How to import AI servers and GPUs without a local company

Yes, you can import AI servers into a country where you have no company, as long as a qualified party acts as Importer of Record, files the customs entry and pays the local duties and taxes. Before anything ships, confirm the US export control status of the GPUs and agree who owns each document.

That short answer hides most of the work. A single rack of GPU servers can be worth more than a small office building, it often contains chips that the United States controls for export, and it usually has a hard installation date attached to a customer contract or a data center opening. When the buyer, the seller and the end site sit in three different countries, the question of who is legally importing the equipment becomes the question that decides whether the project runs on time.

This guide walks through how the import of AI servers and GPU clusters works when your company has no legal entity in the destination country. It covers the role of the Importer of Record, the export control checks that come before the shipment, what changes in the United States, Brazil, Argentina, the European Union and Uruguay, and the steps that keep a high-value delivery from getting stuck at the border.

Can you import AI servers without a local company?

In most of the destinations we work with, yes, but not on your own. Customs needs a party that is registered locally, or accepted by the local authority, to declare the goods and answer for the duties, taxes and compliance that follow. If your company is not that party, someone else has to take the role.

That someone is the Importer of Record (IOR). If you want the full background on the role, our guide on what an Importer of Record is and when you need one explains it from the ground up. For AI hardware, the short version is that the IOR is the name on the customs declaration, the party that pays or guarantees the import duties and taxes, and the party the authorities contact if something is wrong with the classification, the value or the paperwork.

Some countries let a foreign company register directly as an importer, and others require a locally established entity. Even where a foreign company can technically register, the time it takes to get tax numbers, customs registrations, bonds and bank arrangements in place is often longer than the project schedule allows. That is why many hardware sellers, integrators and cloud providers use an IOR service for the first shipments into a new country, and sometimes keep using it for good.

Who is legally responsible when GPU servers cross the border?

Responsibility is split between the exporter in the country of origin and the importer in the destination country. The exporter answers for export controls and export declarations. The Importer of Record answers for the import declaration, classification, customs value, duties, taxes and any product approvals the destination requires.

This split matters because AI servers are almost never a simple box sale. A typical shipment includes the servers themselves, high-speed networking cards, switches, optical transceivers, cables, power distribution units and sometimes liquid cooling parts, each with its own tariff classification. The commercial invoice has to describe every line clearly enough that a customs officer in the destination can classify it, and the declared value has to be consistent with the purchase order and the transfer price, if the buyer and seller are related companies.

Incoterms also shape who does what. If the seller agrees to Delivered Duty Paid terms, the seller is promising to handle the import in the destination, which means the seller needs an Importer of Record there. If the buyer takes over at the port or the airport, the buyer, or the buyer's IOR, takes the import role. Getting this wrong in the contract is one of the most common reasons a GPU shipment sits in a warehouse while two companies argue about who should have registered where.

Do US export controls apply before your AI servers even ship?

Often, yes. Servers built around high-end accelerators can fall under US export control classifications such as ECCN 3A090 and 4A090, and the rules follow the chips even when the servers are assembled or shipped from outside the United States. Whether a license is needed depends on the exact item, the destination and the end user.

The rules for advanced computing items have changed several times in a short period. The Department of Commerce rescinded the AI Diffusion Rule on May 13, 2025 and, at the same time, published guidance warning companies about diversion tactics and about letting US chips be used to train AI models for parties of concern. In January 2026, the Bureau of Industry and Security published the Revision to License Review Policy for Advanced Computing Commodities, which moved certain chips below defined performance thresholds to case-by-case review for China and Macau, with conditions such as third-party testing. The license requirements themselves sit in Section 742.6 of the Export Administration Regulations, which is the text to check when someone tells you a shipment is "no license required".

For an importer in Latin America or Europe, the practical consequence is simple. Ask the manufacturer or the seller for the export classification of every major item before you book freight, and expect to answer end-use and end-user questions. Sellers of top-tier GPU systems routinely ask for end-user statements, the address of the final data center and sometimes a description of who will access the compute. An Importer of Record cannot replace the exporter's compliance work, but a good one will check that the documents line up so that the import side does not contradict what was declared on the export side.

What changes when you import AI servers into the United States?

A foreign company can act as Importer of Record in the United States, but the rules for doing so are getting stricter in 2026. Expect a customs bond, a licensed customs broker, careful classification and a close look at which tariffs apply to the specific chips inside the servers.

Two federal actions from 2026 are worth knowing. Proclamation 11002 of January 14, 2026 imposed a 25 percent Section 232 duty on a narrow set of advanced semiconductors and derivative products, effective January 15, 2026, while excluding imports for use in US data centers, US research and development and several other uses. If your servers carry the covered chips, the declared end use becomes part of the customs file. Then, Executive Order 14411 of June 3, 2026 directed CBP to stop foreign importers of record from filing informal entries, to limit their use of continuous bonds, and to require them to be validated in CTPAT or to use a CTPAT-validated customs broker, with most changes due within 180 days.

