The cost of AI is not only hidden in data centres. It is hidden in the systems we are learning to depend on.
Artificial intelligence still feels cheap. A manager pays for a monthly subscription. A developer opens a coding assistant. A student asks a chatbot to summarise a report. The interaction is instant, polished and priced in a way that makes the cost feel almost trivial.
That impression is unlikely to last.
The first phase of consumer AI was built around the prompt: one user, one question, one answer. The next phase is being built around agents. These systems do more than respond. They search, plan, call tools, inspect files, write code, check their work, retry failed steps and continue until a task is complete.
That shift changes the economics. The relevant unit is no longer the prompt. It is the workflow.
This matters because much of the public debate still treats AI as a software product. In practice, it is becoming an infrastructure layer. It depends on data centres, electricity, water, chips, cloud providers, model vendors, monitoring systems and compliance processes. As adoption spreads, the cost moves beyond the subscription price.
The early signs are already visible.
GitHub has said Copilot will move to usage-based billing from June 2026. Instead of counting premium requests, usage will be calculated through AI credits linked to token consumption, including input, output and cached tokens. Paid plans will include a monthly allowance, with additional usage billed separately.
For users, this may look like a pricing update. For the market, it is a signal. AI is beginning to move away from the flat-rate psychology of software-as-a-service and toward the metered economics of cloud infrastructure.
A short chatbot reply and a long coding-agent session are not the same thing. The latter may involve multiple model calls, large context windows, file analysis, code edits and repeated attempts to solve the problem. The user sees one task. The system may perform hundreds of computational steps behind the interface.
That is where the idea of cheap AI starts to fray.
The subsidy phase
The current AI market has many features of a subsidy phase. Large technology companies are competing for market share, user habits and strategic position. Prices are simple because simple pricing accelerates adoption. Free tiers reduce friction. Flat-rate plans make experimentation feel low-risk. Enterprises are encouraged to deploy before they fully understand the cost of scaled usage.
There is nothing unusual about this. Search, cloud storage, ride-hailing and food delivery all used cheap access to change behaviour before their economics became clearer. AI may follow the same pattern, but with a much heavier infrastructure burden.
The counterargument deserves serious attention. AI is becoming cheaper at the unit level. Model inference costs have fallen sharply. Hardware is improving. Smaller models are becoming more capable. Energy efficiency is rising. A narrow claim that AI costs can only go up would be wrong.
But lower unit costs do not guarantee lower total costs. When technology becomes cheaper and easier to use, demand often expands. Once AI is embedded into office software, coding tools, customer support, finance, logistics, marketing, compliance and public administration, total consumption can grow faster than efficiency gains.
The important issue is not whether one query becomes cheaper. It probably will. The issue is whether aggregate AI demand grows faster than the systems built to support it.
Also Read: The transformation ecology crisis: How AI is exposing the hidden fragility of high-performing teams
The data centre bill
The International Energy Agency expects global electricity consumption from data centres to more than double by 2030, reaching about 945 terawatt-hours. That would represent just under 3 per cent of global electricity consumption. From 2024 to 2030, data centre power demand is expected to grow at roughly 15 per cent a year, more than four times the rate of demand growth in other sectors.
AI is one of the main drivers. In the United States, data centres are expected to account for nearly half of electricity demand growth between now and 2030.
Those figures do not mean AI will overwhelm power grids everywhere. They do show that AI is no longer only a software story. It is becoming a major industrial load.
That distinction matters. Software scales quickly. Power systems do not. Data centres can be planned, financed and built faster than new transmission lines, generation assets and regulatory approvals can be delivered. The North American Electric Reliability Corporation has warned that large data centre loads can arrive faster than the infrastructure needed to serve them.
The mismatch is obvious. AI companies move at software speed. Grids move at infrastructure speed.
This is where the cost may move outward. If utilities need to build new generation, strengthen transmission and reserve firm capacity for data centres, someone has to pay. In some cases, the technology companies will bear much of that burden. In others, the cost may be spread through electricity rates, tax incentives, public investment or delayed infrastructure priorities.
The electricity bill for AI will not always appear on the AI invoice.
From convenience to dependency
The larger cost may be structural.
AI is already being considered or deployed in banking, identity verification, cybersecurity, fraud detection, public administration, healthcare operations, logistics and infrastructure management. In those settings, the risk is not limited to a chatbot producing a poor answer. The risk is that institutions redesign important processes around systems they do not fully control.
The Financial Stability Board has warned that AI use in finance can create vulnerabilities linked to third-party dependency, service-provider concentration, cyber risk, model risk, data quality and governance. These are not abstract concerns. Modern AI depends on a narrow stack of model providers, cloud platforms, chip suppliers, data pipelines and application layers.
