Selling Data to US AI Companies From Outside the US

Yes. A company outside the United States can sell or license its data to US AI companies, and many already have: of the 78 disclosed licensing deals whose licensor's head office we could place, 32 were with a licensor outside the US, in 13 countries. What changes is the price. The buyers that publish an estimator value the same company at 0.71 times its US figure when it is based in Europe, the UK or Canada, and at 0.27 times anywhere else, before they have seen a single record. Handshake's calculator shows the mechanism in its own settings: 1.00 for the USA, 0.75 for Canada and Europe, 0.30 for everyone else.

Price is one of five things that decide how a sale from abroad turns out. The others are which buyers take holders from your country at all, what your own privacy law asks before records leave, the US tax taken from the fee, and how the money reaches you past sanctions screening. None is a reason not to sell: each is a known cost with a known answer, best settled before the first buyer call. This guide puts them in order, links the deeper guide for each, adds a table by country, and walks four companies (in the UK, Germany, India and Singapore) through the whole journey.

It is written for the founder, CEO, CFO or counsel of a company outside the US that holds data an AI buyer might want: support tickets, chat, email, documents, code, call recordings, machine logs, archives. Where we stand: we run sales only for holders based in the United States, the United Kingdom and Canada today; our valuation works for a holder anywhere and binds no one.

This guide is research from public sources, not legal, tax or investment advice.

Contents

1. The short answer
2. The five things that decide the outcome
3. Factor 1: the buyers that take holders from your country
4. Factor 2: the price buyers put on your headquarters
5. Factor 3: your own privacy law, and getting the data to the US
6. Factor 4: the US tax on the fee
7. Factor 5: getting paid, and sanctions
8. Country by country
9. Language, and what sells from abroad
10. Deals already done by licensors abroad
11. Contracts and disputes across borders
12. Worked examples: four companies, four countries
13. Questions people ask
14. What this means for you

The short answer

Most AI buyers are in the United States, but almost none of them shuts out a company abroad. Of the 43 buyers and labs' programmes whose own pages we read on 18 October 2026 for where a holder may be based, 13 take holders from anywhere or already work with holders abroad, 2 take holders only from a listed set of places, 2 take holders abroad but rank the US first, 1 takes US holders only, and 25 say nothing at all. The deal ledger agrees: licensing across the border is ordinary, from the Financial Times with OpenAI to GrabMaps' map data from Singapore.

The five things that decide the outcome

The order matters. Buyers and price decide whether a sale is worth pursuing; privacy law decides what you can deliver and how long preparation takes; tax and payment decide how much of the fee arrives. Start with the law before you know any buyer wants your records and you spend money in the wrong order; sign before the law is settled and you give warranties you cannot keep.

The table is the whole guide in one place, in that order. Each row is a section below, with the guide or page that goes deeper.

Factor 1: the buyers that take holders from your country

Ask a buyer "do you take companies outside the US?" and most answer nothing in writing. Among the 24 buyers of a company's own records whose pages we read (the buyers that license support tickets, email, chat, documents and code), 5 take holders from anywhere or show they already work with holders abroad, 1 keeps a closed list of countries, 2 take holders abroad but rank the US first, 1 takes US holders only, and 15 say nothing about where a holder may be based.

The open end is real. SuperAnnotate's page says "Any size, any industry, anywhere." Turing takes code and records from "most regions", with strongest demand in the US, UK, Canada and the EU, and is the one buyer that names the legal reason a country can be hard: "Some jurisdictions carry data-export restrictions we'll flag early." Ooak runs a French entity that keeps EU partners' data in the EU. The General Data Company, based in San Francisco, courts German industry with euro ranges in its German calculator.

The closed end is narrow. Mode's referral pages take holders from the US, Canada, the UK or Australia only, though its calculator still offers "EU" and "Other". micro1 "accepts referrals globally" but says "We currently prioritize US companies", with demand strongest in the US, UK and Canada; its separate track for companies that have shut down lists ten countries, India, Singapore, Germany and Australia among them, in its own posting. The one buyer of company records that takes US holders only does so through payroll, not registration: Nyne's programme wants at least 20 full-time U.S. W-2 employees.

