Türkiye's AI Action Plan 2026-2030: What It Means for Developers and Companies

Announced by the President of Türkiye on August 2, 2026, the Türkiye AI Action Plan 2026-2030 sets the country's artificial intelligence roadmap on four axes: recognize, utilize, produce, govern. It is more than a vision paper — the plan carries measurable targets: AI training for 5 million citizens within two years, at least 2,000 public datasets opened to local developers, national data center capacity scaled up to 1 GW, and an expectation of more than 1 trillion TL in added economic value. So what does any of this mean in practice if you write code for a living, or run a company deciding where AI fits? This article reads the plan through a developer and business lens: what you can build with APIs and LLMs once public data opens up, why the 1 GW infrastructure target matters, and how a team that wants to start today can move — with KVKK, Türkiye's data protection framework, considered from day one.

What the Plan Says: Four Axes and Hard Targets

The plan was presented publicly by the President on August 2, 2026, with details published via iletisim.gov.tr and covered widely by national media the same day. The framework rests on four axes: recognize (public awareness and education), utilize (putting existing AI to work across the public and private sectors), produce (building local models, data, and infrastructure), and govern (safe, ethical, regulated use).

The targets underneath those axes are where the ambition shows:

• Education: AI training for 5 million citizens within two years.
• Open data: at least 2,000 public datasets opened to local developers.
• Infrastructure: scaling Türkiye's data center capacity to 1 GW.
• Talent: training 10,000 advanced AI specialists and 100,000 application professionals.
• Economic impact: an expectation of more than 1 trillion TL in added value from AI.

The visible face of the production axis is Bilge, the homegrown AI model developed by TÜBİTAK, Türkiye's national research agency. Notably, the plan does not treat the model as a standalone effort: it sits in the same frame as the open data that would feed it, the infrastructure that would run it, and the talent pipeline that would build on it.

When 2,000+ Public Datasets Open Up: What APIs and LLMs Unlock

For developers, the most concrete commitment in the plan is opening at least 2,000 public datasets to local developers. Open data is valuable on its own — but the real multiplier appears where that data meets large language models:

• Natural-language querying (RAG): index statistics, transport, agriculture, or meteorology data, put an LLM on top, and you get assistants that answer questions like "how much was it in which province last year?" with sources attached.
• Structuring messy records: converting PDFs, tables, and free-text records into clean JSON schemas with an LLM is the first step toward building products on open data.
• Citizen and municipal services: multilingual chat assistants that ground frequent requests in open data.
• Sector-specific decision support: planting plans in agriculture, route and demand analysis in logistics, consumption reporting in energy.
• Startup opportunity: the open data + LLM combination lays the groundwork for a Turkish wave of what the world calls "civic tech".

Turning a free-text open-data record into structured output, for instance, is a single request:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.onysoft.com/v1",
    api_key="sk-ony-YOUR_KEY",
)

response = client.chat.completions.create(
    model="google/gemini-3.6-flash",
    messages=[
        {"role": "system", "content": "You convert open-data records into clean JSON."},
        {"role": "user", "content": "Split this record into province, year, indicator, and value fields: [open data record]"}
    ],
)
print(response.choices[0].message.content)

JSON mode and function calling follow the same schema, so the output can go straight into your database. And because a single sk-ony- key reaches the entire 708+ model catalog, using a fast model for bulk structuring and a flagship for analysis and reporting is a one-line model-name change.

The 1 GW Data Center Target: Why Infrastructure Matters

Modern AI workloads — training and high-volume inference alike — run on GPU clusters, and the binding constraint on those clusters is usually not chips but power and cooling. That is why the target of scaling data center capacity to 1 GW is not "more server rooms"; it is a signal of scaled investment in domestic AI infrastructure.

Three practical consequences for developers and companies:

• Latency: as in-country inference capacity grows, real-time applications serving users in Türkiye get faster.
• Data residency: being able to process data domestically creates an option that simplifies data-transfer questions from a KVKK perspective.
• Local model capacity: training and serving homegrown models like Bilge at scale is only possible with infrastructure in this class.

The key point: infrastructure targets play out over years, but there is no reason to wait before building applications. You can ship your product today on world-class models over an API, then widen your options as the domestic infrastructure and model ecosystem matures. Starting with an API layer that does not lock your architecture to a single provider makes those later transitions painless.

A "Start Now" Playbook for Companies

The plan's targets of 100,000 application professionals and AI training for 5 million citizens mean one thing for the market: over the coming years, your customers and your competitors will both be more AI-literate. The cost of waiting is rising while the cost of starting stays low. Here is a sequence that works:

1. Pick a low-risk internal use case: document summarization, support-reply drafting, extracting action items from meeting notes, internal knowledge-base search. Scenarios that never touch a customer are the safest place to learn.
2. Create a free account: registration on Onysoft AI Gateway is free and billing is pay-as-you-go in Turkish lira — no foreign credit card, VPN, or overseas contract needed, and compliant corporate e-invoicing is standard.
3. Compare models in the Playground: running the same prompt against different models shows, without writing code, which class of model each job actually needs.
4. Run a small but measurable pilot: one team, one scenario, clear criteria — response quality, latency, cost per request.
5. Write your KVKK policy alongside the pilot: which data may be entered, how outputs are verified, who has access (details in the next section).
6. Scale and train the team: roll the pattern that worked in the pilot out to other teams; the plan's application-professional target is exactly this kind of in-house capability investment.

The first request takes about two minutes:

curl https://api.onysoft.com/v1/chat/completions \
  -H "Authorization: Bearer sk-ony-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-3.6-flash",
    "messages": [{"role": "user", "content": "Extract the action items from these meeting notes."}]
  }'

Because the endpoint is OpenAI-compatible, your existing SDKs work as-is; one key, one balance, and one invoice cover the whole catalog, with 24/7 support when you get stuck.

The KVKK Angle: Three Guides, One Compliance Roadmap

The plan's govern axis signals that the regulatory side is not standing still — and in Türkiye the most concrete output so far has come from KVKK, the Personal Data Protection Authority. Across 2025-2026 it published three AI guides in quick succession:

• November 2025: a guide covering generative AI in 15 questions — core concepts, risks, and data responsibility.
• March 2026: a guide on agentic AI — oversight and accountability for systems that can act autonomously.
• June 2026: a guide on generative AI in the workplace — rules and policy recommendations for everyday employee use.

The shared message of all three is clear: using AI is fine, but sharing personal data as input requires a legal basis, transparency, and data minimization; outputs should not enter decision processes unverified; and there should be a written internal usage policy. Build your pilot on these principles and no compliance debt accumulates by the time you scale — which also keeps you aligned with the action plan's governance emphasis.

For a practical checklist and sample policy clauses, see our KVKK-compliant AI guide (in Turkish); to compare provider options under Türkiye's conditions, the AI API providers in Türkiye article is a good starting point.