The same model, same weights, same quality, is now also sold with a location promise: your data stays in the region. OpenAI, AWS Bedrock, and Google Vertex all sell region-pinned inference endpoints, and the pinned SKU prices at a roughly 10 percent premium over the global SKU. Sovereignty stopped being a compliance checkbox. It became a line item with a margin.
The tariff analogy
International trade has a standard device for this exact situation: the border premium. The same steel, copper, or microchip sells for more on the far side of a border - not because the product changed, but because crossing a jurisdiction has friction. Customs, compliance paperwork, localized inventory, duplicated certification. The seller prices that friction into the regional price, and lets buyers who do not need the formalities pay the lower global price.
Data residency pricing has the same shape:
- The pinned SKU is real, not cosmetic. A genuine region-pinned endpoint runs inference on in-region compute, with the model replicated in-region, keys held in-region, and telemetry scoped to the region. It is a second copy of an expensive plant, not a proxy setting.
- The 10 percent is replication cost. A provider that replicates the model and the operational stack across N regions charges the regional price against the regional copy. The global endpoint keeps its cheaper, denser utilization. The premium is roughly the regional fixed cost divided by regional volume - and it compresses as volume grows.
- Why not 50 percent? Because the pinned SKU has no scarcity. The weights already exist; the marginal invite of one more pinned call is just another regional inference, and the provider competes internationally on the same model. The premium is recovery of the replication cost, not a monopoly rent - which is exactly why it sits near single-digit-plus-zero rather than double its listing.
The accounting question: regulated or not
The buyer's decision is binary per traffic class, and the error is expensive in one direction only:
- Regulated traffic (personal data, sector-regulated content, a contract that names a jurisdiction): pay the premium. When a contract or a law names where data may live, the global endpoint is not a discount option - it is an exposure. The 10 percent is cheaper than the regulatory notice.
- Commodity traffic (public-content summarization, fixed-prompt classification, dev tooling at work): the premium is insurance with no policyholder. Paying 10 percent on this traffic forever is a tax on a form nobody asked for.
Most production mixes are blends: a minority of requests are truly regulated, a majority are commodity. The unmanaged default - all requests going to the one endpoint you configured first - pays the premium on the whole spend. A clean routing split recovers it.
The routing split
The mechanics at a gateway are unglamorous:
- Tag traffic classes - regulated versus commodity - from contracts and sector rules, not from whichever endpoint was configured first.
- Regulated class goes to the region-pinned endpoint for the jurisdiction in question.
- Commodity class goes to the best global rate for the needed model tier.
- Track the premium as an explicit cost line: effective rate = global rate x 1.10. It is a monthly policy choice, not a constant - and the premium itself moves both ways as regional volumes grow.
One caveat: re-verify the residency claim whenever the provider's documentation changes. "Data does not leave the EU" and "inference happens in the EU" are different statements, and providers have historically refined exactly this wording. The compliance-side mechanics are covered in the EU AI Act routing checklist; this post is about the hundred-dollar question the checklist assumes.
What to do
- Enumerate which of your request classes are legally regulated (by contract, by sector), and tag them in the router.
- Route only tagged classes to region-pinned endpoints; keep everything else on the global rate.
- Budget the residency premium explicitly - about 10 percent of pinned-SKU spend - and re-verify the ratio at every vendor price-page refresh.
- Store the provider's residency claim, dated, as part of the control evidence. The claim is a control; the date is the audit.