Research and version note This is a version 0.1 research draft in The Future Has No Representative. AI technology, policy, and markets change rapidly, and the article does not forecast the long-run position of a model or supplier. Policies and technical frameworks were reviewed to August 2026 and should be checked again in later versions.
The Future Has No Representative · Article 10
Quick read
AI infrastructure is more than a data centre or a large model. It includes compute, cloud contracts, data formats, identity and access systems, application interfaces, evaluation methods, workflows, and staff capability. A model may be replaced within a year while data pipelines, procurement relationships, and administrative procedures built around it last much longer. Future choice is often constrained at these less visible layers.
Constraint is not necessarily bad. Common interfaces, shared security standards, and long investment can reduce costs and stabilise service. Every usable system closes some possibilities. The questions are whether an institution knows what it has closed, who bears exit cost, and whether later institutions can still migrate, audit, or stop high-impact uses.
Supplier lock-in is only one form. Training and running AI affects electricity, land, water, and semiconductor supply chains. Data-centre siting and grid investment can outlive a model. If a public agency outsources its records, judgment processes, and professional capability, changing vendor may not restore the internal ability to understand the service.
Australia’s Policy for the responsible use of AI in government 2.0, effective in December 2025, requires covered entities to appoint accountable officials, make transparency statements, keep use-case registers, conduct impact assessments, train personnel, and monitor use. This is an important move from a one-off purchase toward life-cycle governance. Registering a use case does not by itself solve contractual portability, energy impacts, or the ability to retire a system.
Changes in purpose must also trigger review. If a model bought for document retrieval later enters staff screening, its name may be unchanged while rights consequences and evidential requirements are entirely different. Prior approval should not travel automatically merely because the system already exists.
AI infrastructure should be governed as a long-term public capability rather than a short software purchase. High-impact systems need readable and durable records, data and interface exit plans, replaceable components where appropriate, continuing human competence, and specified retirement conditions. The future need not receive every technical option. It should not be unknowingly locked into a system that only the original vendor can explain and that can be sustained only by continuing expansion under changing conditions.
The model is often not the durable part
Discussion about AI futures concentrates on capability: which model is stronger, how many parameters it has, and when another generation will arrive. Change is genuinely rapid and makes prediction fragile. Once an organisation deploys AI, however, the durable elements may form around the model.
Data are cleaned, classified, and connected to specified fields. Staff work is reorganised so some judgments pass through the system. Procurement defines service levels, log access, and intellectual property. Identity, network, and security arrangements are configured around a platform. Training and performance indicators stabilise the use. When a model name changes, this surrounding structure does not disappear at the same time.
Infrastructure is consequently physical and institutional. Facilities, chips, grid connections, and cooling supply computation. Data governance, interfaces, contracts, and professional habits cause it to enter decisions. Attention to hardware alone omits organisational lock-in; attention to algorithms omits commitments of energy and place.
The structure has legitimate value. Standardisation reduces duplicated work, long contracts may obtain reliable service, and specialised training improves quality. No system can remain completely open because compatibility with every future technology has unlimited cost. Governance should address the degree, significance, and visibility of lock-in.
How path dependence forms
Sunk cost is the most obvious mechanism. After an agency pays for migration, training, and customisation, changing course means paying again. A later technology may be superior while the established platform survives through conversion cost.
Data schemas create deeper constraint. A field initially introduced for a model enters forms, reports, and law. People begin to describe reality through it, and an implementation choice acquires institutional meaning. When the model is replaced, the input structure survives and the next tool is forced to accept an inherited classification.
Training and evaluation data can create benchmark lock-in. If an organisation repeatedly tests new models against the same set, suppliers optimise toward known measures while untested languages, disability needs, and rare situations remain invisible. A benchmark supports comparison but needs versioning, blind evaluation, and updates connected with real complaints. Otherwise “improvement” may mean only a better fit to yesterday’s measurement environment.
Network effects reinforce a path. Government, hospitals, or schools use one interface; consultants and suppliers organise around it, and labour is trained in the same technology. A common standard improves interoperability but makes alternatives difficult to establish. An open standard should be distinguished from a single implementation: a shared interface may expand choice, while one actual provider narrows it.
Organisational competence can decline in response. Outsourcing initially addresses a skills gap, but continuing dependence may leave the institution unable to define requirements, identify error, or manage updates. A contract can formally be competed while only the incumbent understands the data and system. This epistemic lock-in is harder to correct than a licence fee.
The non-digital commitments made by compute
Training AI and running inference at scale requires data centres, electricity, and networks. The International Energy Agency’s Energy and AI begins from the fact that AI requires electricity and examines demand over the coming decade alongside energy security, emissions, and affordability. Projections will change, but physical connections already make AI policy part of energy policy.
Data centres need sites, grid upgrades, backup supply, and cooling. Water and environmental effects vary substantially by technology, climate, and electricity source, so a global average cannot be assigned to every development. Local communities may bear land, noise, and network costs without receiving a proportional share of service benefits.
Approval should also consider whether facilities can be repurposed. A general grid enhancement may support several industries; highly specialised connections and buildings can be difficult to reuse if demand changes. Designing physical adaptability reduces the chance that one mistaken projection of AI demand becomes a long local liability.
Grids and buildings generally last longer than chips. If AI demand falls below projection, specialised assets may be stranded; if it exceeds projection, it can compete with housing, industry, and decarbonisation for electricity. Approval should stress-test several demand scenarios instead of treating one industrial forecast as destiny.
Domestic compute can contribute to sovereignty, resilience, and research capacity. Imported cloud service may reduce cost and obtain advanced technology rapidly. Neither is universally preferable. A plan should identify capabilities requiring local control, services suitable for external purchase, and how critical operations continue under outage or geopolitical change.
