Why the next generation of AI data centres may need to be designed together with their energy infrastructure

Ahilan Raman | Managing Director, Clean Energy and Water Technologies Pty Ltd (CEWT)

The global AI race is usually discussed in terms of GPUs, models, semiconductor supply and computing capacity.

But another constraint is rapidly becoming just as important:

Where will the megawatts come from — and how quickly can they be delivered?

An AI data centre can increasingly be constructed using repeatable, modular infrastructure. The electricity network supplying it operates on a very different development timescale.

That mismatch is beginning to reshape data-centre engineering.

The power system is becoming part of the data centre

Australia provides a particularly interesting example.

AEMO reported that, by the end of the March 2026 quarter, 11 large data-centre projects above 5 MW, representing 5.4 GW of maximum demand, were already progressing through the National Electricity Market transmission-connection process. Around 40% of that capacity was in Victoria. AEMO said current experience suggested approximately a two-year application-to-energisation timeframe, although individual projects vary.

The policy direction is also changing.

In August, the Australian Energy Market Commission recommended that data centres should bring new clean, firm capacity, operate flexibly and make efficient use of the network. It specifically identified co-location with generation as one way of reducing network pressure.

This points toward an important change in thinking.

Instead of treating electricity as a utility connection that is considered after the data centre has been designed, perhaps the data centre and its energy system increasingly need to be engineered as one infrastructure project.

Why 20 MW is an interesting scale

Not every AI facility needs to begin as a 500 MW or 1 GW hyperscale campus.

There is growing evidence for modular deployment at much smaller increments.

Australian developer QORINAI describes a delivered modular project comprising eight transportable 2.5 MW modules, with capacity contracted in stages from 1.9 MW to 20 MW over five months. Its current development model describes repeatable campus blocks of approximately 10–50 MW, with dedicated substations, cooling plants and data halls.

That is significant.

A 20 MW facility is large enough to support substantial AI computing infrastructure, but still small enough to consider a dedicated energy island and staged deployment.

And there is no reason that the modularity of the computing infrastructure must correspond exactly with the modularity of the power plant.

A data centre might comprise several independently deployable computing blocks while a multiple-engine power island feeds a common electrical bus.

Modularity should occur where it makes engineering and commercial sense — not because every subsystem must have the same module size.

AI is also changing cooling

Increasing rack densities are making the old distinction between electrical infrastructure and cooling infrastructure less useful.

QORINAI, for example, is designing Australian AI infrastructure around direct-to-chip liquid cooling and cites design points as high as 250 kW per rack.

As rack density rises, cooling becomes an increasingly important part of the total energy balance.

That raises another engineering question:

Why design power generation and cooling independently if the power plant is simultaneously producing useful thermal energy?

A conventional assessment may look primarily at electrical efficiency.

A trigeneration assessment asks a broader question:

What useful outputs can we obtain from every unit of primary energy entering the system?

Electricity is one output.

Recoverable heat is another.

Cooling produced from otherwise recoverable thermal energy can become a third.

For AI infrastructure, that distinction matters.

Behind-the-meter power is moving into the mainstream

The trend is already visible internationally.

Reuters reported in late September that demand for smaller gas turbines is rising as data-centre developers pursue rapid behind-the-meter generation to avoid grid-connection delays and long lead times for large turbines. Enverus projects 29.6 GW of behind-the-meter gas generation additions by 2030, with data centres accounting for 88% of that capacity.

A few days later, Enerflex announced a contract to engineer and assemble 450 MW of behind-the-meter natural-gas generation for a North American data-centre developer.

The attraction is straightforward.

Instead of asking:

“When will the grid be ready for my data centre?”

the developer can begin asking:

“Can I bring firm power to the data centre?”

But there is an important problem.

Simply moving conventional fossil generation behind the meter solves the time-to-power problem without necessarily solving the carbon problem.

That is where CEWT is exploring a different architecture.

Can the carbon itself be recycled?

