Overview

The May 2025 announcement that NVIDIA would supply 18,000 GB300 (Blackwell) GPUs to Humain — Saudi Arabia’s PIF-backed AI company — marked a turning point in US-Saudi AI compute relations. Getting Blackwell-class hardware into Saudi Arabia requires navigating the US export control framework, which underwent a major restructuring in early 2025. This guide explains the current AI Diffusion framework, the BIS export license process, what the Humain precedent means for other entities, and practical steps for companies seeking to import or supply Nvidia GB200/GB300 chips into the Kingdom.

Understanding the AI Diffusion Framework (2025)

The Three-Tier System

In January 2025, the Biden administration finalized the AI Diffusion Rule, which replaced the ad-hoc case-by-case review process for AI chip exports with a three-tier country classification:

  • Tier 1 (18 countries): US allies with full access. Includes UK, Japan, South Korea, Netherlands, Australia, Canada. No license required for most AI chip exports. Examples: TSMC in Taiwan (Taiwan is Tier 1 equivalent under a carve-out).
  • Tier 2 (~120 countries including Saudi Arabia): Enhanced access with conditions. Entities can import AI chips under a license exception (called the “Validated End User” or VEU program) if they meet specific requirements, or they can apply for an individual export license.
  • Tier 3 (arms-embargoed countries and adversaries): Effectively prohibited. China, Russia, Iran are Tier 3 for advanced AI chips.

Saudi Arabia is Tier 2. This means:

  1. Government-to-government deals (like Humain/NVIDIA facilitated through the Trump-MBS summit) can move faster via diplomatic-track approvals
  2. Commercial entities not covered by a G2G framework still need to apply for export licenses
  3. The total compute that can be exported to Saudi Arabia is subject to national-level “compute caps” that BIS tracks annually

The Trump-MBS Summit Framework

The May 2025 state visit resulted in what effectively functions as a framework pre-approval for certain AI chip exports to Humain and affiliated entities. The mechanism was:

  1. US Department of Commerce and Saudi MCIT signed an AI cooperation agreement establishing joint cybersecurity and end-use monitoring protocols
  2. This agreement created the predicate for BIS to approve Humain as a Validated End User
  3. The NVIDIA/Humain 18,000 GB300 deal was the first transaction under this VEU designation

The VEU designation is not a blanket license for unlimited GPU imports — it covers specific quantities, end-use locations, and customer categories. Entities seeking to import AI chips to Saudi Arabia under a different structure (not Humain, not VEU) still need individual export licenses.

Step 1: Determine Your Export Control Classification

EAR99 vs. ECCN Classification

Not all AI hardware is equally controlled:

  • EAR99: No license required for most destinations (including Tier 2). Most standard server hardware, standard CPUs, networking equipment falls here. No license needed.
  • ECCN 3A090: The classification for advanced AI accelerators including NVIDIA H100, H200, A100, GB100, GB200, GB300. Requires license for Tier 2 countries unless covered by a license exception.
  • ECCN 4A090: Applies to systems containing multiple 3A090 chips (i.e., GPU servers). Same licensing requirement.

Run your hardware manifest through the SNAP-R screening tool (snapr.bis.doc.gov) before starting a license application to confirm classification. NVIDIA provides EAR classification data for all products in their export compliance portal at nvidia.com/en-us/about-nvidia/export-compliance.

Step 2: Identify the Applicable License Exception

License Exception APP (Applications)

Under the 2025 AI Diffusion Rule, Tier 2 countries can receive up to 1,700 H100-equivalent GPUs (approximately 26,240 petaflops of BF16 compute) per year per end-user entity under a streamlined notification process rather than a full license. This covers smaller deployments.

For the Blackwell GB300 NVL72 unit: each unit delivers approximately 1.44 exaflops BF16, which is 1.44 petaflops per unit (note: BIS uses a specific “Performance Density” metric under the rule — consult the Federal Register notice at 90 FR 4544 for the exact calculation methodology). An 18,000-unit deployment (like Humain’s) far exceeds the notification threshold and requires a full export license.

Validated End User (VEU) Program

If your Saudi entity can qualify as a VEU, you receive a multi-year, multi-transaction authorization that eliminates the need for individual licenses per shipment. VEU application requirements:

  1. Demonstrated history of US export law compliance
  2. Established cybersecurity program meeting NIST CSF standards
  3. Physical security measures at data center facilities (CCTV, access controls, audit logging)
  4. Agreement to on-site inspections by BIS personnel or authorized auditors
  5. No ownership or control by entities on the BIS Entity List or Denied Persons List

Humain qualified for VEU status partly because PIF itself is not on any restricted party list and because the G2G AI cooperation agreement provided BIS with assurance of end-use monitoring. For non-PIF entities, the VEU path requires demonstrating equivalent compliance infrastructure.

