Saudi Arabia’s Year of AI 2026: What HUMAIN, Hexagon and a Million Trained Citizens Actually Change for Enterprise Buyers
Saudi Arabia designated 2026 the Year of Artificial Intelligence. The compute, the capital and the training were already built. Here is which of it changes an enterprise AI decision, and what a national programme still cannot answer for you.
In March 2026 the Saudi Council of Ministers designated the year the Year of Artificial Intelligence, under the patronage of the Crown Prince, who also chairs the Saudi Data and Artificial Intelligence Authority. Designations of this kind are easy to dismiss as ceremony. This one is not, because it arrives attached to compute, money and a training programme that have already been built.
This article sets out what actually exists as of 2026, what it changes for an organisation buying AI in the Kingdom, and which questions a designation year does not answer for you.
What the designation actually sits on top of
SDAIA has run the National Strategy for Data and Artificial Intelligence since 2019, organised around six pillars: ambition, competencies, policies, investment, innovation and ecosystem. The 2026 designation is the point at which several of those pillars stopped being roadmap and started being infrastructure you can rent, hire from, or be regulated by.
- 2019SDAIA establishedThe national authority for data and AI, with regulatory as well as promotional responsibility — the same body that now enforces the Personal Data Protection Law.
- May 2025HUMAIN launchedA Public Investment Fund company positioned as the national AI champion, with a 500-megawatt compute mandate and chip-supply arrangements with NVIDIA and AMD.
- 2025$9.1bn committedRaised by Saudi AI-sector companies across roughly 70 investment deals in a single year.
- Early 2026Hexagon inauguratedDescribed as the world’s largest government data centre, at 480 megawatts of capacity.
- 2026Year of AI declaredA Council of Ministers designation, with SDAIA issuing guidelines to unify national effort under it.
Read that sequence in order and the designation looks less like a launch and more like a ribbon cut on something already running. The compute, the capital and the national champion were in place first.
One million people, which is the number that matters most
The SAMAI initiative, led by SDAIA, reports having trained more than one million Saudi citizens in AI technologies between its launch in September 2024 and November 2025. Alongside it, the count of companies working in data and AI in the Kingdom has reached 664.
It also changes a practical thing about procurement: the assumption that an AI project must import its operators is weaker in 2026 than it was in 2023. That affects total cost, and it affects who is left holding the system when the integrator’s contract ends.
What a designation year does not decide for you
A national programme sets direction. It does not answer the questions a specific deployment turns on, and it is worth being precise about the boundary.
| Set by national programme | Still a per-deployment decision | |
|---|---|---|
| Availability of in-Kingdom compute | Yes | No |
| Whether your data may leave the Kingdom | Yes | No |
| Availability of trained Saudi practitioners | Yes | No |
| Which model your assistant runs on | No | Yes |
| Whether prompts reach a third-party API | No | Yes |
| Where conversation content is stored, and for how long | No | Yes |
| Who can read a transcript, and under what authority | No | Yes |
The first three arrived with the national programme. The last four arrive with your architecture, and no designation changes them.
The distinction matters because the two are easily conflated in a sales conversation. “Hosted in Saudi Arabia” is a statement about the first column. It is not a statement about the fourth row.
PDPL is the part with teeth
SDAIA enforces the Personal Data Protection Law, and in practice that makes data residency a precondition rather than a preference for any system handling citizen data. The relevant question at procurement is not whether a vendor can host in the Kingdom — most now can — but whether answering a question requires a call to a model outside it.
That is an architectural property. It is decided when the system is designed and cannot be switched off afterwards. We wrote the longer version of this argument in our comparison of the vendor categories, including the five different things “on-premise” is used to mean.
Arabic is a capability question, not a translation question
The Kingdom now has native Arabic models with real standing — ALLaM 34B has been ranked the most advanced Arabic large language model built in the Arab world on the MMLU benchmark, and its commercial deployment sits within HUMAIN, which SDAIA describes as the production and distribution channel for advanced Saudi models.