Why public data intensifies lock-in
Public bodies hold tax, health, welfare, education, and regulatory information. Using it in AI does not merely purchase efficiency; it reorganises how the state sees people. Choices about data, labels, and linkage errors can flow through several systems for years.
If a supplier retains derived data, embeddings, feedback logs, or tuning work, returning the original dataset at contract end may be insufficient. A later service can need those artefacts to understand historical decisions. Portability should include format, metadata, version, and transformation documentation, within privacy and security limits.
Nor is retaining everything appropriate. Indefinite retention increases privacy, breach, and wrongful reuse risk. A retirement plan states which evidence must survive for accountability, for how long, which personal data must be deleted, and how deletion is demonstrated. Future audit and present privacy have to be designed together.
Public records should allow reasons to be recovered. If a person challenges a benefit or licence years later, an agency needs to know the model version, inputs, and policy rule then in use. The model may no longer run, yet sufficient logging and decision documentation can permit reconstruction. Without it, technical updating becomes a means by which responsibility disappears.
Where the current Australian framework has reached
Australia’s Policy for the responsible use of AI in government, version 2.0 took effect on 15 December 2025 and applies to non-corporate Commonwealth entities, subject to stated exclusions. It requires accountable officials, transparency statements, adoption strategy and operational governance, internal AI use-case registers, training, and impact assessments for in-scope uses.
The policy requires continuing monitoring and revalidation when scope, use, or operation changes materially. A high-risk use case must be reported to an accountable official and governed by a designated board or senior executive, with review at least annually. Vendor-initiated changes are specifically among developments that an agency should monitor.
These obligations recognise that AI is not finished after pre-deployment approval. Models drift, suppliers update, and uses expand. Accountability and inventories give an organisation a chance to know which systems exist and create an entry point for incidents, complaints, and termination.
The framework mainly governs use-case risk and cannot automatically address national compute concentration, grid choices, or market dependence. An agency may complete an impact assessment and still enter a contract that is hard to leave. A transparency statement saying only that AI is used tells the public little about a high-impact decision. Procurement terms and quality of implementation determine whether options are actually preserved.
Is interoperability a complete answer?
Open interfaces, standard formats, and data portability usually reduce conversion costs. A modular design can replace a model without reconstructing a whole service, and shared tests permit comparison of vendors. These are important methods for preserving future choice.
They carry costs. Interface abstraction may support only common functions and impede a new technique. Open components increase integration and security responsibility. Maintaining several suppliers is expensive. Some high-performance systems rely on close hardware–software co-design, and compulsory separation can reduce reliability.
The objective should not be maximum modularity, but avoidance of unreviewable dependence at critical control points. Identity, authoritative records, policy rules, and appeal processes will often need to remain controllable by the public body. A model service may change; responsibility for the decision should not move with the interface.
Exit should be tested in practice at procurement. Can the agency export material? Can another team understand the documentation? Is there a manual path during outage? Does the contract permit independent audit? A clause labelled “portable” that has never been exercised may reveal dependence only during crisis.
When should a system be retired?
AI projects define deployment success more often than retirement. Falling performance, legal change, disappearance of purpose, end of supplier support, or risk exceeding benefit can each justify termination. Without prior criteria, sunk cost and organisational reputation favour continued use.
Stopping also introduces risk. Employees may have lost old skills, manual processes may not sustain volume, and other systems can depend on AI output. Alternative capability and treatment of unfinished cases, records, and appeals are required before closure. Exit planning for an essential service should begin before deployment.
Later institutions also need control over use expansion. A model procured for document retrieval may move into staff screening or fraud investigation. The technology is similar but consequences and evidential requirements have changed. Skipping new review because “the system already exists” is a common route by which path dependence becomes expanded power.
Use boundaries should therefore appear in access controls, logs, and staff training rather than exist only in an original project description.
The NIST AI Risk Management Framework places govern, map, measure, and manage across the life cycle and includes third-party hardware, software, and data. The framework is voluntary and under revision, but its life-cycle approach is valuable: change, appeal, override, and retirement are elements of risk management, not additions to an initial accuracy test.
Provisional judgment: preserve the capacity to replace, not every product option
AI infrastructure built today will constrain the future. If investment, standards, and skills formed no path, a system could not provide stable service. The task is not to eliminate lock-in but to prevent present convenience from creating inescapable dependence disproportionate to public consequence.
A high-impact system should preserve several capabilities. Its operators need to know what is in use, including the relevant version, while critical records and policy rules remain outside a supplier’s monopoly. Migration and retirement should be technically possible, supported by contracts and by staff who still understand the service and accept responsibility for it. Physical investments should remain useful under more than one demand scenario. Each of these capabilities needs practical verification rather than existence in principle.
Future people have no right to demand that we guess their preferred AI architecture. They can reasonably object to losing basic political and technical capacity through current procurement. If changing a model makes public service impossible, or auditing an old decision requires a company that no longer exists, infrastructure has acquired power beyond software.
A technical choice becomes an institutional environment. It affects which data can be seen, which judgments can run automatically, and who can modify them. Present decision-makers should appraise this formative effect. The useful inheritance is not an advanced model. It is a technical structure that later people can understand, correct, and leave when necessary.
Primary sources and further reading
- Australian Government Digital Transformation Agency, Policy for the responsible use of AI in government, version 2.0, effective 15 December 2025.
- Australian Government, National framework for the assurance of artificial intelligence in government.
- NIST, AI Risk Management Framework and AI RMF Core.
- International Energy Agency, Energy and AI, 2025.
- OECD, AI compute and Measuring the Environmental Impacts of AI Compute and Applications.
Series navigation: The Future Has No Representative — series overview
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.