GPU for Everyone: Türkiye will open compute to SMEs and startups
Per Anadolu Agency on August 31, 2026, a program called GPU for Everyone will open access to AI compute for researchers, startups and SMEs. At least 2 million GPU hours a year will be allocated as credits in the first phase, and the program's first call will open within 6 months.
Compute is tied to a national program
On August 31, 2026 Anadolu Agency reported that the compute behind AI will be opened to researchers, startups and SMEs through the GPU for Everyone program. The report rests on information obtained by its correspondent rather than a published ministry announcement.
The program will be implemented as the National AI Research Resource Program and will run under the 2026-2030 Türkiye AI Action Plan.
Coordinated by the Ministry of Industry and Technology, the program will be carried out by the Ministry together with the Cybersecurity Directorate.
The action plan itself took effect on August 18, 2026, and the figures of 2 million a year and 20 million by the end of 2028 were already public then. What the August 31 report adds is the number of data centers, the quota share, the call timeline and the consolidation of infrastructure under one roof.
Five accredited data centers and at least 2 million GPU hours a year in the first phase
Per the report, capacity contracts will be signed with at least 5 accredited domestic data centers in the first phase.
From that capacity, at least 2 million GPU hours a year will be allocated as credits to researchers, startups and SMEs. At least 40 percent of the capacity will be reserved for tech startups.
Separate access windows and allocation quotas will be defined for startups, SMEs, university researchers and public projects.
The annual total allocation is targeted to rise to at least 20 million GPU hours by the end of 2028. The program's first call will open within 6 months.
The credits will not be uniform. Different credit packages will be prepared for training and inference, and flexible access models including unit usage credits will be offered to users who demonstrate competence in AI development.
GPU hour credits will be provided to eligible startups and SMEs as packages alongside software, cloud and hardware support. The packages will also include technical consultancy and mentoring.
The service catalog will not be limited to raw compute capacity, and will also cover model hosting and inference services from accredited providers.
The program will bring together, under one roof, capacity from accredited domestic data centers, the national high performance computing infrastructure including TRUBA at TÜBİTAK ULAKBİM, and the National Computing Portfolio.
Data center accreditation will look for energy efficiency, data security, service continuity and protection of users' intellectual property rights over their model, data and code assets.
Per Webrazzi's reporting on the action plan, the targets include allocating at least 1,000 AI vouchers to SMEs in the first 12 months and raising installed data center capacity to at least 1 GW by 2030.
Compute turns into a piece of infrastructure
The assessment in this section is ours. In an AI project the most invisible and most expensive line is usually compute. The idea is free, the experiment is not.
A subsidized GPU hour lowers the cost of trying. That touches learning speed directly, because it decides how many rounds you can afford before you find out which idea works.
The quota architecture is notable. Reserving at least 40 percent for tech startups and defining separate access windows for different groups is a design meant to stop one segment from consuming the whole resource.
It also matters that accreditation criteria include protecting users' intellectual property rights over their model, data and code assets. That is usually the first question a business asks about shared infrastructure: what do I lose by putting my work on someone else's machine.
The presence of technical consultancy and mentoring inside the package is meaningful too. On the ground the real bottleneck is usually not hardware, it is the shortage of people who know what to do with it.
One reminder is due: all of this is still in the future tense. Contracts are not signed, quotas are not set, the call is not open. The calendar rewards whoever is prepared.
What it means for businesses in Türkiye
The assessment below does not appear in the source, it is our reading. This story concerns businesses in Türkiye directly.
Getting a credit requires an application and a project. What you describe in that application is not a model, it is a business problem: which process, with which data, improved toward which measurable result.
That is why the real preparation is on the data side. A business with scattered spreadsheets, processes nobody records and rules that live only in one person's head stays exactly where it is, GPU hours or not.
The most concrete preparation is this: pick one process, start collecting the data it produces in an orderly way today, and write the success criterion in a single sentence.
Because credits will go to eligible startups and SMEs as packages, the eligibility conditions will only become clear with the call text. Time will be tight once the call is published, so a business that prepares now will be comfortable at that moment.
A GPU hour credit is a tool, not a goal. The tool being accessible does not remove the responsibility of knowing which job you are doing with it.
The UNALSOFT take
On the agentic AI side the first question we ask is not which model will be used, it is which job will be automated. In a company an AI project usually stalls because of an undefined process and scattered data, not because of missing compute. To the extent this program lowers the cost barrier, it will make that reality even more visible: when the resource is shared, the difference shows up in preparation. Our advice is to use the waiting period to pick a single process, put its data in order and tie success to one measurable criterion. Sitting down with a ready file when the call opens is far cheaper than working out what to do after the capacity arrives.
Compute is coming, is your data ready?
Let us work out together which process suits AI and which data you should start collecting today.