Contents
Three parts. Why it matters, what we build, and how we work together. Every claim is sourced on the last page.
01Why it matters
02What we build
AI that speaks your language
Specialist Arabic–English models, built on your own data and run inside your borders.
Most AI in use today was built for English first and trained on the open internet. For a bank in Dubai or a ministry in Abu Dhabi, that creates three problems at once. The model is weakest in the language customers actually use. It knows nothing about the organisation’s products, policies or regulator. And it often runs somewhere the data is not allowed to go.
Sharq closes those gaps with specialist models. We start from a strong open-weight model and post-train it on the organisation’s own documents, workflows and terminology, in Arabic and English together. That includes Gulf dialect and the mid-sentence switches between languages that real work involves.
The result is a model the client owns. It is tested on the client’s real tasks before it goes live. It is trained and run on compute inside the country: the client’s own servers, or sovereign cloud already operating in the UAE, such as an OCI Dedicated Region or Core42.
Two things we do not do. We do not build a foundation model from scratch, because strong open-weight models already exist and the value is in adapting them. And we do not ask anyone to take our word for results. Every engagement starts with one workload and a scorecard agreed in advance, so the case for going further is made on the client’s own data.
We don’t ask anyone to take our word for results. We agree the test first.
Where general-purpose AI falls short
Language
Gulf teams mix Arabic, English and dialect in one sentence. Most models are tested one language at a time, in standard form, so the test rarely matches the work.
Location
Regulated data often has to stay in the country. In the UAE, banks must hold and store all consumer and transaction data within the UAE, which rules out many foreign-hosted AI services. Since February 2026 the Central Bank has run its own sovereign financial cloud.
Fit
A general model does not know your products, policies or regulator. Staff re-explain that context in every prompt, or correct the output afterwards, by hand.
ممكن تشيك الـ KYC status للعميل قبل الـ onboarding؟
“Can you check the client’s KYC status before onboarding?” Gulf Arabic with English terms mixed in: ordinary at work, and hard for models tested in one language.
The cost shows up as extra compute, manual correction, and sensitive work that never reaches AI at all.
None of these gaps is a flaw in the models themselves. General-purpose systems are built to be good at everything for everyone. Gulf organisations need something narrower: a model that handles their mix of languages, knows their business and runs where their regulator expects.
The regulator is moving the same way. In February 2026 the Central Bank of the UAE launched a sovereign financial cloud for licensed financial institutions, built with Core42 as dedicated, isolated infrastructure. When the sector’s own regulator builds sovereign infrastructure, local models stop being a preference and become the baseline.
The first of the costs above is the easiest to measure, because it shows up in every request. The next section sets out what we found when we counted it.
The Arabic token premium
The same text takes more tokens in Arabic than in English. How many more depends on the model.
more tokens for identical Arabic text on the least efficient tokenizer we tested than on the most efficient.
| Tokenizer | Arabic | English | Ratio |
|---|---|---|---|
| GPT-4 · cl100k | 5,251 | 2,016 | 2.60 |
| Llama 3 | 2,856 | 2,016 | 1.42 |
| Gemma 3 | 2,609 | 2,071 | 1.26 |
| GPT-4o · o200k | 2,378 | 2,017 | 1.18 |
What it means
Tokens are what models compute and what most providers bill. Choosing and adapting the base model for your language mix changes the running cost of every Arabic request, before any other saving.
Method. Official Arabic and English texts of the UN Universal Declaration of Human Rights (NLTK udhr2 corpus), counted in full with each model’s published tokenizer, September 2026. Caveats. Token counts are not prices; per-token prices vary by provider. Formal Arabic only: Gulf dialect and mixed-language text were not measured. Independent research points the same way: Petrov et al. (NeurIPS 2023) measured a 4.40× premium for Standard Arabic on GPT-2 using FLORES-200.
Arabic tokens for the same text: GPT-4 tokenizer, then GPT-4o tokenizer. The model choice alone cuts the count by more than half.
What Sharq builds
An open-weight base model, post-trained around your organisation on compute inside the UAE, and delivered with everything needed to run it.
Bilingual
Trained on Arabic, English and Gulf dialect together, including requests that mix all three.
Specialist
Learns your products, policies and terminology from your own documents, so each request needs less instruction.
Sovereign
Runs on your premises or on sovereign cloud inside the UAE. You own the weights and can inspect the whole model.
Local training
Post-trained on GPUs inside the UAE, linked by RDMA networking. Your data is never exported to train it.
Customer-facing AI is the front door. The savings usually sit behind it.
Start where the work is high-volume, narrow and regulated, and where the data should not pass through a general-purpose API. Sharq sits alongside the assistants customers already see, not in place of them.
- KYC and compliance review
- Credit memo drafting
- Fraud alert triage
- Policy and procedure search
- Citizen enquiries in Arabic and dialect
- Policy and legal research
- Bilingual correspondence
- Case summaries for officers
The Abu Dhabi Judicial Department launched the first phase of its Judicial AI Platform in September 2026: case analysis and judgment drafting, with every output verified by a judge. A regulated workflow, running today.
