A needs analysis approach for artificial intelligence infrastructure operated in Türkiye

Purpose and scope of the study

Infrastructure decisions for artificial intelligence applications are often discussed in terms of hardware capacity. Yet the capacity required depends on how many users will use the infrastructure, for which tasks, and with what response time and quality expectations. An investment made without defining the need can result either in idle capacity or in an infrastructure that falls short during the service's first period of growth.

This study sets out how a needs analysis for an artificial intelligence and data infrastructure operated in Türkiye should be structured. Its scope covers the components for running models, storing application data, logs and backups that will serve education, research and SME applications. The study does not offer a capacity proposal or a cost calculation; it defines the questions, information sources and evaluation headings on which these decisions should be based.

Operating the infrastructure in Türkiye is treated in this study as a design requirement. Hosting model execution, application data, logs and backups in Türkiye makes it easier to define responsibilities for data security, access control and service continuity clearly. This requirement is taken into account at every stage of the needs analysis as a criterion that narrows the options and guides the assessment.

Defining use cases

The basic unit of the needs analysis is the use case. A use case describes how a particular user group will carry out a particular task with the support of artificial intelligence: for example, a teacher using an assessment assistant, a researcher screening a large set of documents, or a business classifying incoming customer requests. For each use case, the following information is recorded in the same format:

• User group and expected range of user numbers
• Type of task: text generation, summarisation, classification, search or document analysis
• Intensity of use: daily number of requests, peak hours and concurrent use
• Acceptable response time and quality expectations
• Nature and confidentiality level of the data to be processed
• Criticality of the use case in terms of service continuity

Use cases are defined together with the teams in the relevant programmes. Applications developed in the digital education, free software services and digital transformation programmes are the first users of the infrastructure; demand information therefore comes directly from the design of these programmes. Use case records are also used for prioritisation: use cases with a high number of users, clear data confidentiality requirements and critical importance for service continuity are considered among the first to be included in the pilot.

Evaluation headings

Once the use cases have been defined, the need is assessed under five interrelated headings. These headings must be considered together rather than separately. For example, choosing a larger model may improve response quality, but it has a direct impact on hardware, energy and latency.

• Model suitability: Turkish and multilingual support, context length and task performance required by the use cases; licence terms of models suitable for local deployment
• Computing capacity: The processors, memory and storage required for concurrent demand, model size and response time targets
• Data management: Keeping application data, logs and backups in Türkiye; access rights, retention periods and deletion rules
• Service continuity: Monitoring, updates, restoration from backup in the event of failure, and responsibility for maintenance
• Operating costs: Linking hardware renewal, energy, hosting, licence, maintenance and specialist staff costs to the cost per transaction

Of these headings, the one most easily overlooked at the decision stage is operating costs. Set-up costs are one-off and visible; energy, hosting, maintenance, update and specialist staff costs recur in every period and can make up the decisive share of total costs over the life of the infrastructure. The assessment therefore considers set-up and operating costs together over the service life.

Information sources and method

The needs analysis draws on three types of information. The first is use case information obtained from programme teams and user institutions. The second is technical measurements obtained by testing candidate models on defined tasks. The third is financial information on hardware, hosting and operating costs. These three types of information are combined so that demand, capacity and cost can be seen in the same table for each use case.

To reflect uncertainty, demand estimates are expressed with lower and upper bounds rather than as a single value. Capacity needs are calculated separately for different points within this range. Decision-makers can thus see both the capacity that would sit idle at low demand and the bottleneck that would arise at high demand. In technical measurements, each model is tested with the same set of tasks under the same conditions; results are recorded together with information on the hardware and settings used for testing.

Alongside technical and financial information, human resource needs are assessed separately. The areas of expertise and the workload required to set up, monitor and update the infrastructure and to support users are defined. If the continuity of the specialist team cannot be ensured, even a technically adequate infrastructure cannot maintain service quality.

Pilot evaluation and decision criteria

The needs analysis is validated with a limited pilot before full-scale investment. The pilot involves operating selected use cases with real users over a defined period. The indicators monitored in the pilot should be the same as the criteria used at the decision stage:

• Response quality and latency in defined use cases
• Resource use and operating cost per transaction
• Service availability and successful restoration from backup
• Whether users continue to use the service in their daily work

The duration of the pilot is set to reflect the usage cycle of the use cases. Busy periods, such as the beginning and end of terms for education applications and month-end for business applications, are when capacity is put to the test; covering at least one of these periods ensures that the pilot's measurements are realistic.

Pilot findings make it possible to increase capacity on the basis of measured demand rather than assumptions. Capacity decisions are assessed not only on technical findings but also in the light of the sustainability of operating resources.

Relationship to the programmes

This approach offers a working framework for the demand and technical design work package of the Türkiye-Based Artificial Intelligence and Data Infrastructure Programme. Teams developing applications in the digital education, free software services and digital transformation programmes are the main contributors to defining use cases. The participation of researchers, technical experts and institutions able to provide infrastructure capacity is considered under the coordination of the Digital Transformation and Artificial Intelligence Directorate.

The main limitation of the approach is that demand estimates initially rely largely on the expectations of use case owners. The needs analysis is therefore treated not as a one-off exercise but as an assessment that is updated with findings from the pilot and the first period of service.