arguments against AI.

They’re arguments for building it properly.

Make enormous new loads bring new electricity supply. Make developers pay their fair share of infrastructure. Don’t leave households carrying the risk if projects disappear. Use water-efficient cooling where water is scarce. Build the transmission and smart grids needed to support them. Make data centres flexible participants in the electricity system. Put them where they make sense. Reject the projects that don’t. And make sure Australia captures more than just the electricity bill.

But we also need to count both sides of the ledger.

It’s easy to measure the megawatts, litres of water, hectares of land, tonnes of concrete and dollars of infrastructure because they’re physically visible.

The output of all that computation is harder to see: better medicines, new materials, better weather and climate models, more efficient energy systems, scientific discoveries, automation and potentially enormous productivity gains.

Those benefits aren’t guaranteed. Neither are they zero.

So the real question isn’t:

Do data centres have an environmental footprint?

Of course they do.

It’s:

Can we make the benefits bigger than the costs?

I think we can.

And there may be one final irony. The enormous electricity demand being used as an argument against AI could itself help finance the renewable generation, storage, transmission and smarter grids needed to transform the electricity system.

The energy problem could become an energy accelerant.

But that requires something better than either blindly opposing data centres or writing Big Tech a blank cheque.

Australia needs to build the right infrastructure, in the right places, under the right rules.

Bring the demand. Build the energy. Build the grid. Pay your way. Carry your risk. Build Australian capability. Protect communities.

And build it properly.

6 questions about AI, energy and our future

THE SIX QUESTIONS

SOURCES & FURTHER READING