Uzbekistan defines subsidised AI infrastructure access for socially significant projects

New rules set separate token and GPU-memory allowances, competition windows and appeal procedures for eligible institutions and startups.

Uzbekistan has set out a programme for free or subsidised access to government AI infrastructure for socially significant projects. The Ministry of Digital Technologies published the [rules summary on 21 October](https://gov.uz/en/digital/news/view/235892), following Government Resolution No. 581 of 20 October 2026.

The programme covers eligible higher-education institutions, scientific organisations, startups and specified regional digital-technology advisers. Large-language-model projects can receive API access for up to six months, with ten million tokens per calendar month. Additional usage of up to fifty million tokens receives a fifty-percent tariff discount; unused allowances do not roll over.

Two routes reflect different computing needs

The ministry describes a separate allowance for non-LLM model training, testing, validation, evaluation and inference: up to six months of access with no more than 120 GB of video memory. Selection competitions are scheduled for July and November, using a hundred-point assessment scale.

Separating the routes makes the programme more specific than a general promise of AI support. An application that calls a hosted language model needs a measurable request allowance. A project training or serving a different model needs access to accelerator memory and associated computing resources.

Those are different project budgets. A token allowance measures the amount of text processing available through an API; accelerator memory determines which workloads may fit within a computing allocation. Neither figure alone describes an entire project's capacity, but each gives applicants a concrete planning input.

A project proposal becomes a resource plan

For a university or startup, the useful preparation is to connect the intended public benefit with an actual workload. A document-assistance project might estimate the number and length of documents it expects to process. A computer-vision project might describe its model, dataset and training or evaluation stages.

That planning helps answer a basic question: what can the team complete during the access period? A finite allocation can support a defined prototype, a validation study or a pilot. Setting the milestone alongside the resource request makes it easier to distinguish research progress from an indefinitely running service.

The programme also limits users to one simultaneous project. That encourages an identifiable unit of work with a responsible applicant, rather than an open-ended collection of unrelated experiments under one allocation.

Selection and appeals are part of access

The published summary includes an electronic appeal route, with three working days to challenge a result and five working days for review. It also sets restrictions on use, disclosure and commercial processing, subject to applicable legislation or contractual provisions.

For participants, those details make the administrative process as important as the technical allowance. An institution needs to know who applies, which project is being assessed and how the eventual allocation relates to its intended activity.

The announcement adds a structured compute-access mechanism to Uzbekistan's AI development programme. Its immediate value is specificity: eligible organisations, distinct technical allowances, scheduled selection and a route to question the result. Applicants can use those boundaries to prepare a focused project rather than infer eligibility from a general national AI ambition.

Sources

- [Free and Preferential Access to AI Infrastructure Established for Socially Significant Projects](https://gov.uz/en/digital/news/view/235892) — Uzbekistan Ministry of Digital Technologies