Why data stays put | AlienGiraffe
Research

Why data stays put.

Banks and telcos keep their most important data out of the cloud for reasons a platform can't change: mainframes, localization rules, cost, and the people who own production. Every figure here links to its source, and we say when a source is a vendor.

Platform vendors

Databricks describes the gap itself.

In June 2026, Databricks wrote that enterprises "cannot — and will not — move all of their data to the cloud", and that Tier-1 banks have adopted "Hybrid Forever" strategies. It also reported requests from hundreds of customers for on-premises and hybrid connectivity to Unity Catalog.

Legacy

Core banking still runs on mainframes, and will for years.

Modernization happens on the mainframe as often as off it. Senior executives interviewed by Coalition Greenwich prefer to migrate about 5% of applications a year.

Rules

Some data is required by law to stay in the country.

The firmest rules are outside Europe. India and Saudi Arabia have the clearest localization requirements; in Europe the driver is sovereignty and the US CLOUD Act; in the US, it's mostly legacy.

  • India's RBI requires payment system data to be stored in a system only in India. Processing abroad is allowed only if foreign copies are deleted within one business day or 24 hours.

    RBI circular, 2018 (summary) ↗
  • Saudi banks should use, in principle, only cloud services located in the Kingdom, with explicit SAMA approval for services outside it.

    Second-hand: from Governance Docs' guide to the SAMA Cyber Security Framework (subdomain 3.4.3). We couldn't verify the primary text.

  • EU traffic-data retention for telcos runs 6 to 24 months depending on the member state.

    Cisco ↗
  • Most operators cannot freely aggregate or centralize network data, which is why federated approaches are being proposed for telco AI.

    arXiv 2506.17259 ↗

Today's tools

Ingestion copies the data. Federation sends it through the cloud.

Both are good tools for data that may move. Neither is built for data that can't leave the network, even briefly.

  • The Oracle CDC connector doesn't support Oracle RAC (including Exadata RAC), standbys, or read replicas; change capture needs the primary.

    Databricks docs ↗
  • It needs archive-log mode, supplemental logging, and a user with LogMiner privileges on the source.

    Databricks docs ↗
  • Federation queries are read-only, pushed to the source on every run, with no result or disk cache.

    Databricks docs ↗
  • Federation logs in with one stored credential, so the source sees a service account.

    Databricks docs ↗
  • Databricks compute for federation always deploys in the cloud.

    Databricks docs ↗
  • Log files on SFTP servers can be ingested, which copies them into the lakehouse.

    Databricks docs ↗

The people waiting

The cost shows up as waiting.

There are no public time studies of DBA queues at banks or telcos. The general figures point the same way.

By region

Different reasons in each region.

The reasons differ by region. The result is the same: a list of sources that won't go to the cloud.

India

Hard localization rules for payment data (RBI).

Gulf

In-Kingdom cloud by default for Saudi banks, under SAMA guidance.

Europe

Sovereignty, and concern about the US CLOUD Act.

United States

Mostly legacy and mainframe, not law.

Counter-evidence

Regulated data does move, sometimes.

Lumen Technologies moved its 133 TB call detail record warehouse from an on-prem platform to Databricks SQL on Azure. Outside the localization jurisdictions, regulators rarely forbid the cloud outright; the barrier is usually internal risk appetite, legacy systems, and cost. If your data can move, move it. This research is about the data that can't.

Databricks customer story ↗