Once AI is embedded into a critical workflow, the institution becomes dependent on that stack. Each layer can fail. Each layer can change its pricing. Each layer may sit outside the direct control of the organisation relying on it.
This dependency is manageable when AI supports an existing human process. It becomes more serious when AI replaces the process.
A bank that uses AI to help analysts review suspicious transactions still has a human workflow. A bank that restructures its compliance function around automated agents has a different risk profile. If the model changes, the vendor fails, the API goes down, regulation shifts, compute prices rise or power supply tightens, the bank may discover that the previous manual capability no longer exists.
The failure scenario is not only that AI breaks. It is that the fallback has disappeared.
Also Read: Why Malaysia’s AI Nation 2030 plan matters for B2B startups
The loss of institutional muscle
The labour debate around AI is often framed around jobs lost or productivity gained. That framing is too narrow. A more subtle risk is the erosion of competence.
Organisations learn through repetition. Junior staff become useful by drafting, checking, reconciling, researching, coding, correcting mistakes and handling exceptions. They build judgement because they spend time with the work.
If agents absorb too many of those tasks, companies may appear more efficient in the short term while weakening their own talent pipeline. The organisation keeps senior decision-makers and automated systems, but loses the layer of people who understand how the work is actually done.
This is difficult to measure. It would be careless to claim that AI has already hollowed out institutional competence across the economy. But the risk is credible, especially in fields where errors are detected through experience: finance, law, medicine, engineering, cybersecurity, aviation and public administration.
The danger is not that AI makes people incapable. The danger is that organisations stop creating people who know how to operate without it.
The cost of no return
The most expensive technology choices are not always the ones with the largest upfront price. They are the ones that remove optionality.
A company can test AI safely when it remains a layer on top of an existing process. The risk changes when the process itself is rebuilt. Teams are reduced. Vendor contracts are signed. Data flows are reorganised. Interfaces are redesigned. Compliance procedures are rewritten. Management dashboards begin to assume AI availability. New employees are trained on the AI-native workflow rather than the older one.
At that point, stepping back is no longer a matter of cancelling a subscription. The company would need to rebuild skills, processes, documentation, software and confidence.
This is one of the least visible costs of AI adoption. The bill is not only for compute. It is for irreversibility.
The counter case
A serious assessment must recognise that AI can also reduce costs. It can improve demand forecasting, detect fraud, support grid management, accelerate software development, strengthen customer service, optimise logistics and support scientific research. In some areas, AI may help reduce emissions or make infrastructure more efficient.
Nor are all AI systems equal. A small model running locally for a narrow task is not comparable to a frontier model powering an autonomous agent across multiple systems. A document summariser is not comparable to an always-on compliance, trading or cybersecurity agent.
The distinction matters. AI as a tool is one thing. AI as a dependency is another.
The answer is not to reject AI. That would be unrealistic and, in many cases, commercially damaging. The answer is to treat AI as infrastructure when it functions as infrastructure.
Also Read: The builders of AI are selling a future their own technology destroys
What should change
Companies need to measure AI by workflow cost, not subscription cost. A monthly plan tells management little about the economics of large-scale agent deployment. The relevant calculation includes compute, cloud services, monitoring, human review, compliance, security, error correction and fallback capacity.
Critical institutions also need continuity plans. Banks, public agencies, healthcare operators and infrastructure providers should be able to explain what happens if a model provider, API, cloud region, vendor contract or power supply becomes unavailable.
Regulators need better visibility into concentration risk. If many institutions rely on the same small group of model providers and cloud platforms, the exposure becomes systemic. The Financial Stability Board’s warnings on third-party dependency should be treated as an early signal, not a theoretical footnote.
Companies should also protect human competence. Keeping people “in the loop” cannot mean asking them to rubber-stamp machine output. It means preserving the ability to challenge, audit and replace AI-driven work. It also means continuing to train junior staff on fundamentals, even when automation appears faster.
Finally, AI infrastructure should face the same scrutiny as other strategic infrastructure. Data centres require power, water, land, transmission capacity and political permission. If the gains are private while the infrastructure costs are spread more widely, the public bargain will eventually be questioned.
The real price
The first phase of AI was about access. The next phase will be about dependency.
Cheap AI has encouraged experimentation, and much of that experimentation has value. But cheap access can hide expensive commitments. Once agents are embedded into the machinery of business and government, the cost of AI will no longer be measured only in tokens, subscriptions or data centre bills.
It will be measured in grid pressure, vendor concentration, institutional fragility, lost human capability and reduced room for reversal.
AI can produce useful work. That point is settled. The harder question is whether companies and governments are building around it with a clear view of what happens when it becomes more expensive, less available, less reliable or too deeply embedded to remove.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
Image credit: Tima Miroshnichenko
The post The true cost of AI is beginning to surface appeared first on e27.