What "says nothing" means

A silent page is not a refusal. It means the answer arrives late, in the buyer's own agreement, after you have scoped your data, signed a confidentiality agreement and sent a sample. micro1's published data licence master terms, one of the few a holder can read in advance, show the usual shape: the holder warrants compliance with the laws on cross-border transfers that apply to it. Treat silence as "not decided in public" and ask in writing before any sample leaves your systems. The buyers guide has the eight questions to send, and every buyer's rules, in its own words, sit on one table at /countries#buyers. The buyer match tool checks your size, region and language against each buyer's published rules.

Marketplaces and buyers based abroad

Marketplaces publish country rules more often, because they pay holders in many places. AWS Data Exchange, Amazon's data marketplace, takes providers from thirteen listed places (the US, the EU, the UK, Switzerland, Norway, Israel, Japan, Australia, New Zealand, Hong Kong, the UAE, Bahrain and Qatar), leaving out Canada, India, Brazil and Singapore, though AWS's general page on who may sell adds India and South Korea. Datarade in Germany takes registered businesses from a full world country list, and Opendatabay in the UK "serves users globally". The buyers guide also lists buyers and brokers based outside the US.

Factor 2: the price buyers put on your headquarters

This is the finding that changes the most money. Four buyers publish an estimator that asks where your company is based, and each cuts the estimate for a company outside the US, whatever its data. Two companies with the same head count, years of records and software tools get different numbers because one answered "Europe" or "Other".

Handshake shows it most plainly, because its calculator's settings are readable in the page's own script: its four country answers carry the weights 1.00, 0.75, 0.75 and 0.30 for the USA, Canada, Europe and everywhere else. For a 20-person company with three years of records, visitors see $100,000 to $296,551 in the US and "Up to $100,000" for the same company elsewhere. micro1, which runs the same formula, shows $379,290 to $888,088 for a US company of 100 people with ten years of records, $284,468 to $666,066 for the same company in Europe and $113,787 to $266,426 anywhere else.

Our valuation fits all four estimators together: averaged over every pair of outputs that differ only in location, a company in Europe, the UK or Canada is valued at 0.71 times the US figure and one anywhere else at 0.27 times. The buyers guide sets the four estimators side by side.

What the discount is, and how to answer it

The discount is a location factor: one multiplier on the whole estimate, applied before anyone has read a record. It is not a judgment of your data or a price (every estimator calls its output indicative, and buyers price only after a sample), but it sets the anchor, the first number you hear, and first numbers are sticky. Our negotiation page treats it as a lever, the headquarters discount.

The buyers' own words point to three reasons for it, and each can be answered. Demand: micro1's is "strongest for companies based in the United States, United Kingdom, and Canada", so a company in those places should insist on being priced as one. Language: micro1 asks for English records and contacts, so records kept in English should be shown to be, in a datasheet. Law and export risk: Turing flags "data-export restrictions", and a buyer that must work through your privacy law prices that work in, so a holder that arrives with its rights file settled and its records de-identified has removed most of what the factor pays for.

Answer all three, put one datasheet before several buyers, and the factor becomes an assumption to test. A valuation priced against sales and offers, not one buyer's form, gives you a number to hold the anchor against.

Factor 3: your own privacy law, and getting the data to the US

Your records were collected under your law, and it follows them out of the country. In every country we read, a licence to an AI buyer is a new purpose for personal data the business collected to run itself, and that purpose needs a lawful footing before anything leaves. The footing differs, which is why preparation takes weeks in one country and months in another.

Records about companies rather than people are the cheap win everywhere. The GDPR does not cover data about legal persons, including the name and the form of the legal person and the contact details, and much of what buyers value (how a product works, how a problem was solved, the code) can be separated from the people in it. The rest needs a written balancing test where the law allows one, consent where it does not, or de-identification to the point that the lot no longer identifies anyone.

