Decision guide AI under the company's control
Local AI or cloud AI in Malta
A fair comparison of the three ways a company can run AI, with costs, rules and a decision you can follow step by step.
- Published
- Updated
- Reading time
- 7 min read
- Editor
- David Lahuta
In short
- The deciding question is which data may leave the company, not which model is best.
- Sensitive data and steady use point to your own machines or a European cloud, light use or the need for the strongest model to a provider's service.
- The AI Act's duties for a company using AI are the same in every option.17
What are the three options?
- Your own machines (local, or on-premise). An open model, which anyone may download, on hardware the company owns or rents for itself.
- A European cloud. A provider runs the model in EU data centres under EU law.
- A US provider's service. A US AI company's business plan or programming interface, with or without a European data option.
In 2025, 19.95% of EU enterprises with ten or more staff used AI.1 Of those that considered it but did not, 70.89% named a lack of expertise, 52.52% unclear legal consequences and 48.83% data protection.1
Comparison
| Criterion | Your own machines | European cloud | US provider's service |
|---|---|---|---|
| Cost over several years | Hardware or rent, then power and staff time. Flat as use grows | Per use or reserved capacity. Grows with use | Per user or per use. Grows with use |
| Data and rules | Data stays in the house. Personal data rules still apply | Data stays in the EU, no transfer abroad to assess | Data may go to the US, on a valid legal basis under appeal35 |
| People needed | Someone to install, update and watch it | Someone to set up access and records | Little beyond setup and rules of use |
| Speed | Weeks to a working state | Days | Days, the newest models at once |
| Risk | Quality gap, hardware failure, security is your job | Changed terms or prices, fewer models | Terms, prices and data location set by the provider |
| The way out | You keep everything | Easy with standard connections | Easy with standard connections, hard if built on its own tools |
What are the pros and cons?
Your own machines
Pros
- No company data leaves the building.
- Cost hardly grows with use, within capacity.
- No provider can change terms or switch a model off.
Cons
- Open models lag the best ones.6
- Hardware, updates and security are your job.
- It needs skilled staff.
European cloud
Pros
- Data stays under EU law, in EU data centres.
- No hardware to buy, running in days.
- You pay for what you use.
Cons
- Fewer models than the largest providers.
- The bill grows with use.
- Terms can change.
US provider's service
Pros
- The strongest models on release.
- Nothing to install, ready in days.
- Low cost for light use.
Cons
- Personal data may go to the US.
- The bill grows with use.
- Easy to tie your work to its own tools.
What does each option cost?
Prices are per token, a word or part of one.
Costs
Your own machines. To start, a gaming graphics card under USD 2,500 runs open models as good as the best of six to twelve months earlier.6 A dedicated server rented in a European data centre costs from EUR 214 a month for one card with 24 GB of memory to EUR 3,599 for four cards with 384 GB, plus a setup fee, before VAT.9 To run, power and part of a person's time. Hardware costs have fallen 30% a year.8 To leave, little.
European cloud. To start, almost nothing. To run, a German provider charges EUR 0.65 per million tokens in and out for a large open model, EUR 0.10 in and 0.30 out for a small one.10 A European model maker charges USD 0.5 in and 1.5 out for its large model, and USD 24.99 per user a month for its team plan.11 To leave, little with standard connections.
US provider's service. To start, almost nothing. To run, one provider's business plan costs USD 25 per user a month, or 20 billed annually.14 Its programming interface costs from USD 0.10 in and 0.50 out per million tokens for its smallest current model to USD 8 in and 40 out for its most capable.12 To leave, little with standard connections, much where the work rests on its own tools.
The price of a given level of AI fell over 280-fold from November 2022 to October 2024.8
How demanding is each option?
Complexity
One scale, low, medium or high.
Your own machines, high. Skilled people for models, hardware and security. Running in weeks. Risks are a weaker model than hoped, a failed card or an unwatched security gap.
European cloud, medium. The provider runs the machines, you set up access, records and connections. Running in days. Risks are a model that misfits the task, or changed terms.
US provider's service, low to medium. Low for a business plan, medium once connected to company systems. Running in days. Risks are personal data in the wrong place, or staff in private accounts.
When are your own machines the right choice?
When data must not leave the company, use is steady and heavy, and someone can run it. A model a year behind may do for documents and internal knowledge, so test it on your tasks. In March 2026 the top closed model led the top open one by 3.3% in the Arena ranking of human votes.7 Large US providers offer or have announced versions for the customer's own data centre, connected or cut off.1516
When is a European cloud the right choice?
When confidential or personal data goes in, you want no transfer outside the EU to assess, and nobody in the house should run servers. US providers sell European offers too, one processing customer data in the EU for selected models.15
When is a US provider's service the right choice?
When use is light, no sensitive personal data goes in, and no contract or regulator keeps data in the EU. Then a business plan serves well, with data terms such as no training on your content by default.14 One provider's own interface offers processing only in the US or worldwide, and stores data at rest in the US.13
How do you decide, step by step?
