What a national AI budget changes for your business

Every few months a government publishes an AI number with a lot of zeros in it. South Korea’s proposed 2027 budget is a clean example: 821 trillion won in total spending, about $599 billion, with 21.3 trillion won of that going to three megaprojects the government has named as semiconductors, physical AI and AI data centers, up 97.2 percent from this year [1]. Roughly $15.5 billion, aimed at chips and AI, in one country, in one year [2]. Then you go back to your two contractors, your Claude subscription and your API bill, and the announcement sits there looking important without telling you what to do.

Almost none of that money reaches a small operator, and none of it reaches you in the form the headline implies. What does reach you is narrower and more useful than the number: a short list of settings in your own account, and a shorter list of rules that bind you whether or not you noticed them. This guide is about finding both quickly. It is not for anyone with a compliance department, a data-residency clause already negotiated into a customer contract, or a decision to make about where to site a data center. Those readers have lawyers, and the lawyers are correct.

A budget line is a supply signal, not a product you can buy

Look at what the money in Korea’s proposal is actually for. Semiconductor funding rises from 1.3 trillion won to 3.4 trillion won, and spending on semiconductor production infrastructure specifically rises tenfold, from 200 billion won to 2 trillion won [2]. A separate 2.6 trillion won goes to physical AI demonstration projects, and the budget minister said the government will supply 10,000 graphics processing units, plus data and workers, to help a national AI foundation model achieve globally competitive performance [1].

None of that is a product with a signup page. It buys fabrication capacity, demonstration sites, research staff and hardware for a state-backed model. The effects that eventually reach a solo operator are indirect and slow: more memory chips in circulation, more data-center capacity in a particular time zone, possibly a domestic model you could call in three years if it turns out to be good.

There is a second reason not to plan around the figure. It is a proposal. Under the National Finance Act the government must submit the budget to the National Assembly 120 days before the fiscal year starts, and the ruling and opposition parties are expected to fight over it once it arrives there [1]. A number that has not passed a vote is a statement of intent with a press release attached. Treat it the way you would treat a vendor roadmap slide: useful for direction, useless for scheduling.

The rules arrive years before the capacity does

Spending programs run on multi-year timelines. Obligations do not. The European Union’s AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026, with exceptions that landed earlier: the prohibited practices and AI literacy obligations from 2 February 2025, and the governance rules and general-purpose model obligations from 2 August 2025 [8]. The transparency rules came into effect in August 2026, and the Commission said it would start enforcing them on 2 August [8].

The high-risk provisions are the slow part, and they got slower. The AI Omnibus, a simplification package adopted on 19 November 2025 and in force since 27 July 2026, extended the rules for high-risk use cases in sensitive areas under Annex III to 2 December 2027, and the rules for high-risk systems embedded into regulated products under Annex I to 2 August 2028 [8].

Read those two sections next to each other. Korea’s money is a proposal for a fiscal year that has not started. Europe’s transparency duty already applies. The Act’s disclosure obligations say that when people use AI systems such as chatbots they should be made aware they are interacting with a machine, that providers of generative AI have to ensure AI-generated content is identifiable, and that deep fakes and text published to inform the public on matters of public interest have to be clearly and visibly labelled [8]. Whether any of that lands on you depends on your role and your market, which is exactly the question worth ten minutes of your afternoon.

That asymmetry is the durable lesson. Announcements about capital are slow, conditional and mostly about someone else. Announcements about rules are fast, and they carry dates. So when a country makes news for its AI budget, the first thing worth checking is not the budget. It is whether the same government has passed anything that changes what you must disclose, log or refuse. The budget is the part that gets covered. The rule is the part that has a compliance date on it.

Where your prompt is processed is a different setting from where it is stored

Here is the part you can act on today. Every major vendor now lets you pin part of your traffic to a geography, and every one of them distinguishes between storing your data somewhere and running the model there. The distinction is the whole game.

OpenAI’s data residency is configured per project within your API organization, and you send traffic to a prefixed domain such as us.api.openai.com, eu.api.openai.com or kr.api.openai.com [3]. Ten regions are listed. Only three of them, the United States, Europe including the EEA and Switzerland, and the United Arab Emirates, support regional processing. Australia, Canada, Japan, India, Singapore, South Korea and the United Kingdom are marked for regional storage only, and OpenAI states plainly that if your selected region does not support regional processing, it may also process and temporarily store your customer content outside that region to deliver the services [3]. Using any non-US region also requires that you be approved for abuse monitoring controls and execute a Modified Retention amendment [3].

Sit with that for a second. South Korea is proposing $15.5 billion for chips and AI [2], and on OpenAI’s own residency table it is a storage-only region [3]. Sovereign compute capacity and a selectable inference region are different things, and they arrive years apart.

If you do not touch an API at all, and most solo operators do not, the equivalent controls live in the subscription products and they are about retention rather than geography. OpenAI says it does not use your business data for training its models by default, that it may do so if you have explicitly opted in, that workspace admins can control how long your data is retained, and that deleted conversations are removed from its systems within 30 days unless it is legally required to retain them [4]. That is the realistic set of levers for a seat-based plan. Pinning a country is an API-tier decision, and if someone is asking you to guarantee where their data is processed, you are being asked to move onto that tier and pay for it.