Three families of law

A written balancing test: the EU and the UK. The usual basis is legitimate interests, backed by a written compatibility test; people must be told in advance and be able to object; sensitive data must be kept out. A company holding its customers' data as their processor cannot license it on its own decision: one that does shall be considered to be a controller and has broken the rules. Switzerland has no free-standing legitimate interests basis: processing that breaches personality rights is unlawful unless justified. The GDPR guide covers all three.

Expectation or consent: Australia and New Zealand. Australia allows a secondary use with consent or where the individual would reasonably expect it and it is related, and its regulator says that will be difficult to establish for service data used to train a model sold elsewhere; even a small business loses its exemption when it discloses personal information for a benefit. New Zealand allows another use only if it is directly related, authorised, or in a form that does not identify anyone.

Consent first: Canada, South Korea, Israel, the UAE, Malaysia, Mexico. These laws have no balancing test a licence can use. Canada requires consent before personal information is used for a purpose not identified at collection, excludes a deal whose primary purpose is the sale or lease of personal information from its business-transaction exception, and judges anonymisation strictly: no serious possibility of re-identification. South Korea needs separate consent to give data to a third party for a new purpose. Israel lets a company process a database only for the purpose it was established for, and its regulator's draft AI guidance points to a separate opt-in. The UAE prohibits processing without consent outside listed exceptions; Malaysia requires consent and written notice before a new use. Mexico is the lightest of these: a new purpose needs consent again, which for ordinary data can follow from a revised notice, and dissociated data needs none.

The privacy law by country guide takes these countries one by one, with Japan, India, Singapore, Brazil, Türkiye and China, and each country page puts the law in its own words beside the tax and the buyers.

Getting the data to the US

Sending identifiable records to a US buyer is a cross-border transfer, and most privacy laws have a rule for it. In the EU, a buyer on the Data Privacy Framework list can receive personal data with no further tool, and one that is not needs appropriate safeguards, in practice the standard contractual clauses, whose controller-to-controller module binds the buyer to the specific purpose(s) of the transfer. UK data goes on the same list only for buyers participating in the UK Extension; Swiss data goes to buyers certified under the Swiss-US framework. An appeal against the EU framework is pending, so write the clauses into the licence as a fallback.

Outside Europe, none of the laws cited here names the United States as an approved destination, so the route is a contract or consent. Canada keeps no list of approved countries and relies on the sender's accountability and its contract. Australia expects a contract making the buyer handle the information as the Australian principles require and keeps the holder accountable for the buyer's breaches. Malaysia accepts consent or a set of listed routes, contract clauses with an assessment among them; Israel, consent or an agreement binding the buyer to Israeli conditions. South Korea is the hard case: it has recognised only the EU so far, so for a US buyer the practical ground is separate consent to the transfer.

Why de-identifying first is the common answer

Every one of these rules gets easier when people are taken out of the records before anything leaves. Truly anonymous data is outside the GDPR, and Australia treats properly de-identified information as no longer personal information. A claim of anonymity must be shown, not asserted, and where it falls short, the de-identified lot is still treated as personal data with a transfer tool in place. Our de-identification standard says what counts, and the de-identification tool runs on your own computer, so the originals never leave it.

Factor 4: the US tax on the fee

When a US buyer licenses your records and trains on them in the United States, the fee is a royalty for using property in the United States, so US income. The buyer must hold back 30 percent of every payment to a foreign company for the IRS, unless a tax treaty lowers the rate and you give the buyer a Form W-8BEN-E claiming it before the payment is made. A buyer that fails to withhold is made liable for the tax, so it asks for the form before the first invoice.

