Skip to main content

GDPR

Sub-Processors Under GDPR
What your ML vendor must disclose — and what to reject

·5 min read·By the GhostNexus team

When you approve an ML or GPU vendor, you are not just approving them. Under Article 28 GDPR, you are approving every party they hand your data to — the hosting provider, the queue, the billing processor, the email service. That is the sub-processor chain, and it is the part most reviews miss.

What a sub-processor is

A sub-processor is any third party your processor engages to help deliver the service, that may process personal data in doing so. Article 28(2) requires your processor to get your authorisation before adding one, and Article 28(4) makes the processor fully liable for its sub-processors' compliance. In practice: your vendor's cloud host, managed database, log service and payment processor are all potential sub-processors.

Why the chain — not the vendor — is what you approve

Your data does not stop at the vendor's front door. If a GPU platform runs on US infrastructure, that infrastructure provider is in your chain — and so is the jurisdiction it sits in. Approving the vendor while ignoring the chain is how teams end up with data on servers they never evaluated. The chain is the real surface area of your risk.

Red flags in a disclosure

No sub-processor list at all — or 'available on request' that never arrives
Named vendors but no location column (you can't assess transfers)
No distinction between who touches workload data and who only handles billing/email
No commitment to notify you before adding or changing a sub-processor
A hosting provider outside the EEA with no transfer mechanism stated

A transparent example

Here is how a disclosure should read — GhostNexus's own chain. Note the two things most vendors omit: a location column, and an explicit statement of what each party does and does not touch.

Sub-processorRoleLocation
Hetzner Online GmbHCompute infrastructure hostingNuremberg, DE — EU
Self-hosted RedisJob/state queueNuremberg, DE — EU
ResendTransactional email only — no workload dataUS (provider terms)
StripeBilling only — no workload dataUS (provider terms)

Workload data — your scripts, datasets, model weights, results — stays on EU infrastructure and is deleted after each job. Billing and email vendors are named honestly, with an explicit note that no workload data flows to them. That is what a DPO can actually approve.

Questions to send every ML vendor

  • List every sub-processor, its role, and its location.
  • Which of them can access workload data vs only metadata/billing?
  • Where does workload execution physically happen?
  • How and when will you notify us of a change to this list?
  • What transfer mechanism covers any non-EEA sub-processor?

General information, not legal advice. Whether a given chain fits your obligations depends on your processing and should be confirmed with your DPO or counsel.

See a fully disclosed chain in a real DPA

Generate a GDPR Art.28 DPA pre-filled with GhostNexus's complete sub-processor table — location column and all. A draft ready for your legal team.