Running an Arabic LLM on-premises in Saudi Arabia: what it takes
Many organisations in the Kingdom want AI that understands the Arabic their customers write and keeps their data within the country. Cloud model APIs partly meet the first need, but not the second. This guide explains what it takes to run an Arabic language model on your own servers in Saudi Arabia, covering rules, infrastructure, language, and questions to ask before signing. It is not legal advice.
Why organisations in the Kingdom keep AI in-house
The Personal Data Protection Law has been fully in force since September 2024, with SDAIA responsible for enforcing it. Personal data can currently be transferred abroad, provided safeguards are in place. There is not yet a general requirement to store it in the Kingdom. However, SDAIA’s draft amendments to the Implementing Regulation explicitly introduce such a requirement. NCA cybersecurity controls and SAMA’s outsourcing rules also set conditions for government entities and banks. For details, see Saudi Arabia’s AI and data rules in 2026.
If the model, document index and logs all run on your servers, data is not transferred at all, so there is no transfer to justify. To compare the Gulf countries and Egypt in one table, see data rules by country.
What needs to run inside your network
Running the model alone is not enough. To keep everything inside your organisation, these components must run on your own infrastructure:
- The model itself, without contacting any external service while it runs.
- The search index for your documents, with each document’s permissions enforced. We explain this in how an AI assistant should only know what you know.
- The audit log records every question, answer and action, and cannot be altered.
- The safety layer and policy rules your team can inspect.
- Updates: they arrive on media for you to check first, and you apply them when you choose.
How much infrastructure is needed
Base your estimate on the number of users and the volume of documents, not the size of the model. Orchestra runs on one GPU-class server rather than a cluster. For an organisation of 500 users, our planning estimate is two NVIDIA A100 80 GB GPUs. We confirm this with you in the Discover phase before anything is purchased. You buy the hardware from your own vendor using a short specification we provide for pricing. Read more in how we determine deployment size.
Arabic in the way your customers write
Your customers in the Kingdom do not write in textbook Arabic. They use Gulf dialects and mix Arabic with English. Test any model on real messages from your customers, not just against a general benchmark. We built Orchestra from scratch for formal Arabic and the Arabic people write. We have published our results on general benchmarks, and are now measuring its performance on dialects.
Want to see Orchestra work on your own documents? Book a 30-minute call with one of our engineers
If your servers are not yet ready
You can start with Orchestra hosted on our servers, which operate in Saudi Arabia today, using data that is allowed to leave your network. Once your servers are ready, we move it to them and you switch to the fixed annual licence. See the details in the hosted option.
Questions to put to any vendor
- Where are the model, document index and logs stored? Do they all stay inside our network?
- Does anything leave our network when the system runs or is updated?
- Has the model been tested with our messages and the dialects our customers use?
- Does search enforce permissions for each document, so people see only what they are allowed to see?
- Are all questions, answers and actions logged, and can auditors access the log?
- Is the price fixed for the year, or does it increase with each token?
- What does the pilot cost, what are its success criteria, and is the fee credited toward the licence if we proceed?
Working with you
We begin with a paid pilot using your documents and success criteria agreed in advance. If you sign the licence within 90 days after the pilot ends, the full pilot fee is credited against the licence price. We then deploy the system on your infrastructure and operate it with you. See how Orchestra pricing works for details.
Written by Orchestra