Company registration for an AI startup: how to choose a jurisdiction in 2026

The standard approach to choosing a country runs like this: compare corporate rates, check the double tax treaty network, work out the cost of upkeep. For a trading or holding company that works fine.
For an AI company it does not.

The reason is simple. An early-stage AI business usually has no profit to tax. What it has is costs: inference, training, salaries, cloud. For the first two or three years the tax rate changes nothing, because the base is zero or negative. Access to infrastructure, on the other hand, changes everything.

A company incorporated in a country that does not appear on a model provider’s supported list cannot connect to the API legally. A company without a bank account in a jurisdiction where payment processing works cannot take money from customers. A startup without a recognisable corporate form will not get an investment from a fund that works off standard documents.

Hence the working principle: first establish where the business can actually operate, then compare tax within whatever is left. That is the reverse of the usual order, and it is exactly where most founders get it wrong.

There is no single answer to “where should I incorporate my AI company”, because the label covers four different businesses with four different constraints.

The largest category by far. You are not training a model; you are building an interface, a pipeline or an agent on top of the OpenAI, Anthropic or Google APIs, or on open weights.
What matters: API access, payment processing, speed to launch. What does not: research infrastructure, R&D incentives, regulatory sandboxes.

One important caveat. If you are doing substantial fine-tuning of someone else’s model rather than simply calling it over an API, you may be reclassified from deployer to provider under the AI Act. The obligations that come with that are considerably heavier. The boundary is genuinely grey, and this is a case where you want a documented position from day one rather than a scramble when a regulator first asks.

You train your own weights and publish or license the model. Here the obligations for general-purpose AI model providers apply, along with requirements around training-data documentation and copyright compliance. There is also the question of compute access: export controls on accelerators follow the end user’s country, not the country where the entity sits.

You sell to business: CV screening, credit scoring, document processing, analytics. This is the most heavily regulated category. Several of these use cases are listed explicitly in Annex III of the AI Act as high-risk systems. Enterprise buyers in the EU will ask for AI Act documentation as part of vendor due diligence, even where the obligations have not formally kicked in yet.

Your main costs are researcher salaries and compute. Here R&D tax regimes produce real money rather than paper savings. Ireland raised its R&D tax credit from 30% to 35% in Budget 2026. The UK returns roughly 27% of qualifying expenditure through Enhanced R&D Intensive Support, even to a loss-making company.

Before comparing rates, run the business through three checks. Fail any one of them and the jurisdiction is out, whatever the tax looks like.

Check one: your model provider. OpenAI and Anthropic both publish lists of supported countries. Access from outside them leads to a blocked or suspended account. Anthropic’s list runs to around 158 countries as of mid-2026; OpenAI maintains its own. Russia, Belarus, Iran, Cuba, Syria and North Korea are supported by none of the major providers. Hong Kong is also absent from OpenAI’s list, despite no local prohibition, as a commercial decision by the provider.
The practical consequence: your country of incorporation, the country behind your payment card and the IP address you connect from all need to sit inside the supported perimeter. Working around geoblocking is not a viable basis for a business that intends to sell to enterprise customers and raise capital. An account suspended for breaching the terms of use takes the whole product down with it.

Check two: payments. Stripe supports businesses in roughly 46 countries. If yours is not on the list, you have two options: incorporate in a supported jurisdiction, or use a merchant of record arrangement in which the platform is the legal seller (Paddle, Lemon Squeezy, Dodo and similar). The second is more expensive at 4-5% against 2.9% plus a fixed fee, but it removes the need to register for VAT in dozens of countries. For a product selling to consumers internationally, that is often cheaper than staffing a tax function.

Check three: cloud and provider startup programmes. Inference and compute credits are a meaningful part of early runway. These programmes require a legal entity in a supported country, and the larger tiers usually depend on a referral from a fund or accelerator. A project without a corporate wrapper does not qualify.

Nobody chooses Delaware for the tax. The 21% federal rate is higher than Singapore or Hong Kong, and a second layer of tax lands on shareholder dividends. People choose it for the investment infrastructure.
Most companies in the portfolios of US accelerators are Delaware entities, and the entire standard financing toolkit, SAFEs, convertible notes, option plans, is drafted for a C-Corp under Delaware law. A fund offered a Cypriot company will either require a reorganisation or price in the extra legal risk. Restructuring after a round costs more than getting the structure right at the start.

