Data Stored in Taiwan Is Not the Same as Data in Your Hands
Data residency is not data sovereignty. What Taiwan's and the EU's timelines require you to prove before 2027.
As AI moves into core operations, the real question is not where the data sits, but who has the final say when something goes wrong
When signing a cloud contract, companies usually check a few things carefully: which data center the data sits in, whether it is encrypted, how service levels are calculated. Very few ask: if we want to replace this in three years, how long would it take? That question doesn't matter most of the time. It starts to matter once AI reaches core operations — because what you would be replacing is no longer a service, but something that has grown into the business itself. Taiwan's most heavily regulated industry recently put this very plainly.
“We cannot supervise the party receiving it”
The Bankers Association is preparing to raise a request in its financial policy white paper: financial institutions cannot directly supervise large international cloud providers, and those providers are not going to restructure global operations or amend contracts for an individual institution. The association therefore proposes compensating through internal controls and third-party audits, rather than being unable to use these services simply because the contract cannot be changed. Put plainly: the data has been handed over, but the party receiving it is beyond reach. This is not unique to banking. Banking was just the first sector forced to say it out loud. That other industries have not yet been asked the question does not mean they are free of it.
The question is not “where,” it's “who can oversee it”
When enterprises discuss data sovereignty, the first question is usually whether data leaves the country. That question points in the wrong direction. As legal practitioners have noted, regulators have not in fact prohibited financial institutions from using offshore cloud services. What regulators are genuinely concerned about is whether their supervisory authority is compromised once data passes to a third party. That is the difference between data residency and data sovereignty. The first answers which machine the data sits on. The second answers who has the right to access it, who can supervise it, and who is accountable when something goes wrong. The U.S. CLOUD Act is one example: it allows U.S. authorities to require a cloud provider headquartered in the United States to produce data within its control, regardless of where that data is physically stored. None of this means international cloud providers are untrustworthy. It establishes one thing only: “stored in Taiwan” does not answer “who can oversee it.” Taiwan's own regulators have been asking the second question for years. The FSC's rules on cloud outsourcing by financial institutions require that institutions retain full ownership of outsourced data, and that on termination they must be able to migrate to another cloud provider or bring the workload back in house. In other words, the question nobody asks at signing — how long would it take to replace this — is one regulators have been asking all along.
A global reallocation is under way
This question is redirecting budgets worldwide. Gartner forecasts worldwide sovereign cloud infrastructure spending will reach $80 billion in 2026, growing 35.6% year over year, with mature Asia-Pacific among the fastest-growing regions at 87%. Gartner estimates this geopatriation demand will move roughly 20% of existing workloads from global providers to local ones, with the primary buyers being governments, followed by regulated industries and critical infrastructure. Taiwan is on the same path. At its first meeting, the Executive Yuan's National AI Strategy Committee elevated sovereign AI from a technology project to a core national objective.
The timetable is already set
For Taiwanese enterprises, this is not a distant trend. It is already on the calendar. The first date has passed. The Basic Act on Artificial Intelligence took effect on 14 January this year. Under the Executive Yuan's implementation plan, government agencies were to complete risk assessments of official AI use by July this year, and to establish internal control frameworks within a year. The second date is January 2028. By then, each competent authority must issue management rules and industry guidance covering AI risks in its own domain. Finance, healthcare and manufacturing each have their own regulator — and the FSC is already planning to bring agentic AI into its guidance for the financial sector. The third date applies only to some. For enterprises operating in the EU, obligations for high-risk AI systems take effect on 2 December 2027. What all three dates require is not a policy statement but evidence: whether you can explain what happened, whether you can replace a supplier, whether you can demonstrate that someone was supervising. And evidence cannot be reconstructed after the fact. That is why 2027 is a year for preparation, not observation.
“So let's bring it all back in house”
Read this far, and that is the easy conclusion. It is not the right reading. The same Gartner forecast notes that 80% of sovereign cloud spending comes from new digital initiatives or legacy systems not yet migrated — not from a large-scale exit from public cloud. What is happening is a split in workloads, not a retreat. Nor does sovereignty mean building everything yourself. Almost every “sovereign” offering still rests on the same chips and the same supply chain, and some locally branded cloud services remain subject to the laws governing their parent company — which is the origin of the “sovereignty washing” debate. Frankly, “sovereignty” has been used rather loosely of late. But the thing behind the word is real — and for most enterprises it means something far more concrete: where this data is processed, who can access it, what record remains when something goes wrong, and whether there is a second option when a supplier changes terms tomorrow. These questions do not require a hundred billion dollars. They require architectural decisions.
Three questions to ask now
First, which data must never be processed by a third party under any circumstances? Is that line written into policy, or enforced by the system? Second, if your primary AI supplier changes pricing, changes terms, or discontinues service next year, how long would it take to switch? If the answer is “we don't know,” that is the risk. Third, when a regulator or auditor asks you to reconstruct a specific AI decision, is the evidence in your hands or your supplier's?
All three articles are about the same thing
The first article in this series was about not being able to see what AI costs. The second was about not being able to control what AI does. This one is about not being able to switch away, or to supervise. On the surface, those three questions belong to finance, security and compliance respectively. In substance they are one question: whether a company can be accountable for a system it did not build and cannot fully see into. The market has given this a name: sovereignty. It is not a national posture, and it is not an argument against the cloud. It is the boundary of responsibility. Inference moving on-premises, agents moving into production, regulators beginning to ask for evidence — these three arrive together in 2027, and all three want the same thing: a record that can be explained. The dividing line in 2027 will not be who used the stronger model. It will be who can explain, when asked, what their AI actually did. There is no shortcut to that, and it cannot be reconstructed after the fact — but it can start next week.