Yesterday, the Ministry of Digital Affairs announced the launch of the AI Factory project. This is an IT environment, the basis of which will be the first state AI services: an AI assistant in «Diya» and later — an AI tutor in «Mriya». And in the future — services for the defense industry, science and medicine.
Yesterday, the Ministry of Digital Affairs announced the launch of the AI Factory project. This is an IT environment, the basis of which will be the first state AI services: an AI assistant in «Diya» and later — an AI tutor in «Mriya». And in the future — services for the defense industry, science and medicine.
In order to create such an environment for AI chatbots that will process user requests in real time, you will have to purchase hardware and software. As the editorial staff of dev.ua reported, a corresponding announcement has already appeared on Prozorro. The purchase will be carried out by the State Enterprise «Diya».
We are talking about two sets of «Software and Hardware Complex «Sovereign Factory of Artificial Intelligence». Procurement budget (expected cost): UAH 227.5 million. Delivery must take place by June 1, 2026. The contractor must have an income of UAH 100 million over the past year.
According to participants in the data storage and processing market, with whom the editorial office spoke, AI complexes must be ordered in advance. Due to high demand, there is a very small probability that deliveries will take place by the end of the year. Therefore, apparently, the delivery date is so far removed from the start of the procurement.
So the AI infrastructure at «Diya» will most likely not appear as soon as we would like.
The specifications for what the complex should be, its hardware and software characteristics, are spelled out in great detail in the procurement documents. For example, GPU computing nodes (there will be 4 of them in total on two complexes) must be equipped with at least two next-generation processors on the x86_64 architecture with PCIe Gen 5 support. Each of them has at least 96 physical cores with a base frequency of at least 2.6 GHz.
Memory: The compute node must contain at least 24 modules with a capacity of at least 96 GB each. There must be the ability to scale to at least 6 TB of system memory.
The solution must provide memory operation at a speed of at least 6400 MT/s.
It should be noted that the Ministry of Digital Affairs told DOU that AI Factory will initially include four GPU nodes. Over time, their number may increase to 31.
After installing the complexes in the data center, the contractor must conduct practical tests. This is the retraining of the large language model Gemma-3-12B. For this purpose, the customer provides a text dataset of up to 10 GB.
The criterion for success is the measured time of the full pre-training cycle (Time-to-Train). The system must demonstrate the ability to handle a load equivalent to 310,000 requests per day. In 95 percent of cases, the response time (latency) should not exceed 500 milliseconds.
We will remind you that we wrote a detailed article about the fact that the Ministry of Digital Affairs and the country’s largest telecom operator have agreed to create a national LLM (large language model). And Kyivstar is ready to invest $1.5–2 million in initial investments.
As Mykhailo Nestor, Director of Digital Product Development at Kyivstar, explained to dev.ua, in the case of an IT environment that wants to independently build an AI Factory, it is about purchasing infrastructure for the load that will come, for example, from chatbots, and not from training models (as in the framework of the national LLM project). Because Kyivstar is responsible for the infrastructure for training the national LLM.
What infrastructure will Kyivstar purchase for LLM training: the company is considering a cluster of 64 Nvidia GPUs — H100/H200 in pods that include high-speed switching such as NVLink, etc. One of the cloud market participants told dev.ua that he purchased similar equipment last year. The cost of one GPU of this level was about $30,000. Thus, we can talk about approximately $2 million. This is similar to these figures, which the operator announced at the start of the project.
«For us, this can be both local hardware and a cloud cluster on one of the hyperscalers (AWS, Azure, GCP),» emphasized Mykhailo Nestor.
The final result — a trained language model from Kyivstar — can then be hosted within the AI Factory for the needs of chatbots «Diya», «Mriya», etc. But each request will also require processing on the GPU (even if there is an untrained model at hand). Therefore, according to Nestor, local servers (which are currently purchased by «Diya») are an advantage due to lower signal latency and cost.
«The ministry can also conduct its own experiments with model training for its own needs,» adds Nestor.