Step 3: File the Export License Application

SNAP-R System

All BIS export license applications are filed through SNAP-R (Simplified Network Application Process — Redesigned) at snapr.bis.doc.gov. Create an account and register your company as an exporter (US exporter) or as an in-country recipient working through a US exporter.

The license application for ECCN 3A090/4A090 hardware to Saudi Arabia requires:

Part A — Applicant Information:

  • US exporter’s EIN and BIS account number
  • Saudi end-user’s full legal name, address, and commercial registration number (CR number from MISA)

Part B — Transaction Details:

  • Full hardware manifest: part numbers, quantities, values
  • End-use statement: specific description of what the chips will be used for (“training large language models for commercial AI services in Saudi Arabia” is acceptable; vague statements like “general computing” are not)
  • Facility address where hardware will be installed

Part C — End-User Certificate (EUC): The EUC is the most important document. The Saudi end-user must sign a statement that:

  1. The hardware will be used only for the stated purpose
  2. The hardware will not be re-exported without separate BIS authorization
  3. BIS and authorized agents may inspect the hardware and end-use location
  4. The end-user will notify BIS of any change in ownership or control

The EUC must be signed by a senior officer (C-level or board member) of the Saudi receiving entity, notarized, and apostilled. MCIT’s certification stamp is recognized by BIS as equivalent to apostille for Saudi government entities.

Supporting Documentation:

  • Copy of the purchase contract between the US exporter (NVIDIA, Dell, HPE) and the Saudi entity
  • Facility operator’s Tier III or Tier IV certification (Uptime Institute or equivalent)
  • CST Cloud Service Provider license (if the hardware will be used in a commercial cloud service)
  • Organizational chart showing beneficial ownership up to ultimate beneficial owners (UBOs)

Step 4: Processing Timeline and What Affects It

The statutory processing time for BIS export license applications is 90 days, but actual times vary:

  • Straightforward Tier 2 commercial deal (established entity, clear end-use, no red flags): 60–90 days
  • New applicant without prior BIS relationship: 90–150 days
  • Application requiring inter-agency referral (State Department, DoD, or NSA review): 120–180 days
  • G2G facilitated deals (like Humain/NVIDIA under the AI cooperation agreement): 30–60 days via expedited diplomatic track

Applications requiring inter-agency referral are triggered by: large quantities (above ~3,400 H100-equivalent chips), proximity of the end-user to Chinese-controlled entities, any prior BIS violation history, or applications from entities in sectors BIS designates as sensitive (nuclear, ballistic missiles). Most pure-commercial AI compute deals avoid referral.

Check application status in SNAP-R. If your application has been “Under Referral” for more than 60 days, engage an export controls attorney to request a status update from BIS’s Western Region export licensing office.

Step 5: Major Non-NATO Ally Status — What It Means

Saudi Arabia was designated a Major Non-NATO Ally (MNNA) by the United States in 2023. MNNA status affects export licensing in specific ways:

  1. Defense articles and dual-use items: MNNA status streamlines licensing for items under the International Traffic in Arms Regulations (ITAR). Most commercial AI chips are EAR-controlled (not ITAR), so MNNA status has less direct effect here.
  2. Pre-positioning of defense equipment: MNNA allows Saudi Arabia to pre-position US military equipment, which is irrelevant to commercial AI compute.
  3. Political signal for BIS reviewers: MNNA status is a factor BIS considers in weighing national security risk. It does not create a presumption of approval, but it is a favorable marker that reviewers note.

The more significant designation for AI compute is the AI cooperation agreement from May 2025, which established the specific oversight framework that allowed VEU status for Humain.

Alternative Paths for Non-Humain Entities

If your entity is a Saudi enterprise or a non-Humain Saudi AI company seeking Blackwell hardware:

  1. Buy through a Tier 1 intermediary: Route the purchase through a Tier 1 country entity (e.g., a UK or Singapore subsidiary that holds the hardware on a lease/managed services basis). The chips remain in the Tier 1 entity’s asset register; you consume the compute via a managed service or cloud API. This is legally permissible if the UK/Singapore entity genuinely controls the hardware and your access is software-only.

  2. Colocate in a VEU-licensed facility: If Humain or another VEU-designated Saudi operator offers colocation, your hardware deployment may fall under their VEU umbrella. This is a grey area requiring BIS confirmation — request a binding commodity jurisdiction or advisory opinion before proceeding.