The numbers. The filing fee for the Certificate of Incorporation starts at $89 and is $109 or more in practice, depending on the number of authorised shares and the length of the document. Same-day processing adds $50. The annual report costs $50, and the state franchise tax has a minimum of $175 under the Authorized Shares Method and $400 under the Assumed Par Value Capital Method.

This is where nearly everyone gets caught. A standard startup charter authorises 10,000,000 shares. Run that through the Authorized Shares Method and the bill runs into five figures, at which point the founder’s heart stops. You do not have to pay it: the law lets you choose the cheaper calculation. Set the par value at $0.0001 and use the Assumed Par Value Capital Method, and you land at roughly $400.

The EIN is the real bottleneck. You need an Employer Identification Number for the bank, for payment processing and for filings. A US resident gets one online in minutes. A founder without a Social Security Number files Form SS-4 by fax or post, and that takes four to eight weeks. Planning a product launch without allowing for it is a standard mistake.

What people forget. A foreign-owned company files Form 5472 with its federal return. The penalty for not filing is $25,000. Form 1120 is filed separately, due 15 April. Franchise tax and the annual report are due by 1 March. A dormant company with no revenue pays and files on exactly the same basis.
More on company registration in the USA

A UK limited company is an underrated option for AI projects whose largest line item is the engineering team.
Corporation tax is tiered: 19% on profits up to £50,000, 25% above £250,000, with marginal relief in between. For a profitable company that is not the lowest rate available anywhere. But a loss-making startup is not paying it, it is receiving money back.

Since April 2024 the merged R&D expenditure credit scheme has applied, at 20% of qualifying expenditure. Companies that spend 30% or more of total expenditure on research qualify for Enhanced R&D Intensive Support: an additional 86% deduction on top of the standard 100%, giving 186% in total, plus a payable credit at an enhanced rate for loss-making periods. The effective subsidy comes to roughly 27% of qualifying spend.
This matters for AI specifically. Pure mathematics now falls within the scope of the relief, a change introduced with data science and AI development in mind. Designing your own architecture, solving hard scalability problems, working on how well a model generalises: all of it qualifies.

A word on the risk. HMRC has tightened its scrutiny of software claims and the rejection rate has climbed. The objection is almost always the same: the claimant described ordinary development rather than the resolution of technological uncertainty. Documentation has to start on day one, not be assembled retrospectively.

An advance assurance pilot opened on 18 May 2026. Small and medium-sized businesses can obtain HMRC’s position on two aspects of a single project before submitting, including whether the project meets the definition of R&D. The pilot runs for twelve months. For an AI startup with a large first claim, it is a way to take the risk off the table in advance.

Cost of upkeep. From 1 February 2026 the online incorporation fee doubled to £100, with same-day incorporation at £156. The annual confirmation statement costs £50 online. Note separately that directors now have to complete mandatory identity verification.
More on company registration in the United Kingdom

Singapore. The headline rate is 17% and has not moved since 2010. For a new company it tells you very little: the Start-Up Tax Exemption applies for the first three Years of Assessment, exempting 75% of the first S$100,000 of chargeable income and 50% of the next S$100,000. That takes up to S$125,000 a year out of tax. A company with S$200,000 of profit pays around S$12,750, an effective rate of roughly 6.4%.

After that the Partial Tax Exemption takes over and the effective rate climbs towards 17%. Singapore rewards you early and gets gradually more expensive as you scale.

The detail people miss: the three-year window runs from the first Year of Assessment after incorporation, not from your first profit. A company incorporated eighteen months before the product launches burns half the relief on nothing. Timing your incorporation has a price attached.

The system is single-tier: tax paid by the company is final and dividends are not taxed again in the shareholder’s hands in Singapore. There is no second layer of the kind a US C-Corp carries.

Under Budget 2026, a 40% corporate income tax rebate applies for YA 2026 with a combined benefit cap of S$30,000, and active companies employing at least one local employee receive a minimum benefit of S$1,500.
More on company registration in Singapore

Hong Kong. Two-tiered profits tax: 8.25% on the first HK$2 million of assessable profits and 16.5% above. There are no additional startup reliefs, but the rate does not climb as you grow. The logic is the mirror image of Singapore’s: Singapore rewards the young, Hong Kong rewards those who have scaled.

One material constraint for AI projects: Hong Kong is not on OpenAI’s list of supported countries. This is not a local prohibition but the provider’s own decision. If your product is built on OpenAI models, a Hong Kong structure creates a problem you will end up solving with a second entity.
More on company registration in Hong Kong

A European entity makes sense when your customers are in the EU and want a supplier with an EU VAT number. A German corporate buyer purchasing from a Delaware company handles the reverse charge itself. Plenty of them would simply rather not.

