Stargate UAE and HUMAIN Explained: The Gulf's AI Build-Out

What are Stargate UAE and HUMAIN? Stargate UAE is a planned AI computing cluster in Abu Dhabi, announced in May 2025 by G42, OpenAI, Oracle, NVIDIA, Cisco and SoftBank Group. HUMAIN is Saudi Arabia's state-owned AI company, launched the same month. Both are announced plans for AI infrastructure, and both depend on imported chips.

Updated 25 October 2026. This is a plain-language reading of the announcements by CyphrQ.ai, with links to the sources I could open. It is not investment advice.

Why are the UAE and Saudi Arabia building AI infrastructure?

Training and running modern AI models takes very large amounts of computing power, housed in data centres. Whoever hosts that capacity gets the jobs, the tax base and a say in how the technology is governed. Both governments want it hosted locally rather than rented from abroad.

That is the logic behind the word "sovereign" in the announcements. In this context it means a country wants AI capacity it can regulate and rely on, not that it builds everything itself. As you will see below, both projects lean heavily on US chips and US companies.

One caution before the details. An announcement is a statement of intent. I label each item below as announced, in progress or confirmed, and I treat nothing as finished unless a source says so.

What is Stargate UAE, and what was announced?

G42, the Abu Dhabi AI company, published the launch on 22 May 2025. According to G42's announcement, the plan has these parts:

Item	What was announced
Cluster	A 1-gigawatt Stargate UAE compute cluster, built by G42
Operators	OpenAI and Oracle
Chips	NVIDIA's GB300 Grace Blackwell systems
Security and networking	Cisco, described as "zero-trust security and AI-ready connectivity"
Other partner	SoftBank Group, named as a collaborator with no role described
Campus	The UAE-U.S. AI Campus in Abu Dhabi, which "will provide 5 gigawatts of AI data center capacity and regional compute resources"
First phase	"The first 200-megawatt AI cluster is expected to go live in 2026"

A gigawatt (GW) is a unit of power, not computing. It describes how much electricity the site can draw, and it is the usual way to size a data centre. One GW is 1,000 megawatts (MW).

The same announcement says the campus will be powered by nuclear, solar and natural gas. That is the company's statement of intent, not a confirmed supply contract.

Is Stargate UAE finished?

I found no source confirming that any part of it is live. In an October 2025 update, a G42 spokesperson told Gulf News that "work is progressing rapidly on the first 200 megawatts of capacity, which is expected to be completed by 2026." The same article says the rest of the 1GW cluster and the wider 5GW campus have no confirmed completion timelines.

So as of this writing, 200MW in 2026 is a target, and the year is nearly over. Check the partners' own pages before you quote a status.

What is HUMAIN, and what has Saudi Arabia announced?

HUMAIN is a company owned by Saudi Arabia's Public Investment Fund (PIF), the state wealth fund. Crown Prince Mohammed bin Salman launched it on 12 May 2025, according to the PIF's press release and the Saudi Press Agency's coverage (PIF, SPA).

HUMAIN's stated scope is the whole AI stack in one company: data centres, AI infrastructure, cloud capabilities and AI models, including a large Arabic-language model. That differs from Stargate UAE, which is one campus built by a consortium. HUMAIN is a national company that will do many things.

I am deliberately not quoting HUMAIN's capacity targets or partner deal values in this post. Several figures circulate in the trade press, and I have not read them in a primary source. My rule is not to publish a number I have not read at the source. I found no source confirming that any HUMAIN data centre is operating.

How do the two projects differ?

Structure. Stargate UAE is a consortium of private companies around one campus. HUMAIN is a state-owned company with several business lines.
Role of OpenAI. In Stargate UAE, OpenAI is a named operator. HUMAIN's announcements are broader and do not centre on one model provider.
What they share. Both need advanced chips made in the US, and both rely on US approval to get them.

Can the UAE and Saudi Arabia buy the chips?

Both countries need US export approval for the most advanced AI chips. On 20 November 2025, G42 announced that the White House had authorised exports of advanced AI semiconductors to G42. G42 said deployment is governed by its Regulated Technology Environment, which it describes as approved under US Department of Commerce and Bureau of Industry and Security (BIS) guidelines.

That is the company's own account. The US Commerce Department also published a statement on UAE and Saudi chip exports. Its site blocked my access when I checked, so I don't summarise its conditions here. Read it at the source.

The practical point is that chip supply is a policy decision as well as a purchase. The rules can change, and G42 itself describes its approval as tied to a regulated environment. Any timeline for either project is only as firm as those approvals.