For a foreign hardware seller, this means the "we will just be the importer ourselves" option is narrowing. The trend points to US-based importers, or IOR providers with solid bonding and broker relationships, handling high-value technology entries. Valuation also deserves attention: AI servers are often bought under framework agreements with software, support and installation bundled in, and the customs value has to separate what is dutiable from what is not.

How do the rules compare in Brazil, Argentina, the EU and Uruguay?

Each destination has one or two points that decide the timeline. Brazil and Argentina focus on importer registration and product approvals, the European Union on the EORI number, import VAT and product conformity, and Uruguay on its free zone regime.

In the European Union, the EORI number is mandatory for all customs operations, and a non-EU company gets it in the member state where it lodges its first declaration. Servers placed on the EU market also have to meet the ecodesign requirements for servers and data storage products in Regulation (EU) 2019/424, so the technical file from the manufacturer matters as much as the invoice.

In Brazil, the main question is who holds the Siscomex qualification and whether its limits cover the value of the shipment, which is a real constraint when a single order can run into millions of dollars. Any server or network item with a radio module, or with functions Anatel treats as telecommunications, needs Anatel approval before it can be cleared. Argentina has updated its regime for telecommunications equipment in 2026, so items with radio functions should be checked against the current ENACOM rules before shipping. Uruguay is often used as a regional hub through its free trade zones, which can make sense when the equipment will serve customers in more than one country.

Which costs should you budget for besides the hardware?

Plan for freight and insurance, import duties where they apply, import VAT or equivalent taxes, customs brokerage, storage, local delivery and the IOR fee. On high-value AI servers, taxes and insurance usually weigh far more than freight, so the tax structure deserves attention early.

Many servers and computer parts enter major markets at low or zero import duty because of trade agreements on information technology products, but that does not make the import cheap. Import VAT in the European Union and the cumulative federal and state taxes in Brazil are calculated on values that include the hardware, and in some cases freight and insurance as well. On a shipment worth several million dollars, the cash needed to pay those taxes at clearance can be larger than the whole logistics budget, and the question of who can recover them afterwards depends on who the importer is and how the sale is structured.

Insurance is the other line people underestimate. A shipment of GPU servers is a high-theft, high-value cargo, and insurers will ask about routing, packaging, storage and handover points. If the equipment sits at an airport warehouse for a week because a document is missing, the risk and the storage bill both grow.

Finally, think about timing costs. If the servers are tied to a go-live date, every day in customs has a cost for your customer, even if it never shows up on a customs invoice. That is the strongest argument for preparing the import before the equipment leaves the factory.

How do you import AI servers on schedule, step by step?

The smoothest imports follow the same sequence: confirm the export classification, decide who will be Importer of Record, check destination approvals, align the commercial documents, and only then book freight. Doing these in parallel with production, rather than after the goods are ready, is what saves weeks.

1. Collect the technical data: model numbers, GPU and CPU details, power supplies, radio modules, the export classification from the manufacturer and the country of origin of each item.
2. Define the import structure: who sells to whom, which Incoterm applies, who will be Importer of Record and who will own the equipment after clearance.
3. Check destination requirements: importer registrations, product approvals, conformity marking, restrictions on used or refurbished units and any duty or tax incentives available.
4. Prepare consistent documents: commercial invoice with clear descriptions and values, packing list with serial numbers, end-user statements where the exporter needs them, and certificates the destination asks for.
5. Plan the physical move: packaging suitable for high-value electronics, insurance, routing, temperature and shock considerations, and the final delivery to the data center loading dock.
6. Clear, deliver and close the file: pay or guarantee duties and taxes, release the goods, deliver to site and keep the import records for the period the local authority requires.

One detail that often surprises first-time importers is serial numbers. For high-value GPU systems, both export and import sides may ask to match serial numbers across invoices, packing lists and end-user documents. Having them in a clean spreadsheet before the shipment leaves the factory avoids long holds later.

When does an Importer of Record make sense for AI hardware?

An Importer of Record makes sense when the equipment has to arrive in a country where you have no entity, no customs registration or no time to set one up. It also helps when the destination's rules are new to your team and the cost of a mistake is high.

That covers a lot of AI projects. A hardware vendor closing its first deal in Brazil, a cloud provider placing a GPU cluster in a European colocation site, an integrator delivering a turnkey rack to a customer in Argentina, or a company testing demand in the United States before opening a subsidiary all face the same problem: the goods are ready, the customer is waiting, and nobody local is set up to import. An IOR service fills that gap by taking the import role, coordinating brokers and carriers, and keeping the documentation consistent from the factory to the loading dock.

It is also worth saying when it makes less sense. If you already have a local entity with customs registrations and an experienced broker, and you import regularly, you probably do not need an external IOR for routine shipments. The case for an IOR is strongest on first entries, on unusual destinations, on very high values and on projects with fixed dates.

You can find more practical guides on importing technology in the Comex IOR Services blog on importing without a local company. If you are planning a GPU cluster or a batch of AI servers for a country where you have no structure yet, talk to us about an Importer of Record for AI servers. Tell us what you are shipping, where it needs to go and when it has to be running, and we will tell you plainly whether an IOR structure makes sense and what it would take.