INTHEBLACK November 2026 - Magazine - Page 29
FROM DATA CENTRES TO AI FACTORIES
A 2025 report by Deloitte, The AI infrastructure reckoning:
Optimising compute strategy in the age of inference
economics, found that organisations are quickly discovering
their existing infrastructure may not be designed for
AI’s demands.
including utilisation, scalability, predictability,
electricity, financing, staffing and training,
latency, migration, hardware refreshes,
cybersecurity and regulatory exposure.”
Transaction costs can include data egress
charges [to move out of the public cloud],
migration planning, application refactoring,
integration work, testing, downtime risk, vendor
contract constraints and the internal effort
required to keep workloads portable. Once those
costs are included, a workload that appears
cheaper in the public cloud on a unit-price
basis may become more economical on-premise
if it is large, stable, storage-heavy or subject
to frequent data movement.
Salampasis, who is also a member of
CPA Australia’s Digital Transformation
Centre of Excellence, says transaction costs
can materially change the answer because
organisations need to understand each
workload, the strength of the business case
and how easily applications can be replaced
or moved. Any decision, he says, must be
grounded in financial rationality.
“Cloud can be economically strong
where uncertainty and flexibility matter, but
private infrastructure becomes more attractive
when demand is continuous, predictable,
storage-intensive or latency-sensitive.”
Talent, cybersecurity, compliance and
maintenance costs can move the equation in
either direction. On-premise environments may
require specialist engineers, security monitoring,
patching, hardware refreshes and audit controls,
while cloud can reduce some maintenance
burdens but introduce new requirements around
vendor governance, identity management,
configuration oversight, compliance evidence
and cost-control skills.
Sfondrini says that digital sovereignty
requirements often create additional
costs through local providers, duplicated
capabilities, stronger encryption, dedicated
infrastructure, portability investments and
tested exit strategies.
The report states that AI workloads are driving the emergence
of “AI factories”: integrated infrastructure ecosystems specifically
designed for AI processing that bring multiple specialised
components into a single solution. “These AI factories can also
offer excess computing capacity through service models, allowing
organisations to monetise unused processing power while
maintaining strategic control over critical workloads.”
Data centres will need to continue to evolve to accommodate
increasingly specialised tools for specific applications, it continues.
“The organisations that successfully navigate this infrastructure
transformation are likely to gain sustainable competitive
advantages in AI deployment and operation. Those that fail to
adapt are likely to face escalating costs, performance limitations
and strategic vulnerabilities as AI becomes increasingly central
to business operations.”
“Just as cloud computing reshaped IT strategy over the past
decade,” the report notes, “hybrid AI infrastructure will probably
define technology decision-making for the decade ahead.”
“Data residency can restrict where
workloads run. Provider requirements
can reduce access to global economies
of scale. Portability can require architectural
investments that would not otherwise
be necessary. Exit strategies can introduce
duplication and migration-readiness costs.”
THE CHALLENGE FOR CFOS
Sfondrini says all of these factors create
a new question for financial operations
areas, because the objective can no longer
be limited to determining the cheapest
technically viable architecture.
intheblack.cpaaustralia.com.au
29