Treaty rates on a data licence run from 0 percent for a holder in the UK, Germany, France, Japan, Switzerland or Canada, through 5 percent for Australia and 10 percent for China, to 15 percent for India. Singapore, Brazil, the UAE and Malaysia have no income tax treaty with the US (the IRS list has no entry for Brazil, Singapore or the UAE), so a holder there bears the full 30 percent.

Three points change the answer, and each is in the withholding guide. What you are paid for: a term licence that ends with a duty to delete is classified as a lease of a copy, a royalty, while pay for work performed abroad to create new data is foreign-source and not withheld. Where the buyer trains: a royalty for use abroad is foreign-source too, so write the place of training into the contract. Which column: in Italy, Israel and South Korea, copyright and know-how carry different rates, so agree how the licensed records are described before the first payment.

Your own sales tax is the easy part: a licence to a US business is outside EU VAT, supplied where the buyer has established its business, outside UK VAT, a zero-rated export in India and zero-rated in Canada.

Factor 5: getting paid, and sanctions

Few buyers publish their currency or payment method, and the route matters. Stripe Connect, the rail many US platforms use to pay suppliers, reaches connected accounts only in "the US, UK, EEA, Canada, or Switzerland". That covers holders in the UK, Canada, Switzerland and the European Economic Area, and not those in India, Japan, Australia, Singapore or Brazil. The route that works for any holder is an international wire in US dollars, for which a US bank asks for the beneficiary's full name, address, account number, bank name and SWIFT/BIC; the conversion spread at your own bank is a real cost to ask about.

Buyers' own rails differ. Mercor pays by wire on delivery, Troveo through Tipalti, and AWS Data Exchange pays monthly to a bank account that can accept US dollars, with a virtual US account from Hyperwallet for a provider without one. Ask the currency, the method and who bears conversion before you sign.

Sanctions: who a US buyer cannot pay

Sanctions decide whether the buyer may pay at all. OFAC, the US Treasury's sanctions office, "does not maintain a specific list of countries that U.S. persons cannot do business with": some programmes are tied to a place, such as the bar on dealing in "goods or services of Iranian origin", and the rest name people and companies. A company owned "in the aggregate, directly or indirectly, 50 percent or more" by blocked persons is blocked whether or not it is named, so buyers screen the holder, its owners and often its bank. Their terms say so: Mercor's users confirm they are not on a US sanctions list or in a sanctioned country, and Prime Intellect names six excluded places.

A separate Justice Department rule limits dealings in Americans' sensitive data with six countries of concern: China, Cuba, Iran, North Korea, Russia and Venezuela. It bites when a deal gives a covered company access to bulk US sensitive personal data; on the rule's own example, a deal in which the US company accesses data the foreign holder already holds "is not a covered data transaction". For most holders abroad, whose records are about people outside the US, the rule is not engaged.

Country by country

The table puts the five factors side by side for the countries holders ask about most. Each country's name links to its page, where the law is quoted in its own words and every buyer's answer for that country is listed.

"The open buyers" means the 5 buyers of company records whose pages say anywhere or show holders abroad; they take holders from every country here, and each cell adds what is particular to the country. Price is the fitted location factor for its region; withholding is the treaty rate on a royalty, claimed on a W-8BEN-E; privacy work is our reading of what a licence needs first.

Three patterns stand out. The UK and Canada sit closest to the US on buyers, price and tax, but Canada's privacy law, like the rest of the consent-first family, offers no balancing test, so its holders need consent or strict anonymisation. The "elsewhere" price band splits sharply on tax: a Japanese holder keeps the whole fee with a form, while a Singaporean, Brazilian, Emirati or Malaysian one loses the full default rate. And none of the laws outside Europe in the table names the United States as an approved destination, so there the transfer rests on a contract or on consent.

Language, and what sells from abroad

Language is the second filter after headquarters, and it splits buyers cleanly. Buyers of company records want English. micro1 is "primarily looking for companies that operate in English", and Turing wants "English-language comments & issues" in the codebases it buys. The other buyers of company records publish nothing on language, so a sample in another language may be judged by reviewers who cannot read it, and none publishes a price, or a multiplier, for company records in another language.