Step by step to your decision
- Will personal or confidential data go into the AI? If not, a US provider's business plan is a sound start. If yes, go on.
- Must the data stay in the company or in the EU, by contract, regulator or customer? In the company, your own machines. In the EU, a European cloud or your own machines.
- How many people, how often? Steady, heavy use favours your own machines, occasional use paying per use.
- Do you need the strongest model, or is six to twelve months behind enough?6 If the strongest, a provider's service where data may leave.
- Who will run it, your IT or a supplier? Without either, a cloud.
- The recommendation. Sensitive data and steady use, your own machines. Sensitive data and uneven use, a European cloud. Neither, a provider's service. Both, a mix under one control.
What to answer before deciding
Before you decide
- Your uses, one line each.
- Which data each use touches, and which is personal or confidential.
- Who may see what.
- Which records you need of questions, answers and actions.
- Expected users and questions a day, and who will run it.
- How you would leave, and what you keep.
What do companies usually get wrong?
- Buying hardware before knowing the uses. The uses decide the model, the model the hardware.
- Taking a European data option as the whole legal answer. It settles where data is processed, not who may see it or why.
- Assuming local means secure. Without access control and records, it is one more place for data to leak.
- Forgetting the model's history. A user of a model is expected to check it was not built on unlawfully processed personal data.2
- Tying everything to one provider's own tools. Standard connections keep the way out open.
- Deciding per department. Five departments, five tools, no overview.
What this means for a company in Europe
Data protection law applies to personal data in all three options. Transfers to certified US companies rest on the EU-US Data Privacy Framework of 10 July 2023.3 The General Court upheld it on 3 September 2025, after the Court of Justice struck down two earlier ones.4 An appeal was brought on 31 October 2025.5 AI literacy and the first bans have applied since 2 February 2025, most rules since 2 August 2026, and the high-risk rules of Annex III apply from 2 December 2027.17
In Malta
Malta's Information and Data Protection Commissioner explains that where an AI system processes personal data, the GDPR applies alongside the AI Act and takes legal precedence. The Commissioner is also a fundamental rights authority and a market surveillance authority for some high-risk areas.18
The Artificial Intelligence Regulations, 2025 (L.N. 226 of 2025) designate the Malta Digital Innovation Authority as the single point of contact for the AI Act.19 We found no Maltese rule on transfers beyond the GDPR and no guidance of the Commissioner's own on choosing an AI service, so the EU rules above hold, and the Commissioner is the authority to ask.
Next step
If it helps, we would gladly sort your planned uses by which data may leave the company, to show which option or mix fits. Our AI Business Harness is one AI for the whole company, in Europe or in your own infrastructure, under your control, every step recorded. The decision is yours.
Sources
- 1. Eurostat, Use of artificial intelligence in enterprises (data extracted December 2025). Read on 17 November 2026
- 2. European Data Protection Board, Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models, adopted 17 December 2024. Read on 17 November 2026
- 3. European Commission, EU-US data transfers. Read on 17 November 2026
- 4. General Court of the European Union, press release 106/25, judgment in Case T-553/23 Latombe v Commission, 3 September 2025. Read on 17 November 2026
- 5. Official Journal of the European Union, C/2025/6610, appeal brought on 31 October 2025 in Case C-703/25 P. Read on 17 November 2026
- 6. Epoch AI, Frontier AI performance becomes accessible on consumer hardware within a year, 15 August 2025. Read on 17 November 2026
- 7. Stanford HAI, The 2026 AI Index Report, chapter 2, technical performance, April 2026. Read on 17 November 2026
- 8. Stanford HAI, The 2025 AI Index Report, top takeaways. Read on 17 November 2026
- 9. Hetzner Online, dedicated GPU servers, price list. Read on 17 November 2026
- 10. IONOS, AI Model Hub, prices per million tokens. Read on 17 November 2026
- 11. Mistral AI, pricing. Read on 17 November 2026
- 12. Anthropic, Claude API pricing. Read on 17 November 2026
- 13. Anthropic, Claude API data residency. Read on 17 November 2026
- 14. Anthropic, Claude plans and pricing. Read on 17 November 2026
- 15. Google Cloud, Sovereign AI on your terms, made with Europe, 17 November 2026. Read on 17 November 2026
- 16. Microsoft, Advancing digital sovereignty with greater control and choice, 20 October 2026. Read on 17 November 2026
- 17. European Commission, AI Act Service Desk, timeline for the implementation of the EU AI Act. Read on 17 November 2026
- 18. Information and Data Protection Commissioner, Artificial Intelligence (updated 31 July 2026). Read on 17 November 2026
- 19. Legislation Malta, L.N. 226 of 2025, Artificial Intelligence Regulations, 2025. Read on 17 November 2026
Change log
- Published.