Cyprus. This is where nine out of ten articles on the subject need correcting. From 1 January 2026 the Cypriot corporate rate is not 12.5% but 15%. The reform was passed by parliament on 22 December 2025, published in the official gazette on 31 December and has applied since the start of the year. The driver was alignment with the OECD global minimum tax: at 12.5%, large groups were topping up the missing 2.5% in other countries anyway, and Cyprus decided to collect it instead.

The reform also brought changes that matter more to an AI business than the rate does. The 120% super-deduction on qualifying R&D expenditure has been extended to 2030. The loss carry-forward period has gone from five years to seven, and up to ten in certain circumstances, which is a genuine asset for a startup running at a loss for several years. The special defence contribution on actual dividends has been cut from 17% to 5%, and the deemed dividend distribution rules have been abolished.

Read it this way: the headline rate went up, while the total burden across the full life cycle of a profit is unchanged or lower for many structures.
More on company registration in Cyprus

Ireland. A 12.5% rate on trading income and the best research incentive in Europe: Budget 2026 raised the R&D tax credit from 30% to 35%. For deep tech with heavy development spend that is significant. The European headquarters of the major technology companies are also physically here, which makes hiring easier.
More on company registration in Ireland

Lithuania. From 1 January 2026 the standard rate rose from 16% to 17%, and the reduced rate for small companies from 6% to 7%. But there is something more interesting for a startup: a new company with fewer than ten employees and annual revenue below €300,000 pays 0% for its first two years, subject to conditions. For a project that spends two years building a product before it earns anything, that is effectively an exemption for the entire development period.
Lithuania also introduced immediate depreciation of long-term assets from 2026, including computer hardware and software, provided they are used in the business for at least three years. One limitation worth knowing about: carried-forward losses can now offset no more than 70% of current taxable profit.
More on company registration in Lithuania

A note on Estonia. It is almost always the first name mentioned in any conversation about technology companies and e-Residency. The model genuinely is unusual: retained profit is not taxed at all, and tax arises only on distribution, at 22/78 of the net amount. The planned increase to 24% was cancelled in December 2025 and the defence tax was scrapped as well, though VAT has been 24% since July 2025.
An honest caveat: Estonia’s strength is the reinvestment of profit, and an early-stage AI startup usually has no profit to reinvest. So the main argument stops working precisely when you would want it. E-residents also run into refusals when opening accounts with local banks if the company has no real connection to the country.

The most expensive misconception in this area goes: “we will incorporate in Delaware, so European regulation does not apply to us.” It does.

The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) applies extraterritorially, on the same logic as the GDPR. Obligations arise where a system is placed on the EU market or its output is used in the EU, regardless of where the provider is established. A company in the US, the UK or Singapore selling a scoring service to a European bank is a provider under the regulation, with everything that follows. A third-country provider must also appoint an authorised representative in the EU: Article 22 for high-risk systems, Article 54 for general-purpose AI models.

The context matters here. In November 2025 the European Commission proposed a simplification package known as the Digital Omnibus on AI. Negotiations were difficult: the first trilogue on 28 April 2026 ended without agreement, political agreement was reached on 7 May, the European Parliament endorsed the text on 16 June and the Council of the EU on 29 June 2026.

The outcome: obligations for stand-alone high-risk systems under Annex III move from 2 August 2026 to 2 December 2027, and those for systems embedded in regulated products to 2 August 2028. National regulatory sandboxes are deferred to 2 August 2027.

But not everything moved, and this is the detail the headlines lose. The Article 50 transparency requirements apply from 2 August 2026, which is days away. Users must be able to tell that they are interacting with an AI system. The obligation to mark AI-generated content in a machine-readable format was separately deferred to 2 December 2026, with a three-month transition for systems placed on the market before 2 August.
The prohibitions in force since February 2025 and the general-purpose AI obligations in force since August 2025 were not deferred at all. The revised text also adds a prohibition on systems generating non-consensual intimate imagery and child sexual abuse material.

Penalties

• Up to €35 million or 7% of worldwide turnover for prohibited practices
• Up to €15 million or 3% for breaches relating to high-risk systems
• Up to €7.5 million or 1% for supplying incorrect information to a regulator

The higher of the two figures applies. Proportionate caps exist for small and medium-sized enterprises. A fine can be directed at the European subsidiary or the authorised representative where the company itself sits outside the EU.

