Multi-Cloud AI Infrastructure Import Compliance: The Complete Guide for AWS, Azure and Google Cloud Deployments

On March 16, 2026, AWS announced at NVIDIA GTC that starting in 2026 it will deploy more than one million NVIDIA GPUs, including Blackwell and Rubin architectures, across its global cloud regions. Reuters confirmed on March 19, 2026, that NVIDIA plans to deliver this quantity to Amazon’s cloud division by end of 2027. Those one million GPUs do not materialise inside AWS data centres by satellite. Every single server enters each country through customs. Every single entry requires a licensed importing entity, country-specific product certifications, BIS export control compliance, and an import declaration filed by an authorised party. AWS is one cloud provider. Add Azure and Google Cloud simultaneously expanding GPU capacity in the same global regions and the scale of the multi-cloud AI infrastructure import compliance challenge becomes clear. This guide covers what that compliance challenge actually involves, country by country, certification by certification, for the organisations deploying AI infrastructure across multiple cloud providers and jurisdictions simultaneously.

Why Multi-Cloud AI Infrastructure Creates a Unique Import Compliance Problem

Single-cloud AI deployments are complex. Multi-cloud deployments multiply that complexity across every dimension simultaneously. Consider the architecture real enterprises have already built:

• Adobe Creative Cloud: GPU infrastructure distributed across AWS, Azure, and Adobe’s own data centres globally according to industry reporting. AWS handles model training, Azure handles inference workloads, private facilities handle proprietary data. Three cloud environments, each with GPU hardware in multiple countries, each country a separate import event
• Airbnb: GPU workloads distributed across AWS, Azure, and Google Cloud simultaneously across multiple regions worldwide according to Airbnb engineering documentation. Real-time job routing to whichever cloud has cheapest available capacity. Multiple regions means multiple import jurisdictions
• Spotify: ML infrastructure distributed globally across GCP as primary compute, AWS as burst capacity, and Azure for European workloads to comply with EU data residency requirements, according to Spotify engineering blog disclosures

The import compliance implication of these architectures is the one nobody has written about. Every physical GPU server that enters Germany for an AWS Frankfurt region expansion is an import event requiring CE marking, WEEE registration, ICS2 advance declaration, and an EORI-registered importer. Every server entering India for an AWS Mumbai or Azure Central India region deployment requires BIS CRS certification obtained 3-6 months in advance, IEC registration, and an ICEGATE-filing importing entity. The cloud provider manages the infrastructure once it is operational. The physical import into each country requires a separate, locally compliant Importer of Record structure that most enterprise cloud procurement teams have never had to think about before.

The Three-Layer Multi-Cloud AI Infrastructure Import Compliance Framework Every Operator Needs

Every physical GPU server deployment across any cloud provider in any country requires compliance across three simultaneous layers. The layers do not interact cleanly. Each must be managed independently and in parallel:

• Layer 1: US BIS Export Controls. NVIDIA H200, Blackwell B100, B200, GB200, and Rubin R100 GPUs are classified under ECCN 3A090 under the Export Administration Regulations. Our AI GPU import compliance 2026 guide covers the full BIS three-tier country framework and enforcement landscape in detail. Every export from the United States to a foreign cloud region requires BIS compliance: licence exception verification, denied party screening, Electronic Export Information filing, and end-use documentation. The cloud provider’s procurement team handles this at the point of ordering from NVIDIA. The enterprise customer deploying within that cloud region carries ongoing compliance obligations for the hardware in their tenancy under current BIS policy and the Chip Security Act 2026 framework moving through Congress
• Layer 2: Destination Country Import Compliance. Every country where a cloud region physically exists has its own import requirements: customs declaration by a locally registered entity, applicable duty and tax payment, and product certifications for the specific hardware categories being imported. These requirements exist regardless of whether the cloud provider or the enterprise customer is the named importer. Somebody must be the Importer of Record in every country
• Layer 3: Ongoing End-Use Documentation. Under current BIS export control policy and the Chip Security Act framework moving through Congress, the end-user of controlled GPU hardware carries documentation obligations extending beyond the import event: access logging, location verification, chain of custody records, and reporting obligations if hardware is diverted or tampered with. In a multi-cloud architecture, the enterprise customer is typically the end-user of record even when the hardware is owned and operated by the cloud provider

Multi-Cloud AI Infrastructure Import Compliance Country by Country

AWS GPU-capable regions include Australia, Dubai (UAE), Germany, India, Japan, Singapore, the United States, and the United Kingdom. Azure and Google Cloud operate overlapping region footprints across many of the same jurisdictions. Here is the full import compliance picture for each major multi-cloud AI deployment market:

