Distributed compute, distributed energy

For most businesses, computing has spent the past decade moving in one direction: away from the premises and into increasingly large cloud data centres.

AI is beginning to complicate that model.

The hyperscale cloud is not disappearing. But enterprises now have increasingly credible options to deploy AI infrastructure on their own sites, at the edge of their networks or within smaller private data environments. HPE Private Cloud AI, for example, is designed for on-premises inference, retrieval-augmented generation and fine-tuning. Lenovo's ThinkEdge SE450 is a compact server designed for AI workloads at the edge, while Schneider Electric's micro data centres integrate power, cooling, security and monitoring in a smaller edge-ready footprint.

The reasons vary. Sensitive information may need to remain within an organisation's environment. Some workloads benefit from lower latency. Industrial facilities may need processing close to machines and sensors. Other businesses may want greater control over models, governance and infrastructure rather than relying entirely on public cloud services.

The result is unlikely to be a wholesale migration away from the cloud. It is more likely to be a more distributed computing architecture, in which hyperscale facilities coexist with private data centres, edge servers and increasingly powerful compute deployed directly at commercial and industrial sites.

That shift has an energy consequence.

A small data centre is no longer necessarily a small load

The physical density of AI infrastructure is increasing unusually quickly.

NVIDIA specifies approximately 120 kW of rack power consumption for a GB200 NVL72 rack. That is already materially different from the power density associated with conventional enterprise IT.

The direction of travel is more significant again. NVIDIA says traditional 54 VDC power distribution encounters practical physical constraints as rack power climbs, and its next-generation 800 VDC architecture is being developed to support racks ranging from around 100 kW to more than 1 MW. Full-scale 800 VDC data-centre production is intended to align with NVIDIA's Kyber rack-scale systems from 2027.

This does not mean every business will install a one-megawatt AI rack. It does mean that the relationship between computing infrastructure and electricity infrastructure is changing.

At these densities, power is no longer simply a service delivered to the server room. Connection capacity, cooling, electrical distribution, resilience and the behaviour of the load itself become part of the infrastructure decision.

As compute becomes more distributed, electricity increasingly becomes part of the compute architecture.

Australia is already seeing the aggregate effect

The growth of data infrastructure is becoming visible at power-system scale.

AEMO estimates that data centres consumed around 4 TWh of electricity across the National Electricity Market in FY2025, equivalent to approximately 2.2% of total grid demand. Under AEMO's Step Change scenario, data-centre consumption rises to around 12 TWh by 2029–30, equivalent to approximately 6% of NEM grid-supplied electricity.

Much of that demand will continue to come from very large centralised facilities. But the same technology trends are also making smaller, distributed deployments more capable. That creates a second infrastructure question alongside the challenge of powering hyperscale data centres: what happens when meaningful new computing loads begin appearing across existing commercial and industrial sites?

Energy flexibility at the same node

A commercial site adding AI infrastructure may not need the scale of a hyperscale data centre to create a meaningful electricity problem.

A new high-density compute load can increase site peaks, reduce spare connection capacity and add demand at exactly the times when the local network is already constrained.

In that environment, distributed battery storage can become more than a standalone energy asset. Located close to the load, storage can potentially help a site:

• manage peak imports by supplying part of a short-duration demand spike locally rather than drawing the full requirement through the grid connection;
• make better use of an existing connection where instantaneous demand would otherwise approach the agreed import limit;
• smooth fast-changing loads, making the electricity demand presented to the network less volatile;
• shift the timing of grid purchases, charging when capacity is more available and discharging when site demand is highest; and
• provide an additional layer of resilience as part of a broader site-power architecture.

This does not make network capacity irrelevant. A battery cannot indefinitely support a sustained load that materially exceeds the capability of the site's grid connection. Large new computing loads still require appropriate network capacity, electrical infrastructure and, in many cases, network augmentation.

But storage can change the shape of the problem. Rather than sizing every element of the electricity system around a site's highest momentary demand, distributed energy infrastructure can help separate instantaneous power requirement from continuous grid requirement. That distinction becomes increasingly valuable as load density rises.

The compute industry is already moving in this direction

The same principle is beginning to appear inside the data centre itself. NVIDIA's 800 VDC architecture includes energy storage to help handle load spikes and sub-second GPU power fluctuations. Its more recent technical work describes multi-timescale storage as a way to buffer volatile, synchronous AI workloads from the stability requirements of the grid.

At the rack level, the objective is to manage extreme power density and changing compute demand. At the site level, the infrastructure question is broader but conceptually similar: can energy storage help isolate the electricity network from some of the volatility and peak requirements created by increasingly concentrated compute?

That question extends well beyond traditional data centres.

From buildings to infrastructure nodes

A commercial or industrial site is increasingly becoming a collection of interconnected energy loads. AI compute may sit alongside electric vehicles, industrial equipment, refrigeration, HVAC, rooftop solar and other electrified processes.

Treating those systems independently can lead to an increasingly complicated relationship with the grid. A different model is to think of the site itself as an infrastructure node.

In that model, distributed energy is not simply about installing batteries at more locations. It is about putting flexibility closer to where electricity is actually being consumed.

Distributed infrastructure for a more distributed economy

The first phase of the AI infrastructure boom has focused heavily on a simple constraint: finding enough power for enormous data centres. That remains an important challenge.

But a second question is emerging alongside it. If AI infrastructure becomes increasingly distributed across factories, logistics centres, hospitals, offices, campuses and other commercial sites, how should the electricity system around those sites evolve?

The answer will differ by location. Some sites will need larger connections. Some will need network investment. Some will remain almost entirely dependent on centralised computing. Others may increasingly combine compute, storage and intelligent energy management behind a single grid connection.