Note. This piece doesn't describe a funding scheme or a procedure: the Plan IA360, presented on 21 September 2026, still has no call or governing rules, and its dates are subject to regulatory development. What you can decide today, with or without a public IA360 business voucher (bono empresarial de IA360), is what model and what infrastructure your own project needs.
The Plan IA360's biggest investment project is an artificial intelligence gigafactory: €5,000 million in public-private investment, of which €719 million is public money, with the goal of «construir la mayor infraestructura de cómputo del sur de Europa» (building the largest computing infrastructure in southern Europe) and 100,000 AI accelerators operational in 2028/2029 (Plan IA360, flagship project 1). The plan earmarks it for the «entrenamiento y despliegue de modelos y aplicaciones avanzadas de IA» (training and deployment of advanced AI models and applications) at country scale. No business weighing up its first AI project needs anything like that, and confusing the scale of a country with the scale of a specific project is the fastest way to oversize — and make more expensive — something that could be solved with a great deal less.
The question that actually determines your project isn't how much compute exists in Spain. There are two: what model your task needs, and where it's going to run.
The first question: what size of model your task needs
Not every task needs the same model, and using the most powerful one available for everything is as inefficient as using the smallest one for something that requires reasoning across several steps. What determines the real size you need is a combination of factors: the type of task — classifying text isn't the same as drafting code — how much text goes into each query, and the consequence of the model getting it wrong. A support ticket classifier with human review before it matters can be solved with a compact, cheap model; a system that drafts the first version of a contract, whose error nobody reviews before it's sent, calls for a different category of model, even if the volume of work is smaller.
Cross-referencing those variables — task type, input length, consequence of an error, target latency and budget — to decide which model profile fits — compact, general-purpose or reasoning — is exactly the kind of decision a maturity diagnosis and roadmap resolves: a starting point for the conversation, not a price list or a pick of a specific commercial AI model.
The second question: where it runs
The second question sets aside model size and moves to location: whether your data can leave your infrastructure without a problem, whether it has to stay inside the European Union, or whether it can't leave your own network under any circumstances. That answer changes what the GDPR requires of you — if the data leaves the European Economic Area, the international transfer rules come into play — and determines what you have to ask any provider for in writing before signing anything: where it processes the data, how long it retains it, whether it uses it to retrain its own model. This question is covered in more detail, with concrete examples, in where your data lives when you use AI.
An example of why size is a question of task, not budget
Think of two projects at the same business. The first classifies the subject of each incoming email to route it to the right department: the input is short, an error is fixed by forwarding the email, and nobody loses anything serious if it gets it wrong once in twenty. The second drafts the reply a salesperson is going to send a customer about a contract's terms: the input can be an entire case file, and an error there goes straight to the customer. The first calls for a compact, cheap model with minimal latency. The second calls for a reasoning profile, even though it handles far less volume. Confusing the two — or, worse, applying the same model to both for convenience — is a common way to overspend without the project improving for it.
How much it costs, with no invented figures
Once the model and where it runs are decided, the next question is how much it's going to cost to build. When the project is a system that answers questions about your own documents — a RAG — the cost depends on your own corpus: how many documents, how much they change per month, how many fragments they're chunked into for indexing, and what rate your provider charges you. That calculation doesn't come from a generic table: it's worked out with your own volumes and your own rates, in the inventory phase a RAG and internal search project starts with, before deciding the architecture.
When you really do need more power
There are projects that fall outside this framework: training your own model from scratch, batch-processing data volumes that don't fit in an ordinary server's memory, or working with information that demands minimal latency at massive scale. These are a minority, and they're almost always spotted at the diagnosis stage, before committing budget: if your case is one of them, you'll normally already suspect it before reading this. For everyone else — the majority of AI projects an SME starts — the question isn't how much compute the country has, but what model and what deployment the specific task in front of you calls for.
Frequently asked questions
Do I need my own GPU to run an AI project at my business?
In most cases, no. Most SME projects call an already-trained model through an API or deploy it on existing infrastructure; training a model from scratch is the exception, not the usual starting point.
What's the difference between a compact model and a reasoning model?
A compact model handles narrow tasks with low latency and cost, suited to cases where volume is high and errors get human review. A reasoning model sustains longer tasks or ones with more intermediate steps, and is justified when the consequence of an error is high or nobody reviews the output before it takes effect.
How much does it cost to implement a RAG system at my business?
It depends directly on your document volume, how much it changes per month and your provider's rate: there's no single figure valid for every business. That calculation — set-up investment and monthly fee — is exactly what's worked out in the inventory phase of a RAG project, described above, with your own data, not a generic table.
Does the Plan IA360's gigafactory give me access to more compute capacity?
Not for now. The plan states it «facilitará el acceso a capacidad de cómputo a centros de investigación, empresas y administraciones» (will make it easier for research centres, businesses and public administrations to access computing capacity), but it doesn't yet set out the conditions or the procedure, and its opening is planned for 2028. The route designed for an SME to reach the point of using AI without building its own infrastructure is the Red NEURONA.
Further reading
The gigafactory, once it exists, will solve a country's compute needs. Your project solves a specific task, and that decision — model, deployment, cost — is made with your own data, not the plan's. The rest of the pieces are in the complete guide to the Plan IA360.