With 28% planning to repatriate workloads, how can infrastructure leaders build for AI without repeating cloud’s mistakes?

Infrastructure leaders are being asked to give businesses faster access to AI, protect sensitive data and keep an increasingly variable compute bill under control.

However, data sovereignty is now the leading constraint to scaling AI, while repatriation intent has reached an all-time high.

ADAPT research captures the tension.

While 57% expect future AI demand to be supported through public cloud GPUs, 28% plan to repatriate workloads in 2026, up from 17% in 2024 and 10% in 2020.

Only 1% say their infrastructure is fully ready for AI, 29% have no defined strategy for future capacity, and a third expect compute spending to rise by more than 20% over the next two years.

Opening ADAPT’s 14th Cloud & Infrastructure Edge in Sydney, Head of Strategy Joey Meynink framed the challenge around three questions: can infrastructure keep pace, what is preventing AI from scaling, and who owns the compute bill?

Across discussions with more than 160 cloud, infrastructure and architecture leaders, one conclusion emerged: there is no single destination for every workload.

Leaders need an infrastructure portfolio that can adapt as economics, sovereignty requirements and AI architectures change.

Rebalance workloads before AI locks in the next cost base

Rising repatriation does not represent a wholesale retreat from cloud.

It signals that organisations are correcting workload placement decisions made without sufficient attention to long-term economics, operational characteristics or control.

ADAPT Head of Analytics & Insights Gabby Fredkin found that organisations are primarily moving monolithic, run-the-business applications such as ERP platforms, HR systems and databases where cost predictability or data sovereignty is important.

The median organisation planning to repatriate expects to move 9% of its workloads.

This points towards selective rebalancing rather than reversing entire cloud strategies.

ADAPT Senior Research Director Matt Boon framed the placement decision around four considerations: sovereignty, security, sensitivity and scale. These factors give leaders a stronger basis for deciding where a workload should run than broad cloud-first or on-premises-first mandates.

David Walker, former Group CTO at Westpac and DBS and an ADAPT Advisor, described how both banks used compute economics to determine placement.

Approximately 70% of their workloads were more economical on infrastructure the banks operated themselves.

These were predictable applications that could be run efficiently at scale using internal capability.

The remaining 30% benefited from cloud economics, particularly bursty workloads that required specialised compute for limited periods.

At DBS, teams could compare cloud pricing and move compute between providers based on the economics at the time.

The 70:30 split is not a universal benchmark. Its value lies in the discipline behind it.

Each workload was assessed according to its operating profile, performance requirements, risk and cost rather than being directed towards a predetermined platform.

AI makes that discipline more important because sovereignty now extends beyond the location of data.

Lisa Yourmann, Acting First Assistant Secretary for Enterprise Systems at the Department of Defence, explained that organisations may also need control over models, training pipelines, model weights, audit trails and the decisions or recommendations systems produce.

For Defence, sovereign control includes the ability to govern, operate, assure, modify and replace AI capabilities throughout their lifecycle. It also requires Australian expertise capable of building, evaluating and sustaining those systems.

This changes workload placement from a one-time infrastructure decision into an ongoing governance process.

As models, commercial terms, geopolitical risks, token economics and infrastructure architectures continue to evolve, workload placement decisions will need to evolve with them.

Infrastructure leaders therefore need to design for reversibility.

Applications, data and models should be able to move between public cloud, sovereign cloud, private infrastructure and edge environments without forcing a prohibitive commercial or business-process redesign.

The organisations that avoid repeating earlier cloud mistakes will treat placement as a continuously reviewed portfolio decision rather than a permanent home.

Back to top

Modernise the production systems that cap AI performance

AI ambitions will remain bounded by the systems on which organisations already depend.

Simon Davies, General Manager of Core Banking at CBA, argued that enterprise AI is a production-system problem before it becomes a model problem.

CBA’s core banking platform processes around 90% of the bank’s accounts and approximately 40% of money movements across Australia.

It handles about 70 million transactions a day for 17 million customers.

The mainframe supporting that environment remained reliable, but its architecture limited scalability, created friction around real-time data and depended on increasingly scarce specialist skills.

One incident exposed the operational concentration risk. A highly specialised fault could be understood and repaired by only two people globally.

Doing nothing would have avoided the immediate risk of migration, but it would have allowed strategic and operational fragility to compound.

CBA migrated the platform in under 18 months, using AI to refactor thousands of code objects, automate approximately 5,500 test cases and analyse downstream outputs for anomalies and variances.

The first attempted cutover encountered a network route-table problem and had to be rolled back.

The team executed that rollback without panic, blame or disruption because it had repeatedly rehearsed both successful and unsuccessful scenarios. It returned and completed the migration on the next attempt.

