According to optimistic predictions, artificial intelligence was supposed to increase productivity, accelerate research, and ultimately reduce costs. Darek Šmíd’s article “AI Was Supposed to Make the World Cheaper. Instead, It Is Making Life More Expensive,” was published on WIRED.cz highlighted the other side of the current AI boom: the heavy losses reported by technology companies, rising energy consumption, the difficulty of measuring return on investment, and the true cost of using AI. We decided to examine the topic from the perspective of an academic institution that operates and uses artificial intelligence.
Lukáš Hejtmánek from CERIT-SC at Masaryk University explains why it is necessary to distinguish between the cost of developing models and the cost of using them daily, how AI consumption can be tracked token by token, and what the technology already offers researchers and students.
The largest companies developing artificial intelligence are reporting heavy losses. Does this mean that the current development of AI is economically unsustainable?
It is true that many companies developing large models are operating at substantial losses. But a lack of profitability alone does not tell us whether a technology is useful or viable in the long term.
We can compare it, for example, to railway infrastructure. Railway operations are subsidized in many countries and may not be profitable in themselves. Yet we do not conclude from this that railways or public transport are useless. In the case of infrastructure, profit does not have to be the only measure of value. What also matters is the service it provides to society.
Some of today’s AI companies or start-ups may disappear. But that does not mean the models they have already developed will disappear. At CERIT-SC, we operate open-weight models. We store them on our own infrastructure, so they do not depend on whether a particular provider continues to give us access to its service.
Could today’s prices for AI services be artificially low?
Some commercial providers may increase their prices over time. When assessing these prices, however, it is important to separate the cost of developing and training a model from the cost of its subsequentday-to-day operation, or inference. Conflating these two items can create a distorted impression of how much the actual use of AI really costs.
Training a state-of-the-art model can cost enormous sums. This is where a significant share of the losses discussed in the media arises. As with other forms of expensive infrastructure, however, it may not be decisive whether the initial investment pays for itself directly. The broader benefit that the technology subsequently brings to society may also be important. The daily operation of an already-trained model is a distinct economic discipline, and we can track it relatively precisely at CERIT-SC.
From 1 January to 14 July 2026, our infrastructure processed approximately 31.3 million requests and 315 billion tokens. Based on comparable prices charged by commercial providers, this level of consumption would correspond to roughly USD 222,000. One request, therefore, costs approximately USD 0.007 on average.
For each of our 2,392 registered users, this represents approximately USD 14 per month. That is less than a standard individual subscription for some commercial chatbots, even though the total consumption also includes demanding agent-based tasks that involve tens of thousands of tokens in a single request.
The total cost naturally also includes hardware, electricity, cooling, facilities, and staff. A single NVIDIA DGX system costs more than half a million dollars. At our current utilization rate, however, our own inference infrastructure could pay for itself in approximately two years when compared with reference commercial prices.
Can the cost of using AI easily get out of control?
It can, if an organization does not monitor consumption. But this is not unique to artificial intelligence. The same risk exists with all cloud services.
In our environment, we continuously monitor token consumption. When a user or application consumes an unusually high volume over a short period, it is usually caused by a programming error that repeatedly sends requests, or by a leaked access key, for example. In such cases, we contact the user and address the problem.
A high bill, therefore, does not necessarily mean that an individual token is extremely expensive. Often, limits, alerts, and operational oversight are simply missing. It is similar to a poorly configured cloud server running without restrictions. The problem is not necessarily the price of the technology, but the fact that no one is monitoring how it is used.
Companies often do not know how to measure the return on investment in AI. How can its benefits be assessed?
Return on investment is difficult to measure when AI is discussed as a universal tool expected to transform an entire organization in an undefined way. Once it is deployed for a specific task, however, its benefits can be evaluated much more precisely.
At CERIT-SC, for example, we use an autonomous agent called AI Karel. Every day, it scans the computing cluster and prepares an operational report. It identifies, among other things, graphics processing units that have been reserved but remain unused for several days, or services that remain in the infrastructure without an active user.
In one report, for example, it detected 12 reserved but unused graphics processing units and more than 500 objects that were no longer in use. Because operating this type of hardware is expensive, identifying such cases early can save high costs and staff time. Similar functionality can be provided by monitoring tools, but in our case, the system combines monitoring with an analysis of the underlying cause. The latter is much more difficult to achieve through conventional monitoring alone.