For a buyer, the practical test is narrower than a benchmark. Type the way your customers type, in dialect, mixing Arabic and English in one sentence, and see what comes back. That single test separates a bilingual product from a translated one faster than any specification sheet.
Positions describe where vendors typically struggle, not a measured benchmark of any product.
What this means if you are buying this year
Three things change in a designation year, and one thing does not.
- 1Budget approval gets easierA national designation gives an internal sponsor cover. The strategy question moves from “why now” to “why not us”.
- 2Vendor supply increases664 companies is a crowded field. More options, and more claims that need checking.
- 3Hiring gets easierA million trained people changes the operating assumption after handover.
- 4Diligence does not get easierA busy market is where unverifiable claims travel fastest. The questions get more important, not less.
If you take one habit from this year, make it the fourth line. We publish the questions we think you should ask every vendor — including us — because the ones that separate serious systems are usually the ones nobody volunteers.
Compute is the part people underestimate
Hexagon at 480 megawatts and HUMAIN’s 500-megawatt mandate are the kind of numbers that get quoted without much sense of what they buy. The useful translation is this: capacity at that scale means the constraint on a Saudi AI project in 2026 is no longer whether national compute exists. It is whether your particular workload needs it.
Most enterprise assistants do not. A retrieval assistant answering from a company’s own documents is a small model doing a narrow job, and it runs comfortably on hardware that fits in a rack in your own building. The national compute story matters enormously for training frontier models and very little for answering a customer’s question about an order.
The question worth asking internally is which of your AI ambitions actually require national-scale infrastructure, and which are better served by something small, local and entirely under your control. For most organisations the answer splits, and the split is not obvious until someone writes it down.
What 664 companies means in practice
A field that size is good news and a diligence problem in the same breath. More options, more specialisation, more competitive pricing — and a much larger surface of claims that nobody has checked.
Ordered by how often we see the claim, not by how often it is untrue. Most are defensible; almost none are self-evident.
One habit covers most of this. Ask any vendor to show the system failing — pose a question adjacent to the demo material but not in it. A grounded system declines. A weak one improvises fluently, and you learn more in ten seconds than in an hour of slides.
The sectors moving first, and why
Adoption is not spreading evenly, and the pattern is not really about technical readiness. It tracks two things: how much repetitive contact a sector handles, and how tightly its data is regulated.
| High contact volume | Tight data rules | Moves early | |
|---|---|---|---|
| Government services | Yes | Yes | Yes |
| Banking and finance | Yes | Yes | Yes |
| Healthcare | Yes | Yes | Yes |
| Telecom | Yes | No | Yes |
| Retail and e-commerce | Yes | No | Yes |
| Real estate | No | No | No |
| Heavy industry | No | No | No |
High volume creates the business case; tight data rules push the deployment on-premise rather than away from it.
The counter-intuitive column is the middle one. Regulation is usually described as a brake on AI adoption. In the Kingdom in 2026 it is closer to a steering wheel: it does not slow the decision down so much as decide the architecture for you, which is why regulated sectors are moving and moving on-premise.
A reasonable plan for the rest of the year
If a designation year is going to be useful to you rather than merely noted, it helps to have a shape for what to do with it.
- 1Pick one workload with a countable costNot “AI strategy”. One process with a number attached — tickets per month, calls per day, documents per week.
- 2Write the data rules down before the demoWhere content may sit, who may read it, how long it is kept. Bring it to the vendor rather than asking what they offer.
- 3Test in your own language, on your own materialDialect, mixed Arabic and English, a real uploaded document. Not the vendor’s dataset.
- 4Decide the exit before the entryExport format, deletion mechanics, and who holds the keys. Settle it before signature, not during.
None of that is specific to 2026. What the designation changes is that steps one and four are now much easier to get approved, because the strategic argument has been made nationally and you no longer have to make it yourself.
