Applying for GPU Credits
The state will hand out computing power as credit. Projects get credit; curiosity does not. The GPU-for-All programme will give startups, SMEs and researchers GPU-hour credits. Two million GPU-hours a year in the first phase, rising to twenty million annually by the end of 2028. At least 40 percent of credits go to tech startups. Separate access windows are planned for startups, SMEs, university researchers and public projects.
You need to be ready when the call opens. This piece explains what ready means.
What Is the Problem?
BU BÖLÜMÜN ÖZETİ
- The application waits for the last week
- The project is undefined
- The data is not ready
Three things happen again and again in support calls.
The application waits for the last week
The call is announced, “we’ll look into it” is said, the final week arrives. A hurried application is both weak and incomplete. The evaluator sees it in the first paragraph.
The project is undefined
“We want to use AI” is not a project. If you have not written which data, which model and which decision you intend to automate, there is nothing to apply with.
The data is not ready
Even if the credit arrives, it cannot be used. Scattered, unlabelled data without clear permissions makes computing power useless. A strong engine goes nowhere with an empty tank.
Why Does It Happen?
BU BÖLÜMÜN ÖZETİ
- Compute needs were never estimated
- Training and inference get confused
- Credit alone is assumed to be enough
Three reasons.
Compute needs were never estimated
Most businesses do not know how much computing power they need. Not knowing, they cannot write what they want. Yet a rough estimate is possible: which job, how often, with how much data.
Training and inference get confused
These are two different workloads. Training means building or reshaping a model with your own data. Inference means running a ready model. They need different resources and are priced differently. The programme’s credit packages differentiate accordingly. Most businesses need the inference side.
Credit alone is assumed to be enough
GPU-hours are one part of the job. Data preparation, model selection and evaluation sit alongside. The programme accounts for this: packages include technical consulting and mentoring, and for eligible startups and SMEs credits come bundled with software, cloud and hardware support. What is offered is a working setup, not raw power alone.
How Is It Done?
BU BÖLÜMÜN ÖZETİ
- Step 1: write a one-sentence project definition
- Step 2: estimate your compute need roughly
- Step 3: sort the data side before applying
Preparation completes in three steps.
Step 1: write a one-sentence project definition
The whole application rests on this sentence. Fill in the template: “Using [data] with [model or tool], we will improve [decision or process] by [measurable target].” If you cannot build that sentence, it is not application time yet. When you can, half the application is written. The other half is evidence for each part: where the data sits, why that model, how the target gets measured.
Step 2: estimate your compute need roughly
Three questions suffice. Will you train a model or run a ready one? What is your monthly processing volume? Will this run continuously or seasonally? A rough estimate beats an empty field. You can get outside help to sharpen the numbers, but make the first estimate yourself.
Step 3: sort the data side before applying
Where the data sits, in whose name it is held, under which permissions it was collected — all written down. If personal data is involved, settle anonymisation or consent status. This table will be asked for at application and needed on day one after the credit lands. The data inventory is the foundation of this work.
How Long, Where to Start?
BU BÖLÜMÜN ÖZETİ
- The file skeleton takes a day
- The real cost is not money but queue position
- First step: write the sentence and the table
Preparation is cheap; timing is critical.
The file skeleton takes a day
Project sentence, compute estimate, data table and a short company profile. All of it comes together in a day. Updating it when a call opens takes a few hours. Whoever has the file applies; whoever is writing it does not make it.
The real cost is not money but queue position
GPU-hours have a market price, and you can value your allocation against it. But the real issue is the quota. The programme will run through capacity agreements with accredited domestic data centres, with at least five in the first phase. Early applicants face fewer competitors and more flexibility. Latecomers pay in waiting.
First step: write the sentence and the table
Write the project sentence; it takes fifteen minutes. If you cannot, what is missing is not the application but the project. Then produce the data table: which data, where, in whose name, under what permission. Third, assign call-watching to someone. The monitoring routine covers that habit.
The Common Mistake
BU BÖLÜMÜN ÖZETİ
- Talking big, planning small
- Leaving the startup certificate to later
- Asking for credit to experiment
Three traps at the application desk.
Talking big, planning small
Writing an ambitious target and leaving the substance thin is the most common error. Evaluation looks at coherence, not ambition. A small clear project beats a large vague one.
Leaving the startup certificate to later
Startups joining AI-related public procurement are expected to hold a tech-startup certificate. The process takes time and does not wait for calls. If you are a startup, start it today.
Asking for credit to experiment
“Let’s take it and try something” weakens the application and wastes the resource. Experiments are run on your own budget. Credit is for scaling work whose experiment is already finished. An application written without a small trial collapses at the first question: how do you know this works?
Frequently Asked Questions
Sık Sorulan Sorular
Businesses that never train a model still consume compute. Image processing, speech analysis, search across large document sets, demand forecasting. All of these run on ready models but need computing power. The programme’s packages also cover model hosting and inference services, not just raw GPU-hours. So you can benefit without ever saying “we will build our own servers”.
No. The credit programme provides computing resources; it does not ask for data. Accreditation criteria explicitly include protecting users’ intellectual property rights over their models, data and code. Still, read the agreement and keep your own record of which data is processed where. The rule is the same everywhere: never put data somewhere whose terms you have not read.
The file remains. The project sentence, data table and compute estimate stay valid for the next call. The same file also works at other doors: SME vouchers, sector pilot calls, excellence centres. A rejection is the result of that call, not of the file.