The practical takeaway for a founder: where you incorporate determines your corporate and tax position, but it does not take you outside product regulation. Regulation follows the market you sell into. If there are no European users and none planned, the question goes away. If they are planned, it needs solving at the product design stage rather than after the first enterprise contract.

The item founders postpone and investors check first.
If your code was written by freelancers or contractors without a proper assignment of rights, those rights may still sit with them. Legal due diligence before a round finds this immediately, and it either destroys the valuation or kills the deal. You need an agreement with everyone who touched the product, assigning the exclusive rights to the company outright rather than licensing them.

The second layer is training data. Obligations for general-purpose AI model providers include a copyright compliance policy and publication of a sufficiently detailed summary of the training data used. A dataset assembled by scraping without regard to restrictions turns from technical debt into legal debt.

The third layer is model weights as an asset. If you are fine-tuning someone else’s model, check the licence on the base weights: some open licences restrict commercial use or require derivatives to be disclosed. A company whose principal asset rests on incompatibly licensed weights is worth less than its founder thinks.

The sequence for a Delaware C-Corp, with the duration of each stage. The logic is the same in other jurisdictions; the timings differ.

Step 1. Structure review, 1-3 days. Who the shareholders are and in what proportions, where the funding comes from, where the founders are tax resident, whether a round is planned. A mistake at this stage costs more than all the others combined.

Step 2. Name check and document preparation, 1-2 days. Register search, certificate of incorporation, bylaws, allocation of founder shares. Par value at $0.0001.

Step 3. Filing the Certificate of Incorporation, 3-5 business days. Same-day processing is available for an additional fee.

Step 4. EIN application, four to eight weeks without a Social Security Number. This is the longest stage and it runs in parallel with everything else. Start it immediately after incorporation.

Step 5. Opening a bank account, 1-4 weeks after the EIN arrives. The decision rests with the bank and nobody can guarantee the outcome. We prepare the pack and support the application.

Step 6. Payment processing setup, 3-10 days. Requires a live account and a working website with published terms and a privacy policy.

Step 7. Setting up the compliance calendar. Franchise tax and annual report by 1 March. Form 1120 and Form 5472 by 15 April.

From start to first payment received, expect eight to twelve weeks on average, more than half of which is the EIN and the bank.

Incorporating before deciding on the market. The company gets set up because a company is needed, and then it turns out the customers are in the EU and want a VAT number. Now you need a second entity and a structure built after the fact.
Ignoring the EIN timeline. The round is promised to the investor in a month, and the account cannot be opened because the number arrives in six weeks.
Optimising for a low rate on zero profit. Savings on paper, losses in infrastructure.
Authorised shares with no par value set. A five-figure franchise tax bill and a panic instead of $400.
No assignment of rights over the code. Found during due diligence, priced in equity.
Assuming that incorporating outside the EU removes the AI Act question. It does not. Regulation follows the user.
Forgotten filings for a dormant company. The project is on hold, the entity keeps accruing obligations and penalties. Form 5472 not filed: $25,000.
Setting up the option plan after the round. Reserving equity for the team is cheaper before you raise, not after dilution.

The honest section that people selling incorporations usually leave out.
A business with no international customers. If all the revenue is domestic and customers pay in local currency, a foreign structure adds cost and filings and gives nothing back.
A founder who is not prepared for the filings. A foreign company means annual returns, accounting to local standards, hitting deadlines and reporting in your country of tax residence. Anyone hoping to incorporate and forget about it should not start: a dormant company creates obligations exactly as an active one does.
Anyone looking for anonymity. It does not exist in 2026. Beneficial ownership registers, automatic exchange of information, director identity verification, source of funds checks at account opening. Structures that promise opacity lead to a refused account and blocked payments.
Projects still at the idea stage. Until there is a prototype and the first users, an entity is a premature expense. The exception: you are already in conversation with a fund that requires the wrapper.

Choosing a jurisdiction for an AI business is not a comparison of tax rates. It is a sequence of three questions: where your customers are, where the money to grow will come from, and what infrastructure the product needs in order to work at all. The answers narrow the list to one or two options, and only then is there any point in doing the tax arithmetic.

Finextwin is an international corporate services firm with offices in Hong Kong and Tbilisi. Over 6+ years we have registered companies in 30+ jurisdictions for more than 1,250 clients and we work with 65+ banking partners. Book a consultation to go through the structure for your product, your market and your fundraising plans. A manager responds within 30 minutes.