Buyers of speech, text and video want other languages. OpenAI's data partnerships ask for data "across any language, topic, and format", working with partners "eager to represent data from their country or industry"; Protege runs speaker tracks in 70+ languages; FileYield says "Multi-language call data is extremely valuable"; Corpus Bridge says languages other than English add value. Read the labs' programmes closely: Meta's one named partner is the Government of Nunavut, sharing Inuit-language data, and it offers workshops, not payment.

Price follows where the data was recorded, not the language

List prices for speech show that a discount follows the speakers' country, not their language. One vendor lists conversational English recorded in the US at $90 an hour and English recorded in South Africa at $60; Spanish from Spain at $85 and Spanish from Mexico at $65; French from the Democratic Republic of the Congo at $65. Same language, different price.

It is not an English premium either. A European publisher of speech data prices British English well above American English recorded the same way, per hour, with Hindi a little below it, and one marketplace vendor lists customer-service calls in English and in Hindi at the same $24,500 for 1,000 hours. These are list prices with no sale behind them; the one average over real sales we hold is a Chinese speech company's price per hour licensed, across Chinese and foreign languages together.

What a holder abroad can still price

Much non-English data is now produced for free: Google funds Project Vaani, open speech in Indian languages, and Cohere's Aya datasets are openly licensed. For commodity text and read speech in a large language the buyer has a free alternative, which caps the price. What stays scarce is what free programmes cannot produce: real conversations (support calls, chats, negotiations), linked records (a ticket, its email thread, the code change that closed it) and clear rights. A holder with records partly in English should split the lot: the English part to buyers of company records, the rest to buyers of speech and text, each in its own datasheet. What labs buy, /data/audio-speech and /data/text-archives carry the prices by kind.

Deals already done by licensors abroad

The deal ledger shows money crossing the border often. Of the 78 disclosed deals whose licensor's head office we could place, 32 were with licensors outside the US, in 13 countries.

Most are publishers and media companies licensing archives to US labs, with a few European and government buyers among them. The table groups the ones a holder is most likely to have heard of by the licensor's country.

Two cautions for a company holder. Almost every licensor here is a publisher, broadcaster or map company, so the deals prove that US buyers pay licensors abroad, not that they pay a company abroad for its tickets and email at the US rate. And terms are rarely stated: the only orders of magnitude for the large publishers come from press reports, such as Axel Springer's reported figure, which the parties never confirmed. The deal ledger has each deal's sources, and the news publisher guide covers that market.

The closest thing to a company-records deal abroad is Protege's GrabMaps placement: data a Singapore company built in running its business, placed by a specialist intermediary as a licence of a defined lot, on terms that stay private. That is the shape a company holder abroad should expect.

Contracts and disputes across borders

A licence across a border raises three questions a domestic licence does not: whose law governs it, where a dispute is heard, and whether a win there can be enforced where the other side's money is.

Choose the law, and know what it cannot override. Parties are generally free to pick the law that governs their contract, and a US buyer's form will propose a US state's. EU law confirms the freedom, with a reserve: an EU court may still apply overriding mandatory provisions of its own law, whatever the contract says. Two rules travel with you regardless. Your privacy law governs your own processing, whichever law governs the licence. And for an EU holder, the Data Act removes from a business-to-business data contract an unfair term on access to and the use of data, or liability and remedies that the stronger party imposed; whether a court would apply it to a licence under Delaware law is untested. If you send personal data under the EU's standard clauses, those clauses are governed by a member state's law and heard by its courts, whatever the licence says for the rest of the deal.