Everything happens inside your boundary
- 01
Define
Agree the tasks and what good looks like, in both languages.
- 02
Prepare
Select and clean your documents, tickets and records.
- 03
Post-train
Specialise an open-weight model on your data, on GPUs inside the UAE.
- 04
Evaluate
Test against your acceptance bar before go-live.
- 05
Run
Deploy, monitor and refresh as your data changes.
Your data and the model weights stay inside this boundary, and inside the UAE, at every step.
The compute is already here
AI compute at national scale is being built inside the UAE. A sovereign model no longer has to trade local control for scale.
Local compute
+ full model visibility
Keeping data in the country is only the start. A sovereign model is also trained and run on hardware inside the UAE, and nothing about it is hidden from you: you hold the weights, can inspect the model, and can trace how it was trained. Data localisation alone gives you none of that.
UAE–US AI Campus, Abu Dhabi, at full build-out. The first 500 MW phase is due online by the end of 2026.
Stargate UAE, a cluster on the same campus from G42, OpenAI, Oracle, NVIDIA, Cisco and SoftBank. First 200 MW due in 2026.
OCI Dedicated Region
Oracle cloud services inside Abu Dhabi’s borders, on Core42 infrastructure with NVIDIA GPUs, so sensitive data does not leave the emirate.
Core42 sovereign cloud
G42’s sovereign cloud and GPU compute, hosted in the UAE. Core42 also built the Central Bank’s sovereign financial cloud.
Your own premises
Your servers, with the model packaged as a self-contained, self-hosted deployment that you control end to end.
Post-training on local accelerators, linked by RDMA networking.
The compute provider for each engagement is agreed at scoping, from UAE-hosted options such as those above.
RDMA lets GPUs exchange data directly, which keeps training at enterprise speed. Because the accelerators sit inside the UAE, your data never has to be exported to train the model, and foreign data-export exposure is removed at the source.
Start with one workload and a clear decision at the end
- Phase 1
Scope
Choose one workload. Agree success measures and today’s baseline.
- You provide
- A business owner and sample data
- Phase 2
Build
Prepare the data, then post-train and iterate on an open-weight model.
- You provide
- Data access, reviewers, and compute in your environment or UAE sovereign cloud
- Phase 3
Prove
Test against the baseline on held-out work, in Arabic and English.
- You provide
- Reviewer sign-off
- Phase 4
Decide
Move to production, extend to a second workload, or stop.
- You provide
- A go or no-go decision
Durations and commercial terms are agreed at scoping. Security review and data-access approvals should start in Phase 1.
We’d rather show it on your data than put it on a slide.
Every pilot ends with a scorecard agreed before work starts. The figures come from your workload, measured against how you do it today. We do not quote savings in advance; we measure them.
- Task accuracyAgainst your current process, on held-out work
- Cost per 1,000 requestsCompute and licences at your expected volume
- Correction rateShare of outputs your team has to fix
- Response timeMeasured at your expected production load
- Data boundaryConfirmation that data and weights stayed with you, inside the UAE
Pick one workload. We’ll make the case on your data.
One scoping session with your business and technology leads. We choose the workload, agree the scorecard, and send a written pilot proposal.
Notes
- Results.
- Sharq does not quote savings in advance. Pilot results are measured on each client’s own workload, against a scorecard agreed before work starts.
- Token analysis.
- Counts come from the official Arabic and English texts of the UN Universal Declaration of Human Rights, tokenised in full with each model’s published tokenizer in September 2026. Token counts are not prices. Formal Arabic only; dialect and mixed-language text were not measured.
- Comparison.
- The options table describes typical patterns, not any named provider.
- UAE compute.
- Capacity figures are planned and phased, as announced by the operators. Named platforms are deployment options, not Sharq partners.
- Images.
- Original renders made for Sharq.
Sources
- 01Central Bank of the UAE Rulebook, Consumer Protection Standards, Article 6: Protection of Consumer Data and Assets, clause 6.1.6.3. rulebook.centralbank.ae
- 02Petrov et al., “Language Model Tokenizers Introduce Unfairness Between Languages,” NeurIPS 2023. arxiv.org/abs/2305.15425
- 03United Nations, Universal Declaration of Human Rights, official Arabic and English texts (NLTK udhr2 corpus).
- 04Sharq token analysis, September 2026: GPT-2, cl100k_base, o200k_base, Llama 3 and Gemma 3 tokenizers.
- 05Central Bank of the UAE, “CBUAE Develops World’s First Sovereign Financial Cloud Services,” 25 February 2026. centralbank.ae
- 06Gulf News, “Stargate UAE: first phase of 200MW to complete by 2026,” 16 October 2025.
- 07Zawya, “Largest AI campus outside of US to open in Abu Dhabi this year,” 2026: first 500 MW phase of the 5 GW UAE–US AI Campus.
- 08NVIDIA, “Oracle and NVIDIA Accelerate Sovereign AI, Enabling Abu Dhabi’s AI-Native Government Transformation,” 14 October 2025.
- 09Nukta, “Abu Dhabi launches first phase of Judicial AI Platform,” 28 September 2026.