India: Highest Multi-Cloud AI Infrastructure Import Compliance Risk

India has three active cloud regions across AWS, Azure, and Google Cloud, making it the largest multi-cloud AI deployment market in South Asia. Every physical GPU server entering India faces the same certification requirements regardless of which cloud provider owns the hardware:

• BIS CRS certification: Mandatory for all servers and IT hardware under the Bureau of Indian Standards Compulsory Registration Scheme. Certification takes 3-6 months minimum with no expedited pathway. Hardware arriving at JNPT or Chennai without valid BIS registration cannot be cleared. This applies to every NVIDIA Blackwell or H200 server entering India regardless of whether it is an AWS, Azure, or Google Cloud procurement
• WPC/TEC type approval: Mandatory for all wireless-enabled components including server management interfaces
• ICEGATE filing: Import declaration through India’s customs portal requires an IEC-registered importing entity. Foreign cloud providers use locally registered entities or specialist IOR providers
• BIS export control context: India operates under current BIS export control policy with quantity considerations on advanced GPU imports. AWS confirmed its India region GPU expansion in its official GTC 2026 announcement. Large-scale Blackwell deployments across multiple cloud providers in India simultaneously may interact with national allocation considerations under current BIS policy
• Demurrage risk: USD 40-90 per container per day at JNPT from day one of any hold. A multi-cloud deployment activating three cloud regions simultaneously without pre-confirmed BIS registration multiplies this risk across all three procurement streams at once

Germany and the EU: Multi-Cloud AI Infrastructure Import Compliance and Regulatory Depth

Germany hosts AWS Frankfurt, Azure Germany West Central, and Google Cloud Frankfurt region, making it the primary EU multi-cloud AI hub. The EU regulatory framework applies consistently across all three cloud providers importing hardware into German facilities:

• CE marking: Mandatory for all electrical and electronic equipment placed on the EU market. NVIDIA Blackwell and H200 servers must carry a valid CE Declaration of Conformity covering the Radio Equipment Directive, EMC Directive, and Low Voltage Directive before any server enters an EU data centre for commercial use
• WEEE registration: Stiftung EAR WEEE registration required for electrical equipment placed on the German market commercially. Applies regardless of whether the equipment is owned by the cloud provider or the enterprise customer
• LUCID packaging registration: Required for all packaged goods entering the German market commercially
• ICS2 Entry Summary Declaration: Mandatory for air freight before loading at origin. A failed ICS2 validation generates a do-not-load instruction with no override. For time-critical Blackwell GPU deployments shipped by air, ICS2 compliance at origin is the critical path
• EORI registration: The importing entity must hold an active EU EORI number. ATLAS customs system filing required for all commercial imports
• CBAM consideration: EU Carbon Border Adjustment Mechanism applies to certain industrial imports. While AI server hardware is not currently in scope, data centre cooling infrastructure and power equipment imports may warrant CBAM review under the proposed scope expansion to industrial derivatives

Singapore: Multi-Cloud AI Infrastructure Import Compliance and BIS Scrutiny

Singapore hosts AWS Asia Pacific (Singapore), Azure Southeast Asia, and Google Cloud Singapore region. It is simultaneously the most logistics-efficient multi-cloud import market in Southeast Asia and the one facing the most explicit BIS diversion scrutiny:

• IMDA type approval: Infocomm Media Development Authority approval required for all wireless-enabled hardware. Singapore is one of the fastest IMDA approval markets in Asia, typically completing in days to weeks rather than months
• TradeNet UEN filing: All imports require TradeNet permit declaration through a UEN-registered Singapore entity. Zero customs duty on server hardware under ITA. 9% GST applies on CIF value, reclaimable by GST-registered businesses
• BIS diversion scrutiny: Singapore has been named in US government communications regarding concerns about third-country access to controlled chips enabling diversion to restricted end users. Multi-cloud operators deploying NVIDIA Blackwell hardware in Singapore must maintain comprehensive KYC documentation, access logs, and end-use records that can withstand BIS verification under current export control policy and the Chip Security Act framework
• Major Exporter Scheme: For cloud providers importing server hardware primarily for regional distribution and subsequent deployment, Singapore’s MES eliminates GST on qualifying import flows

UAE: The Sovereign AI Multi-Cloud Deployment Market

The UAE hosts AWS Middle East (UAE) and Azure UAE North regions, with significant sovereign AI deployments through G42 and other national AI programmes. The November 2025 US Commerce Department approval of GPU shipments to UAE G42 established a framework for large-scale Blackwell deployments in the UAE, but the import compliance requirements remain fully in force:

• TRA/TDRA type approval: Telecommunications and Digital Government Regulatory Authority approval required for all wireless-enabled hardware. Pre-arrival requirement with no retrospective application. TDRA approval must be confirmed before freight departs origin
• ESMA conformity: Emirates Authority for Standardisation and Metrology requirements apply to electrical equipment
• Import duty: 5% ad valorem on CIF value for most IT hardware. 5% VAT applies separately
• BIS entity-specific approval: The November 2025 US approvals were entity-specific authorisations to G42 and Humain, not blanket market-wide permissions. Enterprise customers deploying within UAE cloud regions through tenancy arrangements must confirm their own compliance status under current BIS policy