At Clean Energy and Water Technologies, we are developing a 20 MW grid-independent Carbon Recycling Technology (CRT) Trigeneration concept for data-centre applications in Victoria’s Latrobe Valley.

The principle is different from simply installing gas engines beside a data centre.

Methane provides firm power.

The resulting carbon dioxide is captured.

Instead of treating that CO₂ purely as waste for disposal, CRT treats carbon as a circulating process material.

Captured CO₂ is combined with hydrogen-rich syngas and renewable hydrogen and converted back to methane through methanation.

The methane returns to power generation.

The carbon therefore circulates through:

Methane → Power → CO₂ → Methanation → Methane

Renewable hydrogen supplies new energy into that carbon cycle.

The objective is not to claim that carbon disappears.

Quite the opposite.

Every molecule of carbon must have a destination.

That requires a rigorous carbon ledger covering methane, CO, CO₂, recycle, inventory, purge, vents, and losses independently of the plant’s energy balance.

Trigeneration changes the system boundary

Our current 20 MW concept combines a multi-engine firm-power island with carbon recovery, H₂-rich syngas, renewable hydrogen and methanation.

But electricity is only part of the architecture.

Gas-engine thermal energy and the exothermic heat released during methanation represent potentially useful energy streams.

For a data centre, that heat can potentially support absorption cooling or other thermal services alongside conventional high-density liquid-cooling infrastructure.

The appropriate measure therefore becomes broader than generator electrical efficiency alone.

It becomes:

How much useful computing-supporting infrastructure can we obtain from each MW of primary energy?

That is a different optimisation problem.

Why Latrobe Valley matters

There is another reason we believe Latrobe Valley deserves attention.

The region is already attracting major data-centre interest.

Keppel has secured rights over a 123-hectare site near Hazelwood for a proposed data-centre campus with potential access to 720 MW of gross power. The Victorian Government identifies existing energy infrastructure, industrial land and access to sustainable raw water among the site’s advantages.

Latrobe City Council’s September update says the Hazelwood North project has not yet lodged its planning approval and that community consultation is expected later in 2026.

That is important context.

It suggests Latrobe Valley is not merely a former power-generation region searching for a new purpose.

It could become part of Australia’s next generation of energy-intensive digital infrastructure.

A smaller 20 MW project can play a different role from a 720 MW hyperscale campus.

It can demonstrate an alternative architecture.

From grid connection to energy platform

For decades, the conventional development sequence has effectively been:

Land → Grid Connection → Data Centre.

The emerging model could increasingly become:

Land → Firm Energy Platform → Data Centre → Grid Integration when appropriate.

That does not mean abandoning the electricity grid.

Nor does it mean every data centre should generate all of its own power.

It means recognising that time-to-power has become a fundamental project-design variable.

And once power generation moves closer to the computing load, engineers have an opportunity to reconsider the whole system: electricity, cooling, water, carbon, hydrogen, storage and waste heat.

For CEWT, the engineering philosophy remains straightforward:

Every MW must have a source and destination.

Every molecule must have a source and destination.

Every litre of water must have a pathway.

The AI infrastructure race may therefore not ultimately be won only by whoever has the most advanced GPUs.

It may also be won by those who can provide those GPUs with dependable megawatts, cooling and supporting infrastructure at the speed at which AI capacity needs to be deployed.

Time-to-power is becoming a design parameter.

And that may fundamentally change how we design the next generation of data centres.

Sources

• AEMO, “Digital demand surge”, 2026.
• Australian Energy Market Commission (AEMC), electricity/data-centre policy developments, 2026.
• QORINAI, Australian AI data-centre and modular infrastructure information.
• Reuters, reporting on behind-the-meter gas generation for data centres, September 2026.
• Enerflex, behind-the-meter generation project announcement, 2026.
• ABC News, reporting on Keppel’s proposed Latrobe Valley/Hazelwood data-centre development, January 2026.
• Latrobe City Council, Proposed Developments in Our Region, September 2026.