On 22 October 2026, at Nvidia's GTC conference in Berlin, HUMAIN, the AI company owned by Saudi Arabia's Public Investment Fund, announced an expanded Nvidia collaboration of up to 84MW of AI compute, roughly 27,000 GPUs, according to The National. The headline number matters less than its structure. The deal shows that the Kingdom's compute buildout under Vision 2030 depends on two things Riyadh does not fully control: US export approvals, and a data-sovereignty framework that is still in draft.
What was actually announced
Phase one is 14MW of Nvidia Blackwell HGX B300 systems, about 6,000 GPUs, starting in the fourth quarter of 2026 and serving HUMAIN Cloud customers. Phase two is 70MW of Vera Rubin capacity, about 21,000 GPUs, from mid-2027. Of that, 50MW is earmarked for an unnamed US client and 20MW for HUMAIN Cloud, per Middle East AI News. The companies said the Vera Rubin projects remain subject to final agreements and deployment plans.
Two details stand out. Only the 14MW first phase is on a near-term schedule. And 50 of the 84 planned megawatts, nearly 60%, would serve a single customer whose name has not been disclosed. The same report puts HUMAIN's longer-term goal at about 3GW of data-centre capacity by 2030, so this announcement is roughly 3% of that target.
The case for tight US control
The strongest argument for export conditions is simple: frontier accelerators are dual-use, and Washington wants assurance that they are not diverted to adversaries or used to train systems that undermine US security. Conditional licensing, with reporting and compliance monitoring, is a defensible way to keep that assurance while still letting allies buy.
That is roughly what the US Commerce Department did on 19 November 2025. It announced authorisations for both Saudi Arabia's HUMAIN and the UAE's G42, each allowed to buy the equivalent of up to 35,000 Nvidia GB300 Blackwell chips, conditioned on security and reporting requirements that the Bureau of Industry and Security will monitor on an ongoing basis. The Commerce statement tied the decision to the July 2025 AI Action Plan and the US-Saudi AI partnership agreements. (We could not open the commerce.gov page directly; the terms above match the statement as reproduced by Middle East AI News.)
Where the model strains
The trouble is that approval is per-company and per-tranche, while the buildout is planned in multi-year phases. The National notes that Saudi Arabia and the UAE depend on US export approvals to obtain advanced AI chips, and the coverage we reviewed does not say whether the Vera Rubin generation, or capacity for a US client, has been licensed. Investors and customers therefore price in regulatory risk that no Saudi statute can remove.
A proportionate regime would give allied buyers predictable, published criteria and licence terms that cover a product generation, not a single purchase. Opaque case-by-case approvals reward whichever relationships are strongest in a given quarter and punish planning. That is bad for American suppliers too, who compete with vendors not subject to the same uncertainty.
The sovereignty question nobody has answered
Hosting a US client's workloads in the Kingdom raises a second question: whose law governs the data and the compute? Saudi Arabia has one candidate answer in draft form. The Communications, Space and Technology Commission issued a draft Global AI Hub Law on 14 April 2025, with consultation open until 14 May 2025. As summarised by CMS, it creates "data embassies": private hubs run by a foreign government, extended hubs run by an approved third-party operator, and virtual hubs run by a Saudi-incorporated provider for customers of a designated foreign state, governed by that state's laws.
That design is promising for a deal like this one. A US client could keep its own legal regime, which is the kind of certainty that attracts workloads. But CMS describes the law as a draft, notes that the Council of Ministers must still designate the competent authority, and says it is unclear whether that will be CST, SDAIA or another body. It also records that Saudi authorities keep the right to intervene where national interests are threatened. A customer weighing 50MW of capacity will read that carve-out closely.
We have not verified whether the unnamed US client would be covered by any hub arrangement, and the announcement does not say. That gap is the point. SDAIA's national AI strategy can set ambitions, but ambition is not the same as a rule an enterprise can sign a contract against.
What proportionate policy looks like
For Riyadh, the pro-innovation move is to finalise the hub framework, name the regulator, and publish narrow, reviewable criteria for the national-interest exception, so that hosting is a legal product and not a negotiated favour. For Washington, it is to turn conditional, tranche-based approvals into stable, published terms for allies that meet security requirements. Neither requires heavy new regulation. Both require clarity.
The risk of getting this wrong is not abstract. If the Vera Rubin phase slips, or the US client's data regime stays undefined, the Kingdom will have built megawatts without the legal infrastructure that makes them valuable. Compute is the easy part; predictable rules are what turn it into an AI economy.