Japan: Tier 1 BIS Market, High Certification Depth

Japan hosts AWS Asia Pacific (Tokyo and Osaka), Azure Japan East, and Google Cloud Japan regions. Japan is a Tier 1 country under current BIS export control policy, meaning no BIS licence is required for standard commercial end uses. The certification depth is still significant:

• MIC type approval: Ministry of Internal Affairs and Communications type approval mandatory for all radio equipment and wireless-enabled hardware
• PSE certification: Product Safety Electrical Appliance and Material certification for electrical equipment
• NACCS filing: Nippon Automated Cargo and Port Consolidated System filing through a Japanese-registered entity. Zero customs duty on server hardware under ITA. 10% consumption tax at import

United Kingdom: Post-Brexit Certification Transition

The UK hosts AWS EU West (London), Azure UK South, and Google Cloud London region. UK import compliance for AI hardware has specific post-Brexit characteristics that differ from EU requirements:

• UKCA marking: UK Conformity Assessed marking is becoming the standard for regulated products following the post-Brexit transition from CE marking. Servers imported for commercial use in UK data centres must carry valid UKCA documentation or accepted CE marking within the transitional period
• HMRC CDS: Customs Declaration Service mandatory since November 2023. EORI registration required for the importing entity
• Zero import duty: 0% on server hardware. 20% VAT at import, reclaimable by VAT-registered businesses
• Tier 1 BIS status: No BIS licence required for standard commercial end uses. Standard denied party screening and EEI filing at origin sufficient

The Multi-Cloud AI Infrastructure Import Compliance Gap Nobody Has Solved

The fundamental problem with multi-cloud AI infrastructure import compliance is structural. Each cloud provider has its own procurement team, its own logistics partners, and its own customs broker relationships in each country. The enterprise customer has none of these. When an enterprise signs multi-cloud agreements with AWS, Azure, and Google Cloud simultaneously and those providers begin deploying physical hardware into data centres in India, Germany, Singapore, and the UAE, the enterprise’s import compliance exposure does not sit with the cloud provider’s procurement team. It sits with whoever is named as end-user on the BIS export documentation.

Three specific multi-cloud AI infrastructure import compliance gaps emerge in every deployment:

• End-use certificate fragmentation: When AWS, Azure, and Google Cloud each import hardware independently into the same country, the end-use certifications required by BIS may reference three different importing entities for what is effectively one enterprise customer’s compute infrastructure. The enterprise customer must ensure their compliance status is documented across all three provider relationships, not just the one they most frequently interact with
• Certification timing misalignment: BIS India CRS certification for Blackwell servers takes 3-6 months. If AWS activates its India region in Q2 2026 and Azure activates its India region in Q3 2026, the enterprise customer may find itself with AWS compute available but Azure compute held at port because the Azure-imported hardware’s BIS certification was not coordinated on the same timeline as AWS’s
• IOR entity proliferation: Each cloud provider’s hardware import in each country names a different importing entity on the customs declaration. An enterprise with multi-cloud deployments across six countries could have 18 different IOR arrangements in place simultaneously, each with its own certification portfolio and compliance status, with no consolidated visibility across the full hardware estate

Real Example: Multi-Cloud AI Infrastructure Import Compliance Failure and What It Cost

A European enterprise AI platform deploying NVIDIA Blackwell B200 servers across AWS Mumbai, Azure Central India, and Google Cloud Mumbai simultaneously as part of a sovereign AI partnership. Combined hardware value across all three cloud providers: USD 240 million. BIS India CRS certification status at procurement decision: not assessed.

• AWS hardware: AWS procurement team confirmed BIS registration 4 months ahead. AWS India region Blackwell capacity online on schedule
• Azure hardware: Azure’s logistics partner submitted BIS application at time of shipment. Hardware held at Chennai port. BIS processing: 5 months minimum
• Google Cloud hardware: BIS registration applied for through a different certification agent. Registration pending. Hardware held at JNPT
• Demurrage on Azure and Google Cloud shipments: USD 80-180 per day across containers from both providers. Over 5 months: USD 400,000-900,000 in demurrage alone
• Enterprise SLA exposure: Sovereign AI partner contract included go-live penalties. Only one third of planned compute online at launch date
• Total cost of coordination failure: USD 1.2-2.4 million in demurrage plus project penalties, against a coordination cost that would have been a fraction of that amount

How to Structure Your Multi-Cloud AI Infrastructure Import Compliance Programme Correctly