  3. Use AMD or other alternatives: AMD MI300X and the upcoming MI325X are also ECCN 3A090 and require the same export licenses. There is no regulatory shortcut through alternative vendors, but AMD’s less saturated licensing queue may mean faster processing times in practice.

  4. Engage a Freight Forwarder with Export Compliance Specialty: Companies like Flexport, Expeditors International, or Kuehne+Nagel have dedicated export compliance teams with BIS relationships. They can co-file applications, flag issues before submission, and manage the Saudi customs clearance process (which requires advance submission of the BIS license number to Saudi Customs via the Fasah single window platform at fasah.sa).

The Humain Precedent in Numbers

The Humain pipeline is worth studying closely because it shows how Blackwell-scale approvals actually sequence. The May 2025 framework covered 18,000 GB300 units as the initial tranche of a 600,000-GPU, three-year pipeline supporting up to 500 MW of compute. The BIS approval that operationalized the next phase arrived in November 2025, coinciding with the Crown Prince’s Washington visit, and covered 35,000 GB300 units — under 6% of the announced pipeline. A sub-allocation of up to 5,000 Blackwell GPUs flows to SDAIA for its sovereign AI factory in Riyadh, which moved to deployment on its own schedule.

Three lessons for other applicants. First, approvals arrive in tranches sized to facility readiness, not as blanket authorizations for announced totals. Humain Campus Riyadh (200 MW, sized for the initial 18,000 units, targeting 2026) defines the deployment capacity that justified the approved volume, with the 300 MW Dammam campus (2027) staging the next tranche. Phase your license applications to your construction milestones; an application for volumes your facility cannot energize invites either denial or an approval sized down to what the facility supports. Second, sub-allocations travel under the primary VEU umbrella — the SDAIA tranche is part of the Humain-NVIDIA envelope, which is the working model for the colocation path described above. Third, political-calendar alignment is real: both the May 2025 framework and the November 2025 approval landed inside bilateral political moments, and applications timed to mature during periods of active US-Saudi engagement have historically moved faster than applications filed into quiet periods.

What Licensing Status Does to Your Price and Schedule

Licensing tier is not compliance overhead; it is the largest single commercial variable in Saudi Blackwell procurement. VEU-covered sovereign-anchor buyers land B200-class silicon at $28,000-36,000 per GPU on 2-4 month procurement cycles. Non-VEU commercial buyers pay $36,000-48,000 per GPU on 6-9 month cycles that include specific-license processing. GB200 NVL72 rack-scale systems land at $2.8M-3.6M per rack at the sovereign-anchor tier. Saudi-landed pricing overall carries a 6-14% premium over US-domestic equivalents at matched volume, reflecting licensing overhead, logistics, and local-content layers.

Run the arithmetic before choosing your path: on a 2,000-GPU B200-class procurement, the spread between VEU-tier and non-VEU pricing is $16M-24M — several multiples of the entire licensing-process cost, including counsel, the compliance build-out, and the VEU application itself. For any entity planning more than one procurement cycle, the VEU path (or colocation under an existing VEU umbrella) is the economically dominant strategy even before counting the two quarters of schedule it returns.

Pre-Application Checklist

Before filing in SNAP-R, confirm: hardware classification verified against NVIDIA’s export-compliance data (ECCN 3A090/4A090 confirmed, not assumed); the end-user’s CR number, UBO chart, and facility address finalized — applications naming facilities that do not yet exist stall; the EUC signatory genuinely C-level or board, with notarization and apostille (or MCIT certification for government entities) arranged in advance; the end-use statement drafted at the right specificity — training large language models for commercial AI services passes, general computing does not; facility Tier certification and CST cloud license current where applicable; quantity below the roughly 3,400 H100-equivalent inter-agency referral trigger, or the referral timeline priced into the project plan; and the purchase contract structured with export-license contingency clauses so that no non-refundable deposit rides on a pending application.

Common Application Failure Modes

The recurring reasons Saudi-destination applications stall or die: end-use statements written by sales teams rather than compliance counsel; EUC signature authority pushed down to a project director BIS does not recognize as senior; quantity requests sized to ambition rather than to energizable facility capacity; missing the referral-trigger analysis and being surprised by a 120-180 day inter-agency review; assuming MNNA status or the bilateral framework creates a presumption of approval (it is a favorable marker, not a shortcut); and treating the application as a one-shot filing rather than a managed process — RAI-style clarification responses that take three weeks instead of three days add a month each to the timeline. The applications that clear in 60-90 days are those where the exporter’s compliance team, the Saudi end-user’s counsel, and the freight forwarder’s export desk worked from a shared checklist before submission.

For deeper reading: How to source GPUs under BIS · How US export controls affect Saudi · How to set up a Saudi cloud region · Silicon tracker.