For a clearly defined task, we can therefore compare how much time and money it required before, how much it costs to solve it using AI, and what result it delivers. That is when return on investment becomes measurable.
The energy consumption of data centers is one of the most serious arguments against the expansion of AI. How do you view this issue?
In my opinion, it is the most justified of the concerns being discussed. Data centers consume large amounts of electricity and require cooling, space, and additional infrastructure. It would be a mistake to downplay this problem.
At the same time, it is necessary to distinguish between different ways of operating AI. It does not always have to involve a large global service available online. Part of the future may lie in locally operated models shared by specific institutions or communities.
At CERIT-SC, such an environment serves thousands of users. The models run within academic infrastructure, prompts and responses do not need to leave it, and the content of the communication is not used for further training. We record only the operational metadata that is strictly necessary. This is particularly important when working with sensitive or regulated research data. For some projects, local academic infrastructure is not merely a cheaper or safer alternative to a commercial service. It may be the only legally and ethically acceptable option.
Local operation does not, of course, solve the issue of energy consumption on its own. It does, however, make it possible to share infrastructure more efficiently, monitor its use, and reduce instances ofexpensive hardware remaining idle.
How is AI already helping students and early-career researchers?
During periods when final theses are being submitted, we regularly see a marked increase in our consumption graph. Higher consumption alone does not, of course, prove that AI is genuinely helping students. But from the way the models are actually used, we can see that they are being applied to tasks such as data analysis, programming, work in Jupyter notebooks, and identifying calculation errors — activities in which they can directly simplify or accelerate the work.
Through the academic infrastructure, a student in Czechia can gain free access to high-performance models without their research data leaving the institution. An assistant can work directly with a notebook, code, or a data file and, for example, explain why a particular calculation failed.
The benefit need not be an abstract promise of a technological revolution. It can be a student who completes a thesis more quickly, identifies an error, or manages to use a method that would otherwise have been difficult to access.
Do we already have concrete evidence that AI is advancing scientific research itself?
The results of AI in science are no longer merely hypothetical. AlphaFold fundamentally transformed the prediction of three-dimensional protein structures, and its creators were recognized with the 2024 Nobel Prize in Chemistry. The AlphaFold database contains predicted structures for approximately 200 million proteins.
Another example is rentosertib, a drug being developed for idiopathic pulmonary fibrosis whose design involved the use of generative AI. The results of a clinical trial were published in Nature Medicine. It is not yet an approved medicine, but it is a concrete example of how AI can accelerate certain stages of drug development.
In materials research, the GNoME system proposed 2.2 million potentially stable crystal structures, of which approximately 381,000 were assessed as new stable materials suitable for further investigation.
We are also seeing progress in mathematics. In 2026, for example, an OpenAI model autonomously disproved a long-standing conjecture in discrete geometry concerning the unit-distance problem. Externalmathematicians subsequently verified the proof. This is a concrete example of AI gradually moving beyond solving known tasks and towards working on genuinely open research problems.
It is therefore right to ask how much AI costs and who will pay for its development. But it is equally important to examine what we are actually getting in return.
Are concerns about the current expansion of AI, therefore, exaggerated?
No. Some of the skeptical predictions may come true. Companies will probably disappear, service prices may change, and energy consumption will require serious debate. Only time will show to what extent the current situation is an investment bubble and to what extent it is the early phase of a long-term technological transformation.
I disagree with the idea that nobody knows the true cost of AI or that it is impossible to determine what users receive in return. We track that cost at the level of individual tokens and requests. At the same time, we can point to specific services, research projects, and operational savings that AI is already making possible.
Above all, AI is changing the way we work. This is captured well by a frequently quoted statement by Michaela Liegertová: “AI will not replace you, but you may be replaced by someone who knows how to use it.”
About the author
RNDr. Lukáš Hejtmánek, Ph.D., heads the cloud computing and storage unit at CERIT-SC. His work focuses on the operation of computing infrastructure, cloud services, and artificial intelligence tools for the academic and research community. CERIT-SC is part of the Institute of Computer Science at Masaryk University and the national research infrastructure e-INFRA CZ.
---
The illustrative image was created using generative AI.