1. Map every physical data centre location across all cloud providers before procurement. For each AWS, Azure, and Google Cloud region in your deployment plan, identify the country of physical hardware deployment. Each country is a separate import compliance jurisdiction. Build the country compliance map before signing any cloud agreements, not after the hardware is ordered
2. Coordinate certification timelines across all three cloud providers simultaneously. BIS India, SABER Saudi Arabia, IMDA Singapore, TDRA UAE, and MIC Japan certifications take weeks to months and must be in place before hardware ships. Confirm with each cloud provider’s logistics team which certifications are in place for Blackwell hardware in each target region, and identify any gaps that will create holds in your specific timeline
3. Establish consolidated end-use documentation across all cloud provider relationships. Maintain a single unified record of your controlled GPU hardware estate across all cloud providers in all countries. This record must include: the ECCN classification of each hardware type, the BIS licence exception or authorisation under which each shipment was made, the named end-user on each provider’s export documentation, access logging arrangements in each facility, and the point of contact for BIS reporting obligations
4. Engage a multi-country IOR provider before deployment, not after. An Importer of Record provider with active entities, certifications, and customs credentials across all your target markets provides consolidated multi-cloud import compliance management. Certification timelines are coordinated across all providers simultaneously. End-use documentation is maintained in a single compliance framework. The alternative, separate IOR arrangements for each cloud provider in each country, creates the coordination failure described in the real example above

How Carra Globe Supports Multi-Cloud AI Infrastructure Import Compliance

Carra Globe provides Importer of Record services for NVIDIA Blackwell, H200, and next-generation AI hardware deployments across all major cloud regions including India, Germany, Singapore, the UAE, Japan, the UK, and 170+ additional markets. Our Global Trade Compliance team manages ECCN classification, BIS export licence exception verification, denied party screening, end-use certificate preparation, and multi-cloud end-use documentation frameworks that consolidate compliance obligations across AWS, Azure, and Google Cloud deployments into a single managed structure. Our Delivered Duty Paid service provides full landed cost visibility for Blackwell server deployments including duty, VAT, and certification costs in each target country before any procurement commitment across any cloud provider. For businesses deploying across the AI infrastructure cluster, see our related guides to AI GPU import compliance 2026, Chip Security Act 2026, and importing NVIDIA H200 servers outside the US.

Frequently Asked Questions: Multi-Cloud AI Infrastructure Import Compliance

Does the cloud provider handle multi-cloud AI infrastructure import compliance for hardware in its own regions?

For standard cloud tenancy, the cloud provider imports its own hardware. But the enterprise customer is typically named as end-user on BIS export documentation and carries ongoing monitoring obligations. In co-location or dedicated host arrangements, the customer is the named importer and bears full compliance liability. Confirm which role applies to your arrangement with each cloud provider before deployment.

Do NVIDIA Blackwell GPUs require BIS export licences for all countries?

No licence required for Tier 1 countries including Germany, Japan, UK, Australia, and Canada. Standard denied party screening and EEI filing at origin are sufficient. For India, Singapore, and UAE, end-use documentation and KYC verification are required. Large-scale deployments may need Validated End User certification. Tier 3 countries are generally prohibited. Requirements depend on destination, end-user entity, and deployment scale.

What is the biggest multi-cloud AI infrastructure import compliance risk in India specifically?

BIS CRS certification timing. It takes 3-6 months with no fast track. In a multi-cloud deployment, each cloud provider’s hardware needs its own BIS registration confirmed before shipment. If one provider ships without it, their hardware is held indefinitely while the other providers’ certified hardware is live. The result is partial compute at go-live, which is operationally harder than a full delay.

How does the Chip Security Act affect multi-cloud AI infrastructure import compliance?

If enacted, it would require location verification on covered chips and mandatory BIS reporting obligations. In a multi-cloud setup, the enterprise customer named as end-user across all three providers carries these obligations simultaneously. The 180-day implementation timeline means preparation must start now. See our Chip Security Act 2026 guide for the full framework.

Can one IOR provider manage multi-cloud AI infrastructure import compliance across AWS, Azure, and Google Cloud simultaneously?

Yes, provided the IOR holds active entities, certifications, and customs credentials in each target country. A consolidated arrangement covers certification timelines across all providers, unified BIS end-use documentation, and single-point landed cost visibility. Far more efficient than separate IOR arrangements per cloud provider per country, which is where the most expensive coordination failures occur.

What happens to existing multi-cloud AI infrastructure import compliance if the Chip Security Act is enacted?

The Act’s primary requirement covers chips before export, so existing deployed hardware may not need immediate retrofitting. But Commerce’s authority to verify ownership and location abroad applies to chips already exported under any authorisation. Assess whether your current access control systems and inventory records satisfy those verification requirements. That assessment is the starting point for Chip Security Act preparation in any multi-cloud deployment.