Last updated 13 October 2026
A held shipment of ordinary servers costs a week. A held GPU cluster can cost the go-live date, the installation crew’s time and the launch it was bought for. That gap is what makes AI data centre hardware such a hard IOR problem in IT trade this year.
The volume is real. The Loadstar reported on 26 May 2026 that AI infrastructure is becoming one of the logistics industry’s most important new cargo verticals, with some AI server units worth between 1 and 3 million US dollars. Each of those units crosses at least one border, and every crossing needs an importer of record for AI servers and GPU clusters who has done the work before the goods leave.
Here are the five reasons it is different, and what to check before you book.
Five Reasons AI Hardware Imports Are Uniquely Complex
Reason 1: US Export Controls Apply Before the Hardware Leaves
In short, a standard server only has to satisfy the destination. Advanced AI hardware has to clear the origin first.
Advanced accelerators and the servers built around them fall under ECCN 3A090 and 4A090 in the US Export Administration Regulations. The 2025 AI Diffusion Rule is not the framework anymore. BIS said in May 2025 that it would not enforce the rule’s new compliance requirements.
Licence requirements still bite, though. BIS guidance of 31 May 2026 confirms a licence is needed to export these items to entities headquartered in Country Group D:5 or Macau, or owned from there, wherever in the world they sit. Some destinations carry their own rules too. A July 2026 rule moved the UAE into Country Group A:5 but kept a licence requirement for advanced computing items, except for entities BIS has approved.
The exporter holds the licence, not the IOR. But an importer who accepts controlled hardware without knowing its export status inherits the problem at the destination. Our guides to AI GPU import compliance, export controls on data centre hardware and importing NVIDIA H200 servers go further.
Reason 2: Air Freight Leaves No Time to Fix a Mistake
In short, the paperwork has to be right before the aircraft is loaded, not before it lands.
GPU clusters are expensive to leave idle, so they fly. The Loadstar noted that many DHL deployments relied on charter aircraft for time-critical moves. In our own work, about 95% of what we move goes by air, most of it IT hardware.
Air takes away the slack. Under the EU’s ICS2, a minimum dataset has to be filed before the goods are loaded onto the aircraft. Customs can answer with a do not load notice, and the full entry summary declaration follows before arrival. A WEEE registration still pending, or a description customs rejects, stops the shipment at the origin airport or holds it at the destination airport while the deployment waits. See our ICS2 stop words guide for the descriptions that trigger it.
Reason 3: One Deployment Means Several Countries at Once
In short, the same server faces a different rulebook in every destination, often in the same week.
An enterprise hardware order often lands in one country. An AI build for a region usually lands in three or four in parallel. Here is how the import side can differ for the same rack.
- Germany. CE marking, WEEE registration with Stiftung EAR, packaging registration in LUCID, import VAT and an ATLAS declaration through a licensed broker. See importing IT equipment into Germany without a local entity.
- India. Servers under HSN 8471 sit inside DGFT’s Import Management System for 2026, with applications closing on 15 December 2026. BIS registration is checked per model. See how the DGFT import authorisation works.
- Saudi Arabia. Product conformity through SABER, a shipment certificate for each consignment, CST type approval where wireless modules are present, and 15% VAT at the border. See importing IT equipment into Saudi Arabia without a local entity.
Classification sits under all of it, so check the HS code for servers and the HS code for GPU chips before anything else.
Reason 4: The Go-Live Date Does Not Move
In short, a customs hold on AI hardware is not a delay. It is a missed commitment.
When ordinary servers are held for two weeks, the IT team waits and the business absorbs it. An AI cluster is different. The power is on, the cooling runs and the install team is on site, all against a go-live date that is often tied to customer contracts or a public launch.
A hold then costs storage at the airport, the crew’s idle days, any late-delivery penalty and the revenue the service was meant to start earning. Our guide to a customs hold on servers covers what to do when it happens.
Reason 5: Each Country Needs a Confirmed Importing Structure
In short, the question is not whether a provider has a partner or an entity. It is whether the structure is confirmed before the booking.