Choose how disputes are heard, with enforcement in mind. A court clause sends a dispute to judges in one place; an arbitration clause sends it to arbitrators under institutional rules, with an award that is usually final. The difference shows when you win. A court judgment may not be enforceable where the other side's money is: the Hague convention on enforcing foreign judgments applies in the EU and the UK, but the United States has signed it and not ratified it, so a judgment between a US buyer and a holder abroad has no treaty route under it. An arbitral award travels much better. The New York Convention obliges each contracting state to recognize arbitral awards as binding and enforce them, the US applies it in its own courts, and a US court shall confirm the award unless one of the Convention's narrow grounds for refusal applies.

So the practical defaults are: accept the buyer's US governing law if you must, but ask for arbitration seated in a neutral place, in English, under recognised rules; state the purpose (AI model training), the deletion duty and the confidentiality terms your own law needs; and put the tax terms in writing (a gross fee, the rate the buyer will apply, where it trains, a copy of its year-end form). Our guide to data licensing agreements walks through each clause, and the licence terms register shows what buyers propose.

Worked examples: four companies, four countries

These are examples, not real companies. Each walks the five factors in order.

A 60-person UK software company

Example. A Manchester company sells field-service software to mid-sized businesses. It holds six years of support tickets in Zendesk, documentation in Confluence, Slack, and its Git history, all in English.

Buyers and price. It is in every favoured group a buyer publishes: Mode's four-country list, micro1's strongest-demand group and Turing's strongest-demand regions, and its English records meet micro1's language rule. The estimators start it at about 0.71 times of a US twin; with demand, language and law position all close to a US company's, it has the strongest case of any holder abroad to argue the factor away.

Law. The tickets name its customers' staff, so they are personal data under the UK GDPR. It writes a legitimate interests assessment, leaves out data it holds for customers inside its product, de-identifies and tests, tells customers and waits for objections, then sends only to a buyer in the UK Extension or under the ICO's transfer agreement.

Tax and money. With a W-8BEN-E, nothing is withheld ( 0 percent); the licence is outside UK VAT; a US buyer can pay through Stripe Connect or a dollar wire. See the UK page.

A German manufacturer

Example. A family-owned maker of industrial pumps near Stuttgart, with a works council, twelve years of machine telemetry, quality reports, technicians' service tickets and engineering drawings. Its documents are mostly in German; its telemetry has no language.

Buyers and price. Its best buyers are not the programmes that want English. The General Data Company courts German industry in German with euro ranges, the open buyers take it, and Datarade, a marketplace based in Germany, involves no border at all. Estimators built for office records start it at 0.71 times of a US twin, but telemetry and drawings are priced by kind of data, not by head count (see /data/physical-world).

Law. Telemetry about machines is not personal data and records about companies are outside the GDPR, so most of its lot needs a contract, not a privacy assessment. An export of service tickets showing what each technician did can come within the works council's co-determination right, so the council is consulted first. Its drawings stay trade secrets only if the licence limits use and bars disclosure, and against a one-sided buyer form it can invoke the Data Act's unfair terms rule.

Tax and money. 0 percent withholding with the form, no VAT on the invoice ( supplied where the buyer is established), Stripe Connect or wire. See the Germany page.

An Indian support outsourcer

Example. A Pune company whose agents answer chat and calls on behalf of companies in the US, the UK and Australia. It holds transcripts, quality scores, training scripts and its own operations data, in English, with some internal chat in Marathi and Hindi.

Whose records they are. This decides everything. The transcripts belong to those companies, its customers, and for its UK and EU customers it is their processor, which cannot license their data on its own decision ( it would become a controller and break the rules). What it can license is what it created (training scripts, quality rubrics, de-identified workflow data), plus transcripts only where a customer's contract expressly permits it.

Buyers and price. micro1 says companies in other regions "may still be considered", and its English records meet the language rule; Handshake's calculator starts it at 0.30 of a US twin, and the fitted figure is 0.27 times. Real support conversations are exactly what free programmes cannot produce, which is its argument against the anchor. Indian privacy law applies to its own staff and customer data: see the India page.