What does this mean for the Gulf's tech economy?

I will separate what is documented from what is my reading.

Documented. Two Gulf states have announced national-scale AI infrastructure, with US companies as partners and US approval as a condition. G42 describes its November 2025 approval as moving "from planning into execution," in the words of its Group CEO Peng Xiao.

My reading, which you can disagree with.

Demand will follow supply only if there are customers. A data centre is a fixed cost. The regional AI economy needs businesses and governments that actually use the capacity, not just build it.
Power is the limiting factor. A gigawatt-scale site needs a very large, steady electricity supply. For the wider picture, see our explainer on what the IEA says about AI data centre electricity.
Skills and security work lag buildings. Concrete is the quick part. Staffing, operating and securing a campus takes longer.
Data rules matter for customers. Saudi organisations, for example, must follow cross-border transfer rules. See our guide to what the Saudi PDPL requires.

What should businesses in the region do now?

You do not need a gigawatt campus to act on this. Three questions are enough:

Where does your AI vendor process and store your data? Ask for the regions in writing, and whether prompts and logs stay there.
What happens if local capacity arrives late? Keep an exit plan for any workload that depends on a not-yet-built facility.
Which law applies to the data you send? Check the UAE, Saudi or other rules that cover your customers, not just your own location.

Going deeper: for security and IT teams

This section covers what an architect or auditor can verify. I name only what the announcements actually state, and I flag what they leave out.

What the sources state about security. G42's launch announcement says Cisco provides "zero-trust security and AI-ready connectivity," and Oracle's Larry Ellison said the deployment "sets a new standard for digital sovereignty." Those are descriptions, not specifications. Neither names a control framework, a certification or an audit report. G42's chip-export announcement names one mechanism: the Regulated Technology Environment, deployed under BIS guidelines. The documents I could open do not publish its technical contents.

Trust boundaries in an AI campus. A large AI cluster has a few distinct layers, each with a different owner and a different attack surface:

Facility and power. Physical access, cooling and grid supply.
Compute and network fabric. The GPUs, the interconnect and the management network. Chips are export-controlled, so the operator may owe the exporter reporting and access controls.
Platform layer. The cloud or orchestration software run by an operator such as Oracle or HUMAIN.
Tenant layer. Your models, prompts, training data and logs.

When a press release says "sovereign," ask which layers it covers. Local hosting of the building does not by itself mean local control of the platform, the keys or the support staff.

Threat model. The credible threats are the ones documented for any cloud: tenant isolation failures, stolen credentials and tokens, misconfigured storage, supply-chain compromise, and insider access. AI adds model weights and training data as high-value targets. For real examples of exposed AI data, see our post on the Chat and Ask AI database exposure.

Data residency and sovereign cloud requirements. Residency is a legal and contractual property, not a building. Ask your provider:

Which country hosts your data at rest, backups and logs?
Who can administer the platform, and from where?
Who holds the encryption keys, and can you bring your own?
Which government can legally compel access, and under which law?
How are cross-border transfers handled for support and telemetry?

Saudi Arabia's PDPL has its own transfer rules, covered in our Saudi PDPL guide. Read the UAE position separately, as it differs between federal law and the DIFC and ADGM regimes (see UAE PDPL and AI tools).

Evidence to request. An ISO/IEC 27001 certificate with the statement of applicability and scope (see ISO/IEC 27001 explained), a SOC 2 Type II report if the provider has one, and a CSA STAR entry (see CSA Cloud Controls Matrix). Check that the scope names the specific region and service you will use. A certificate for a different site does not cover yours.

Export controls as officially stated. US approvals for advanced chips are conditional and can be revised. G42 says its approval rests on BIS guidelines. I could not open the Commerce Department statement, so read its conditions at the source. For your own risk register, treat chip supply as a dependency with an owner and a review date.

Detection and testing. For any AI platform you adopt, confirm that you can export admin and API audit logs to your own SIEM, alert on new API keys and unusual egress, and test tenant isolation through the provider's published process. Ask for the right to review the results of penetration tests.

Learn more

Key takeaways

Stargate UAE (1GW cluster in a 5GW campus) and HUMAIN are announced plans, and I found no source confirming either is operating.
The first 200MW of Stargate UAE is a 2026 target, per G42.
Both depend on US chip approvals, which can change.
"Sovereign" needs a definition each time: ask which layers are local and who holds the keys.
Before you adopt local AI capacity, get data location, admin access and certificate scope in writing.

Display and accessibility