HUMAIN, ALLaM and the Saudi Sovereign AI Stack: Which Layer You Are Actually Buying
A Saudi model served over a foreign cloud is still a foreign network call. Here is what HUMAIN provides, where ALLaM actually runs, and the six questions that separate a real residency claim from a hosting address.
“Sovereign AI” has become the most-used and least-defined phrase in Saudi enterprise technology. It gets applied to a data centre in Riyadh, to a model trained on Arabic, to a cloud region with a local name, and to systems that are none of those things. The four are not equivalent, and an organisation that treats them as equivalent will sign for one and receive another.
This article separates the layers. It describes what HUMAIN is and what it provides, what ALLaM is and where it runs, and — the part that decides most procurement outcomes — the application layer sitting above both, which no national programme supplies for you.
The stack has four layers, and they are owned by different people
Almost every disappointment in a sovereign-AI purchase traces back to a buyer securing one layer and assuming the others came with it. Set them out explicitly and the assumption becomes visible.
| Layer | What it is | Who typically provides it in the Kingdom |
|---|---|---|
| Facility | The building, power, cooling, physical security | HUMAIN, national data centres, in-Kingdom cloud regions, or your own server room |
| Compute | GPUs and the scheduling around them | HUMAIN, cloud regions, or hardware you buy outright |
| Model | The weights that produce the text | ALLaM, open-weight models, or a foreign API |
| Application | Retrieval, permissions, escalation, logging, the interface | Your vendor — no national programme supplies this layer |
A contract that secures the first three and leaves the fourth unexamined is the common failure mode.
The fourth row is where the answer your customer actually reads is produced. It is also the only row where the decision about whether a request leaves the country is made in code rather than in a data-centre address.
HUMAIN: the compute and distribution layer
HUMAIN is a Public Investment Fund company launched in May 2025 and positioned as the national AI champion. Its public mandate covers roughly 500 megawatts of compute capacity, supported by chip-supply arrangements with NVIDIA and AMD. SDAIA has described it as the production and distribution channel for advanced Saudi models.
- 2019SDAIA establishedNational authority for data and AI, holding regulatory as well as promotional responsibility.
- 2024ALLaM publishedThe Saudi Arabic large language model, developed under the national AI programme.
- May 2025HUMAIN launchedPIF company with a 500-megawatt mandate and NVIDIA and AMD supply arrangements.
- 2025$9.1bn raisedCommitted across roughly 70 investment deals into Saudi AI-sector companies in one year.
- Early 2026Hexagon inauguratedDescribed as the world’s largest government data centre, at 480 megawatts.
Published national figures. Bars are scaled for comparison within this chart only — the units differ and are not interchangeable.
The practical consequence for a buyer is narrow but real: raw capacity is no longer the binding constraint on a Saudi AI project. If your workload genuinely needs national-scale compute — training a foundation model, serving a very large model to millions of concurrent users — that capacity exists domestically for the first time.
ALLaM: the model layer, and where it actually runs
ALLaM is the Saudi Arabic large language model developed under the national AI programme. It is notable for being trained with Arabic as a first-class language rather than as a translated afterthought, which shows up most clearly in dialect handling and in Arabic instruction-following.
The distribution point matters more than most buyers expect. ALLaM has been made available through commercial cloud platforms including Microsoft Azure AI and IBM watsonx. That is a genuine convenience and a genuine subtlety: a Saudi model accessed over a foreign cloud API is still a network call to that cloud. The nationality of the weights does not determine the path of the request.
| How ALLaM is consumed | Where the request goes | Data-residency position |
|---|---|---|
| Via a foreign cloud API | Out to that provider’s endpoint | Depends on the region and contract; the call leaves your network |
| Via an in-Kingdom cloud region | To that region’s endpoint | Stays in-country if the region and its dependencies are in-country |
| Self-hosted on your own hardware | Nowhere — it runs where you put it | Determined by you, verifiable by inspecting your own network |
This is why “we use a Saudi model” answers a different question from “our data stays in the Kingdom”.
Arabic capability is a model property; data residency is an architecture property
These two are constantly conflated in sales conversations, and they are independent. A model trained in Riyadh can be served from Ireland. An American open-weight model can be run on a server in Jeddah with no outbound connection at all. Neither fact follows from the other.
Two independent axes. Buyers who collapse them into one question get an answer to neither.
The question that separates architectures
There is a single procurement question that cuts through every variation of this discussion, and it is not “where are you hosted”. It is:
“Yes, but to an in-Kingdom region” is a legitimate answer and a defensible architecture. “No” is a different architecture. What should worry a buyer is a vendor who cannot answer at all, or who answers by describing where their servers are — because that is an answer to the hosting question, which was not the question asked.
Three deployment shapes, honestly compared
| Foreign cloud API | In-Kingdom cloud region | On-premises / self-hosted | |
|---|---|---|---|
| Time to first working system | Fastest | Fast | Slowest — hardware and installation |
| Data leaves your network | Yes | Yes, to the region | No |
| Verifiable by your own team | No — contractual assurance | Partly | Yes — inspect the network |
| Cost shape | Per-token, scales with use | Per-token | Capital up front, flat after |
| Ceiling on model size | Very high | High | Bounded by your hardware |
| Who is liable when it is wrong | Shared, per contract | Shared, per contract | You |
No row makes one shape correct. The rows that matter depend on what data you are putting through it.