The build-out in numbers
Put the published figures side by side and the shape of the programme is easier to read than any single announcement conveys. These are national figures reported by the bodies that ran the programmes, not our measurements.
Bar lengths are scaled for comparison within each unit and are not a single common scale — megawatts, dollars, deals, companies and people are different quantities.
The two compute numbers sit at the top because they are the hardest to replicate and the easiest to misread. Capacity of that size is a national asset. It is not, by itself, an argument about where your own customer conversations should be processed.
Three deployment shapes, and who each one suits
Once the national picture is set aside, an enterprise decision reduces to a fairly small number of shapes. The differences that matter are not feature lists but where the data goes and who is left holding the system.
| Public cloud AI | National platform | On-premise, self-contained | |
|---|---|---|---|
| Fastest to a first demo | Yes | Yes | No |
| Content stays inside your estate | No | No | Yes |
| Answering needs no external model call | No | No | Yes |
| Scales past national-level volume | Yes | Yes | No |
| Works with the building offline | No | No | Yes |
| You hold the encryption keys | No | Varies | Yes |
| Suits a regulated pilot in one department | No | No | Yes |
Not a ranking. The right column is where Elbi sits; the first two are better answers for workloads that need national-scale compute or global reach.
Most organisations discover they need more than one of these. The mistake is assuming a single choice has to cover every workload, which usually results in the most sensitive traffic inheriting the architecture chosen for the least sensitive.
The honest limits of this article
Every figure above is a published national figure, and none of it is our measurement. Funding totals, company counts, training numbers and data-centre capacities are reported by the bodies that ran the programmes. They are useful for direction and poor for precision, and we have not independently audited any of them.
We also have no view on which national model suits your workload. What we can say precisely is what our own system does, because we run it: it answers from your material, on hardware you control, with no external model call in the path.
Common questions
The Council of Ministers made the designation in March 2026 under the patronage of the Crown Prince, who chairs SDAIA. It follows several years of build-out rather than preceding it: SDAIA was established in 2019, HUMAIN launched in 2025 with a 500-megawatt compute mandate, and the Hexagon data centre was inaugurated in early 2026.
HUMAIN is a Public Investment Fund company positioned as the national AI champion, launched in May 2025 with chip-supply arrangements with NVIDIA and AMD. ALLaM is the Saudi Arabic large language model, and its commercial deployment sits within HUMAIN, which SDAIA describes as the production and distribution channel for advanced Saudi models.
No. The Personal Data Protection Law and SDAIA enforcement of it are what set the obligations, and they applied before the designation. In practice PDPL makes in-Kingdom data residency a precondition rather than a preference for systems handling citizen data.
Not necessarily. Hosting describes where the application runs. It does not describe whether producing an answer requires a call to a model hosted elsewhere. That is an architectural property, decided when the system is designed, and it cannot be switched off later. Ask specifically whether any external model API is called to answer a question.
The SAMAI initiative led by SDAIA reports training more than one million Saudi citizens in AI technologies between its launch in September 2024 and November 2025, and the number of companies working in data and AI in the Kingdom has reached 664. These are published national figures rather than independently audited ones.
Keep reading
PDPL and AI chatbots in Saudi Arabia
An assistant that talks to customers is a personal data processing activity whether anyone described it that way in the project plan or not. Most of the work is design decisions taken early, cheaply, before anything is built.
Read Article
Arabic AI chatbots for Saudi customers
Almost every assistant sold in the Kingdom supports Arabic. Far fewer understand the Arabic customers actually type — which is not the Arabic of official documents, and is where most deployments quietly fail.
Read Article
The future of AI assistants in Saudi Arabia
Vision 2030 turned AI adoption in the Kingdom from an experiment into an expectation. What follows is less about better models than about a shift in what organisations ask assistants to do — and what they will be held to.
Read Article