No provider owns an entity in every market, and a well run local structure can carry AI hardware perfectly well. The risk is a provider who accepts the booking first and finds out later that the local importer is not registered, not eligible or not approved for that product.
Ask who will be named on the entry in each country, which registrations that party already holds, and what still has to be applied for. Our guide to a paper IOR versus an operational IOR shows how to tell the difference.
Summary: Why AI Hardware IOR Is Different From Standard IT Imports
| Factor | Standard IT hardware | AI data centre hardware |
|---|---|---|
| Origin export controls | Rarely a licence question | ECCN 3A090 and 4A090, licence depends on destination and end user |
| Mode and correction time | Often flexible | Air, with data filed before loading |
| Geography | Often one country | Several countries in parallel |
| Cost of a hold | Storage and delay | Missed go-live, idle crew, penalties |
| What to confirm first | Classification and duty | Export status, approvals and the named importer per country |
Before the air waybill is issued, check five things per country. The export classification and any licence. The destination approvals for that model. The HS code. The party named as importer and its registrations. The ICS2 or local advance data, filed and accepted.
How Carra Globe Handles AI Hardware Deployments
Since 2024 we have run 64 data centre and server deployments into 35 countries, with AI and GPU hardware alongside servers, networking, storage and power equipment. The largest ran to 27 consignments over 14 weeks, and the highest declared value on a single consignment was about 1.8 million US dollars. See our data center solutions for the full scope.
We act as importer of record and, for retired kit going back out, exporter of record. Carra Globe works through importer-side registrations and local importing structures appropriate to the goods and destinations it accepts, its own in some markets and established local structures in others. Approvals falling outside the available structure are flagged before a delivery date is agreed.
For staged rollouts we hold stock through warehouse logistics and finish with white glove delivery to the hall. Power and cooling plant follows its own route, covered in importer of record for data centre equipment. Country detail sits in Germany, India, Saudi Arabia and the UAE. Reach us at info@carraglobe.com.
Frequently Asked Questions
What makes AI hardware harder to import than standard IT equipment?
US export controls apply before departure, air freight needs data filed before loading, builds run in several countries at once, and the go-live date rarely moves, so a hold costs far more.
Is the AI Diffusion Rule still in force?
No. BIS said in May 2025 it would not enforce it. Licence requirements for advanced AI chips still apply, depending on the destination and on who ultimately owns the end user.
Why does AI hardware usually move by air?
Because an idle GPU cluster is expensive. Weeks at sea delay the revenue it was bought to earn, so most time-critical AI deployments fly, sometimes on charter aircraft.
Can one IOR cover an AI deployment across several countries?
Yes, if the importing structure in each country is confirmed before booking. That can be the provider’s own entity or an established local structure, as long as it holds the registrations the goods need.
Does the importer carry any export control risk?
Indirectly. The exporter holds the licence, but end-use and end-user restrictions can follow the hardware after clearance, so the importer should know the export status before accepting it.
Adjacent ground: shipping AI servers internationally, AI server import duties, multi-cloud AI infrastructure compliance and exporting retired AI hardware.
Sources and verification
- Advanced computing licence requirement. BIS guidance of 31 May 2026 on entities headquartered in Country Group D:5 or Macau, which also records the May 2025 non-enforcement of the AI Diffusion Rule.
- UAE treatment. Enhanced Favorable Treatment for the United Arab Emirates, BIS final rule effective 10 July 2026, 91 FR 43034.
- ICS2 air cargo. European Commission, Import Control System 2, on pre-loading data and the complete entry summary declaration.
- India. DGFT Policy Circular No. 08/2025-26 of 17 December 2025 on import authorisation for HSN 8471 goods in 2026.
- Market context. The Loadstar, 26 May 2026, on AI infrastructure as a new cargo vertical.
- Carra Globe figures. Our own deployment records, 2024 to 2026.
Disclaimer: This guide is for informational purposes only and does not constitute legal, customs or export control advice. Export control and import requirements depend on the product, its classification, the destination and the parties involved, and can change at short notice. The position described is as at 13 October 2026. Confirm the current position with the competent authority or a qualified adviser before acting.