The resulting platform operates 30% faster and at 30% lower cost.

t moves around 26 million signals each day in real time and has processed billions of signals since going live.

Processes that previously took days can now be completed while the information remains operationally relevant.

The technical architecture was only one part of the change.

Eight narrow specialist teams were reorganised into cross-functional squads with full-stack accountability.

Mainframe specialists became some of the organisation’s strongest cloud engineers because they combined deep workload knowledge with new platform capability.

Legacy risk cannot be measured by system age alone.

It sits in data friction, concentrated knowledge, manual operations, rigid team boundaries and an inability to respond safely when conditions change.

Modernisation should therefore be judged by whether it increases the capacity of production systems to support real-time data, automation, observability and controlled autonomy.

Moving a workload without changing those conditions may update the technology while preserving the constraint.

Back to top

Govern tokens before AI becomes everybody’s cost and nobody’s responsibility

AI is changing infrastructure expenditure from a largely provisioned cost into a continuously moving meter.

Joey Meynink challenged leaders to determine who should own that meter.

Does the cost sit with infrastructure, the CIO, the AI program or the business unit consuming the capability?

Without a clear answer, AI compute becomes a shared resource that many teams use but no executive fully manages.

The readiness gap is significant.

ADAPT research shows 77% of infrastructure leaders are under board-mandated optimisation targets, yet only 14% have mature FinOps practices with accurate allocation, forecasting and optimisation capability.

Gabby Fredkin explained why traditional reporting cycles are insufficient for AI.

One organisation exhausted its annual token budget within three months while using AI to transcribe meetings and distribute notes.

The immediate technical use case worked, but it automated an activity whose underlying value had not been challenged.

The organisation also discovered the overspend only when its scheduled reporting process surfaced it.

Token consumption requires more immediate visibility because the cost continues to accumulate with every interaction, retry and workflow.

Model costs are also only one component of the total.

David Walker cautioned that an agent may depend on CPUs, databases, storage, retrieval systems, orchestration platforms and other services.

Production costs can also include human verification, exception handling, compliance controls, escalation paths and failed decisions.

Infrastructure and finance teams therefore need to measure the total cost of a successful business outcome rather than the price of an individual token or model call.

This requires clearer decision rights.

The business owner should remain accountable for the outcome and its budget.

Infrastructure teams should provide placement, architecture and consumption guardrails.

Finance teams need near-real-time allocation and forecasting.

Risk and data leaders must be able to see which models are operating, what information they can access and what actions they can take.

That visibility remains incomplete.

Half of the AI models currently being piloted or deployed are not covered by a formal governance framework.

Cost, risk and observability are becoming part of the same infrastructure problem.

Without model inventories, routing controls, usage thresholds and accountable owners, leaders cannot determine whether AI is scaling value or simply scaling consumption.

Back to top

Build for the inference wave without overcommitting to today’s architecture

Much of the current AI infrastructure market is being shaped by a small number of companies training frontier models.

Dean Nelson, Founder and Chairman of Infrastructure Masons, estimated that around 90% of current AI infrastructure demand supports training, while only 10% supports inference.

He expects that balance to reverse as enterprises begin running specialised models, agents and AI-enabled services throughout their operations.

Inference is where organisations apply models to decisions, workflows and customer interactions.

It is also likely to be more distributed than the current training market.

Instead of relying exclusively on gigawatt-scale AI factories, organisations may deploy smaller clusters closer to their users, operational data and physical environments.

That shift could materially change infrastructure requirements.

Gabby Fredkin found that 64% of infrastructure leaders believe onshore inference will be critically important for large language models.

Organisations are also reconsidering locally hosted and smaller language models where frontier-scale capability is unnecessary.

In the Cloud & Infrastructure Edge panel, Dr Amr Hassan, Director of Emerging Technologies and Program Director of MAVERIC at Monash University, argued that Australia’s strongest opportunity may lie in models fine-tuned for specific datasets, industries and problems.

A specialised model can provide stronger control, greater energy efficiency and a clearer business outcome than a general-purpose model, provided the organisation has the capability to operate it.

Large models will continue to play an important role, but they should not become the default for every workload.

The challenge for infrastructure leaders is that the underlying technology is changing far faster than the assets supporting it.

Data centres, electrical systems and cooling infrastructure are long-lived investments.

Chip architectures, rack densities and model designs can shift within a few years.

Dean Nelson highlighted rapid changes in power density, liquid cooling and electrical architecture that could force facilities designed around today’s standards to reinvest much earlier than their original financial models assume.

David Walker raised the same problem at a national level.

