How much hardware does on-premises Arabic AI need? Start with your people
Most on-premises AI quotes start from the model. We size Orchestra from the workload, then hand you a short specification your own vendor can price.
Blog
Notes from the field.
What we learn deploying Arabic AI inside secure environments.
AI and data rules in the Gulf and Egypt, 2026: all seven countries in one table نزّل قائمة المراجعة: قواعد الذكاء الاصطناعي والبيانات في الخليج ومصر 2026 (PDF) Download the checklist: AI and data rules in the Gulf and Egypt, 2026 (PDF)An AI assistant should only know what you know
An assistant that searches every company document under one account can show people what they should never see. How we apply each document’s permissions inside the search, and why an assistant acts with the permissions of the person it serves.
Saudi Arabia’s AI and data rules in 2026: a data law fully in force, no AI law yet, and a stack of drafts
The PDPL has been fully enforced since September 2024 and SDAIA is applying it, cybersecurity and banking rules add their own conditions, and there is no AI law yet. What applies to an AI deployment in the Kingdom today.
Local Arabic AI: what “local” really means, and how to check it before you buy
“Local” means different things to different vendors: hosted in-country, running inside your premises, or fully air-gapped. It should also mean Arabic the way your customers write it. How to tell them apart, and five questions for any vendor.
A ChatGPT alternative for government: a fair comparison between a cloud assistant and Arabic AI that runs inside the entity
ChatGPT is a strong tool, but a government entity asks different questions: where the correspondence goes, who controls the model, and whether it writes formal Arabic and understands citizens’ dialects. A fair comparison on six points.
An Arabic language model inside your organisation: what it actually takes to run one
Running an Arabic language model inside your organisation is not buying a weights file. The six parts it actually needs: the model, retrieval from your documents, hardware, systems and access rights, safety and logs, and evaluation and operation after launch.
Kuwait and AI in 2026: few rules today, and a clear direction towards keeping data at home
Kuwait has no general data protection law and no AI law, and its old data classification policy was repealed in 2024. But the telecoms regulator is signalling new data sovereignty rules. What that means for where your AI runs.
Oman and AI in 2026: a data law now fully enforced, and a new right to object to automated decisions
Oman's data protection law has been fully enforceable since February 2026, and an amendment in September 2026 added a right to object to automated decisions. Cloud and central bank rules require prior approval to move sensitive data abroad. Where running AI on-premises makes sense.
Bahrain and AI in 2026: a data hub, but automated processing of sensitive data needs authorisation
Bahrain encourages cloud and has a unique data embassy law, so location is not the first argument there. But its data protection law requires prior authorisation for automated processing of sensitive data, and an AI bill is still under debate. What that means for your deployment.
What is an on-premises LLM, and when is it the right choice?
How a language model that runs on your own hardware differs from one you call through a cloud API: the data path, the cost model, control, and when each one makes sense.
Saudi Arabia may require personal data to stay in the Kingdom. What it means for your AI deployment
SDAIA's draft amendments to the PDPL Implementing Regulation add an express requirement to store personal data in the Kingdom, open for comment until 5 November. What changes for AI systems.
The UAE’s AI and data rules in 2026: a new federal authority, a law still waiting on its regulations, and free zones that move faster
The UAE has created a federal AI and data authority, its data protection law still awaits implementing regulations, and the DIFC has regulated autonomous systems since 2023. How to design a deployment that holds up however this lands.
Qatar and AI in 2026: a law that lets data travel, and the exceptions that actually matter
Qatar's data protection law imposes no general localisation and tells controllers not to restrict cross-border data flows. But special-nature data, the Qatar Financial Centre and the banking sector have their own rules. Where running AI on-premises makes sense.
On-premises or cloud language model? What actually changes in cost, control and compliance
A practical comparison between running a language model inside your premises and calling one from a cloud provider: where the data goes, how cost behaves, who controls the model, and when the cloud is the better choice.
How to evaluate an Arabic language model before you buy it
Published scores do not tell you how a model will handle your customers. A practical six-step method to test any Arabic model on your own data and dialects before you buy, with a checklist that works for any vendor.
Egypt has set its AI rules: in-country data, risk tiers and frugal models. What it means for your deployment
Compliance with Egypt's data protection law is due from 1 November 2026, and the national AI governance framework asks for small models first. A practical checklist for organisations deploying AI in Egypt.
Jais 3 and Falcon-Emirati: what October’s Arabic model launches mean for a buyer
Two Arabic model releases in one week, open weights and dialect-specific models. What has actually changed, and what to check before you buy.
Half of enterprises saw an AI agent act without permission. Five questions before yours goes live
October’s agentic AI security reports agree on the same gaps: tracing, containment and identity. How to close them before launch, not after.
Sovereign AI in the Gulf: the region tops the index, and the real question is where your system runs
Counterpoint’s index puts the Middle East first, Accenture warns of a compute gap, and a first regional bank adopts locally hosted models. What it means for your decision.
What “air-gapped” really means for an AI deployment
Air-gapped is more than unplugging a cable. Here is what has to stay inside, and how updates and evaluation still happen.
On-premises, air-gapped or private GPUs: choosing where your Arabic AI runs
Three deployment models, one stack. A practical way to choose based on your data, your regulator and your team.
Five layers, with a security control at each one
From ingress to operations: where each security control sits in an enterprise AI stack, and why sensitive actions go to a human first.
Discover, deploy, operate: how a project with NoorAI runs
Three phases, the same security controls throughout: what happens in each, what we need from you, and when users first see the system.
Evaluated before launch: what an evaluation suite is, and why every deployment gets one
Public benchmarks tell you how a model does on someone else’s questions. An evaluation suite tells you how it does on yours, before a single user sees it.
Three commitments in every deployment
We are an applied AI lab, not a reseller. What that means in practice for your project: security, your premises, Arabic at the core.