For a public-facing marketing chatbot answering questions about opening hours, the first column is entirely reasonable. For a bank answering account questions, or a hospital handling patient records, the third column is not paranoia — it is the only shape where the residency claim is verifiable by the people who are accountable for it.
What the national layer does not give you
National compute and a national model solve the two layers that were previously hardest to source domestically. They leave four problems entirely with you, and these are the ones that determine whether the system works.
Relative weight of common failure causes in enterprise assistant deployments, as we encounter them. Directional, not a survey.
An assistant that reaches confidently for the wrong policy document is worse than no assistant, and no amount of national compute prevents it. Retrieval, permissions and escalation are application-layer engineering, and they are where the work is.
What to actually ask a vendor
- 1Does answering require an external model call?Ask for a yes or no. Then ask them to show you, on a network diagram or a firewall rule.
- 2Which model, and can I change it?A vendor locked to one model is selling you that vendor’s dependency as well as their own.
- 3Where do the documents live?Distinct from where the model lives. Retrieval indexes are a copy of your content.
- 4What happens when it does not know?Silent guessing and clean escalation look identical in a demo and nothing alike in production.
- 5Can I read the log for a single answer?If you cannot reconstruct why one answer was given, you cannot investigate a complaint.
- 6What is the exit?Who holds the index, the fine-tunes and the conversation history if you leave.
A worked example of the distinction
Consider two Saudi banks, both of which will tell a regulator they run “sovereign AI”, and both of which are telling the truth as they understand it.
| Bank A | Bank B | |
|---|---|---|
| Model | ALLaM | An open-weight model |
| Where it runs | A foreign cloud’s in-Kingdom region | Servers in the bank’s own data centre |
| Answering a customer question | A call to the cloud endpoint | No call leaves the bank’s network |
| How residency is evidenced | Contract and region attestation | The bank’s own firewall logs |
| If the provider changes terms | Renegotiate or migrate | No effect |
| If the network to the provider drops | The assistant stops answering | The assistant keeps answering |
Both are defensible. They are not the same commitment, and they fail in different ways.
Bank A gains a shorter path to production and access to models larger than it could host. Bank B gains verifiability and independence from a third party’s availability and pricing. The mistake is not choosing either one. The mistake is believing you chose the other.
Honest limits of this article
The figures for HUMAIN’s mandate, Hexagon’s capacity and the 2025 investment total are published national figures. We have not independently audited them, and readers making commitments on their basis should verify current numbers directly with SDAIA or HUMAIN.
We have no measurement of our own comparing ALLaM to other Arabic-capable models, and we do not publish comparative benchmark claims we have not run. Where we describe ALLaM’s Arabic strength we are reporting the published position, not a test result.
What we can state precisely is our own architecture, because we operate it. Elbi answers from your material, on hardware you control inside the Kingdom, with no external model call in the path of an answer. That is a description of where our system sits in the table above — the third column — not a claim that the third column is right for everyone.
Common questions
HUMAIN is a Public Investment Fund company launched in May 2025 and positioned as Saudi Arabia’s national AI champion. Its public mandate covers roughly 500 megawatts of compute capacity, supported by chip-supply arrangements with NVIDIA and AMD, and SDAIA has described it as the production and distribution channel for advanced Saudi models.
ALLaM is the Saudi Arabic large language model developed under the national AI programme, so the weights are sovereign in origin. That does not by itself keep your data in the Kingdom. ALLaM is distributed through commercial cloud platforms including Microsoft Azure AI and IBM watsonx, and a request to a foreign cloud endpoint leaves your network regardless of where the model was trained. Data residency is determined by deployment architecture, not by the nationality of the model.
Hosting describes where the application runs. It does not describe whether producing an answer requires a call to a model hosted elsewhere. A system can be hosted in Riyadh and still send every question to an external API. The precise question to ask a vendor is whether answering a single customer question requires a network call to a model they do not run.
Usually no. National capacity at 480 or 500 megawatts is sized for training foundation models and serving very large models at national scale. A retrieval assistant answering from a company’s own documents typically runs on a single server. Capacity is rarely the constraint on an enterprise deployment; retrieval quality, permissions and escalation are.
The application layer. Retrieval that finds the right document, permissions that control who sees which answer, escalation when the system does not know, and logging that lets you reconstruct why a specific answer was given are all engineering above the model. These are where enterprise assistant deployments most often fail, and no amount of national compute addresses them.
Keep reading
Saudi Arabia’s Year of AI 2026 explained
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.
Read Article
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
ChatGPT vs a custom AI chatbot
The honest answer is that they are not competing products. One is a brilliant general tool for your staff; the other is a narrow, accountable system for your customers. Choosing wrongly costs money in both directions.
Read Article