Google draws the same line more explicitly. Its deployments and endpoints page carries a flat warning that endpoints don’t guarantee data residency or in-region ML processing [6]. The residency page then explains the split: data stored at rest in the location you chose remains at rest in that location, independent of which endpoint your request calls, while the geographic location of the model computation is determined by your choice of endpoint [5]. Jurisdictional multi-region endpoints keep processing inside a boundary such as the United States or the European Union, and Google notes that the EU endpoint strictly covers residency within EU member states, excluding the United Kingdom and Switzerland [5]. Locational endpoints such as europe-west1 keep processing within the jurisdiction associated with that region, and South Korea’s asia-northeast3 appears in the per-model table for Google models alongside Japan, Singapore and India [5]. Global endpoints route and process data anywhere globally, and Google says they don’t provide regional isolation or data residency guarantees [5].

One more line in that documentation is worth your attention. Google publishes a separate table for partner models, the one that covers Anthropic’s Claude models on Google Cloud, and its columns are the US multi-region, the EU multi-region, Belgium, the Netherlands, Singapore, Taiwan and Global. South Korea is not among them [5]. Which vendor you use changes the answer, so check your own vendor’s table rather than the one you read about.

Residency costs about 10 percent, and sometimes a model generation

Both prices are published, and they agree. OpenAI charges a 10 percent uplift on data residency endpoints for models released on or after 5 March 2026 that are eligible for data residency [3]. Anthropic’s documentation for Claude on Google Cloud says regional and multi-region endpoints include a 10 percent pricing premium over global endpoints, applying to Claude Sonnet 4.5 and future models, while global endpoints carry no pricing premium [7] and, on Google’s own account, no residency guarantee [5].

The second cost is less obvious and more annoying. Anthropic’s page notes that specific regional endpoints support Claude Sonnet 4.6 and earlier, and that newer models use the global or multi-region endpoints [7]. Google’s list works the same way, with per-model residency support varying by location [5]. If you pin to a single country, you are choosing from the models that have finished rolling out to that country, not from the ones announced last month. For most small operators that trade is fine. For anyone whose product depends on the current frontier model, it is a real constraint that shows up as a surprise during a launch.

Before you switch anything, price it. A 10 percent uplift on a small bill is a rounding error, and a 10 percent uplift on a large one is a hire.

How to read the next announcement in ten minutes

The method generalises past Korea, and past 2027. Start by establishing whether the figure has been appropriated or merely proposed, because a proposal still has to survive a legislature [1]. Then name the recipient. Fabs, demonstration projects, GPU procurement and university programs are all real spending that does not become anything you can purchase [1][2].

Next, go to your provider’s residency or locations page and check what that country actually offers you today: processing, storage only, or nothing [3][5]. That single lookup converts a geopolitical headline into a fact about your account. Then check whether the same jurisdiction has a rule with a compliance date, because that is the part that can create work for you this quarter [8]. Finally, write down what would have to change for you to revisit the decision, such as your provider adding regional processing, or a customer contract requiring it. Without that trigger you will re-litigate the same question every time a similar headline appears.

The lookup is quicker than it sounds because the vendors publish it in one table each. OpenAI lists every region with a column for regional storage and a column for regional processing [3]. Google lists models against locations, with separate columns for the US and EU multi-regions [5]. Anthropic describes three endpoint types, global, multi-region and regional, and tells you which one each model generation supports [7]. Ten minutes with those three pages will answer, definitively, whether a country that just made headlines is a place your requests can run.

Notice what is missing from the method. There is no step where you estimate whether the country’s AI industry will succeed. You are not investing in it. You are deciding where your requests go and what you have to disclose, and both of those are answerable from documentation in an afternoon.

What still goes wrong

Region lists move faster than guides do. Every fact above came from a vendor page or a government page on the day this was written, and the residency tables in particular change with each model release. Check the source, not the summary, before you configure anything. The same applies to the legal timeline, which was amended once already by the AI Omnibus, in force since 27 July 2026 [8].

Residency is not privacy, and the vendors say so in the same documents that sell it. OpenAI’s data residency does not apply to system data, which it defines as account data, metadata and usage data that do not contain customer content [3]. Abuse monitoring logs are generated for all API feature usage by default and retained for up to 30 days, unless longer retention is required by law or is reasonably necessary to protect the service or a third party, and zero data retention is subject to prior approval by OpenAI and acceptance of additional requirements rather than being a checkbox you tick [3]. Choosing a region tells you where computation and storage happen. It does not tell you that nothing is logged.

The last failure is over-reading. A budget headline invites you to form a view about which country wins, and that view will not change a single line of your configuration. A country can run one of the largest AI spending programs on the planet and still be a place where your requests cannot be processed. South Korea is currently both. Between those two facts there is a gap of several years, and you live in the gap.