Australia needs more AI compute, but current demand does not provide a reliable blueprint for the next 35 years.

Models may become smaller and more efficient.

More inference may move to the edge.

Organisations may favour use-case-specific models or direct access to hardware rather than hyperscaler services.

Building too narrowly around the current training wave could leave organisations with the wrong type of capacity.

Infrastructure plans should therefore use multiple demand scenarios and modular designs.

Leaders need to test how quickly facilities can accommodate new power, cooling and hardware requirements, whether models and workloads can move between environments, and how capital commitments change if inference grows differently from current forecasts.

The objective is not to predict a single winning architecture.

It is to avoid making irreversible investments based on the least stable period of the market.

Back to top

Earn permission to scale sovereign compute

Domestic compute capacity is becoming part of Australia’s economic resilience, but that does not make every proposed data centre strategically valuable.

The rest of the Cloud & Infrastructure leader panellists discussed how new capacity would generate economic value, support host communities and contribute to the energy transition.

Andrew Sjoquist, Founder and CEO of WinDC, argued that the value of data centres should be measured through the industries and services they enable.

Hospitals, banks, airlines, government services and digital platforms depend on the infrastructure housed within them. AI will extend that dependency as more industries use compute to increase productivity, accelerate research and improve decisions.

Greg Boorer, Founder and CEO of CDC Data Centres, described sovereign compute as an economic and national security capability.

A country dependent on overseas models and infrastructure can lose access far more quickly than it could replace a disrupted physical resource.

Local compute provides greater control over the intelligence on which critical organisations may increasingly rely.

Belinda Dennett, CEO of Data Centres Australia, added that hosting infrastructure locally gives Australia greater influence over how its laws, values and security requirements are reflected in the services being developed.

Australia also has the potential to strengthen domestic capability while serving as a trusted regional location for AI infrastructure.

Political stability, renewable-energy potential, a strong rule of law and security relationships make that opportunity credible.

The first test, however, is whether the infrastructure strengthens Australia’s own resilience, research capacity and enterprise productivity.

Regional opportunity should follow from strong domestic capability rather than substitute for it.

That capability also needs public legitimacy.

Dean Nelson identified power, people, public pushback, planetary constraints and policy as five connected pressures on the digital infrastructure ecosystem.

Projects cannot scale when energy is unavailable, skilled trades are scarce, standards are unclear or communities believe they are carrying the costs without receiving the benefits.

The industry’s historically low public profile has become a liability.

Large facilities make AI physically visible, while limited information about their purpose, water use, energy demand and local contribution allows anxiety to fill the gap.

Dr Amr Hassan argued that Australia needs mandatory minimum standards for environmental performance, safety and community engagement, while preserving enough flexibility for different technologies and locations.

Greg Boorer and Andrew Sjoquist also showed how data centres could support the energy transition when they are placed near available generation, underwrite new renewable projects and participate in demand response.

These outcomes are not automatic.

Poorly located facilities, weak operators and opaque planning can impose costs on communities and damage trust across the wider sector.

Sovereign infrastructure therefore requires more than local ownership or local data residency.

It requires credible standards, transparent evidence, appropriate sites, workforce planning and a clear explanation of how the investment benefits the country and the communities hosting it.

Back to top

Recommended actions for cloud and infrastructure leaders

The next stage of AI infrastructure planning requires leaders to replace fixed platform strategies with repeatable placement, funding and governance decisions.

• Classify workloads and set placement triggers. Assess sovereignty, security, performance, scale and economics, then define when workloads should move between environments.
• Design for reversibility. Build portability into architecture and contracts before lock-in makes change costly.
• Extend FinOps across AI. Track tokens, models, agents, infrastructure and human oversight against business outcomes.
• Assign clear ownership and visibility. Give each AI workload an accountable owner, budget and risk profile, supported by a current model inventory.
• Modernise the foundations beneath AI. Improve data flows, automation and team accountability before adding AI to critical legacy systems.
• Plan for multiple infrastructure scenarios. Model different paths for inference growth, power demand, hardware density and edge deployment.
• Build sovereign capability, not only capacity. Pair domestic compute investment with the skills, energy resilience and community outcomes needed to sustain it.

AI is pushing infrastructure planning beyond fixed platform choices.

Workload placement will increasingly depend on how quickly organisations can reassess cost, control, performance and risk as models, architectures and market conditions change.

That puts greater weight on reversibility, clearer ownership of consumption and infrastructure designed to accommodate several possible futures rather than one settled forecast.

The next phase of cloud and infrastructure strategy will be shaped less by where workloads sit today than by how easily they can move, scale and remain governable as AI demand evolves.

Back to top