Tax and money. A royalty is withheld at 15 percent with the form, and India allows the US tax "paid in the United States, whether directly or by deduction" as a deduction against its own tax on the same income. Labelling or evaluation work bundled in can fall under fees for "technical or consultancy services", so price the parts separately. The licence is a zero-rated export for GST, and since Stripe Connect does not reach India, it is paid by dollar wire.

A Singapore e-commerce company

Example. A Singapore online retailer selling across Southeast Asia, with five years of customer chat in English, Malay and Chinese, catalogue data, reviews and order history.

Buyers and price. The open buyers take it; micro1's track for companies that have shut down lists Singapore; AWS Data Exchange leaves it out; Protege has already placed a Singapore company's data. It starts at 0.27 times of a US twin, and should split the lot: English chat to buyers of company records, Malay and Chinese text to buyers of language data. Singapore's privacy law applies to the chat: see the Singapore page.

Tax and money. This is where Singapore differs most. With no income tax treaty, a royalty is withheld at the full 30 percent. The levers are where the buyer trains (a royalty for use abroad is foreign-source), what is paid for (new data collected in Singapore is a service performed abroad) and the gross price. GST is zero-rated only if the service qualifies, and payment comes by dollar wire.

What the four examples show

Questions people ask

Can a foreign company sell data to a US AI company?

Yes. 32 of the 78 disclosed deals whose licensor we could place were with licensors outside the US, and most buyers of company records either take holders from anywhere or say nothing about where. Only one, Nyne, ties its programme to the US, through US payroll.

Do AI companies pay less for data from other countries?

At the estimator stage, yes: buyers' own estimators average 0.71 times of the US figure for a company in Europe, the UK or Canada and 0.27 times elsewhere. That is an anchor set before anyone reads your records, not a price, and English records, a settled rights file and competing interest can move it.

Do I have to pay US tax on data licence income?

The buyer withholds it from the fee: 30 percent by default, lowered by a tax treaty if you give it a W-8BEN-E before payment. Holders in the UK, Germany, France, Japan and Canada pay 0 percent with the form; holders in countries with no treaty pay the full rate.

Can I send personal data from Europe to a US AI company?

Yes, to a buyer on the Data Privacy Framework list for your country, or under the standard contractual clauses, after a lawful basis, notice and a chance to object. Most holders de-identify first, because truly anonymous data is outside the GDPR and needs no transfer tool.

Which AI data buyers accept companies outside the US?

SuperAnnotate says "anywhere", Turing takes "most regions", and marketplaces such as Datarade and Opendatabay take registered businesses worldwide. The buyers guide has every buyer's rule, and /countries#buyers keeps them current.

Does it matter if our records are not in English?

For buyers of company records, yes: micro1 wants companies that operate in English and Turing wants English comments in code, and none publishes a price for other languages. For buyers of speech and text, other languages are wanted, and price follows where the data was recorded more than the language itself.

What this means for you

A company outside the US can license its data to US AI buyers, and many will talk to it. What changes is the starting price, the language a buyer will read, and the paperwork around privacy, tax and payment, little of it written down in advance. Settle them in this order.

1. Find your country. Your page under /countries lists each buyer's answer, your withholding rate and your privacy work; every buyer's own rules sit on one table.
2. Know what your records are worth before a buyer's anchor. A valuation prices your records against sales and offers; how much company data is worth explains the method.
3. Find the buyers that fit with the buyer match tool, the buyer board and programmes compared; which buyers take holders outside the US has the questions to send.
4. Settle your own law. In Europe, the GDPR guide; elsewhere, privacy law by country. De-identify with the de-identification tool, which runs on your own computer, and check the provenance a buyer will ask for.
5. Settle the tax and the route for the money. The withholding guide has your treaty rate, the form and the payment rails.
6. Read the contract for the cross-border clauses: governing law, disputes, purpose, deletion and tax, with what is a data licensing agreement? beside you.

For the whole selling process, start with how to sell company data to AI labs. How we are paid when we run a sale is on our pricing page.

This guide is research from public sources, not legal, tax or investment advice.