Your client's procurement team will ask one question before production: who else touches our data? An inference provider's data processing agreement (DPA) has to answer it with names, and has to tell you before the answer changes.

Why does this come up for you? You build the pipeline, so your client's lawyers ask you, not the model provider. If you can't show the chain of processors in writing, the deal stalls.

The reality of GDPR fines for AI pipelines

The GDPR, the EU's General Data Protection Regulation, sets two tiers of fines. Breaking the controller and processor duties in Articles 25 to 39 costs up to 10 million euros or 2% of worldwide annual turnover, whichever is higher. Article 28, the article on processors, sits in that range.

Breaking the basic processing principles, data subject rights or the rules on transfers outside the EU sits in the higher tier: up to 20 million euros or 4% of turnover.

  • Controller: the company that decides why data is processed. Here, your client.
  • Processor: a company that processes data on the controller's behalf. Here, you.
  • Sub-processor: a processor that a processor engages. Here, the inference provider, and anyone it engages in turn.

An inference call passes prompt text through every link of that chain. If one link has no contract, the whole chain is exposed.

Chain diagram of four boxes joined by arrows, left to right: your client, the controller, which authorises new sub-processors or objects; your consultancy, the processor, which passes the same duties on by contract; the inference provider, a sub-processor, with the same duties and notice before any change; and, in purple, its own sub-processors, for which your client needs a named list with each company's name, address and contact.
Your client's prompt text passes through every link, so your client needs the name, address and contact of every company in the chain, down to the inference provider's own sub-processors.

How Article 28 defines the processor chain

Article 28 sets the rules for every link.

  1. A controller may only use processors that give sufficient guarantees for technical and organisational measures (Article 28(1)).
  2. A processor may not engage another processor without the controller's prior written authorisation, specific or general (Article 28(2)).
  3. Under a general authorisation, the processor must tell the controller about any intended addition or replacement, so the controller can object (Article 28(2)).
  4. A processor that engages another must pass on the same data protection duties by contract (Article 28(4)).

The European Data Protection Board (EDPB), the body of EU data protection regulators, sharpened this in Opinion 22/2024, adopted on 7 October 2024. Controllers should have the identity of all processors and sub-processors readily available at all times: name, address and contact person. This holds regardless of how risky the processing is. The processor should provide that information proactively and keep it up to date.

Role in the chainWho it is hereWhat Article 28 asks of it
ControllerYour clientUse processors with sufficient guarantees, and authorise or object to new ones
ProcessorYour consultancyAct on documented instructions, and pass the same duties on by contract
Sub-processorThe inference providerSame duties in writing, and notice before it adds or replaces its own sub-processors

A click-through privacy notice is not enough. Article 28(3) requires a contract or other legal act that binds the processor, and lists what it must stipulate. The Irish Data Protection Commission's guidance walks through the same list.

  • Processing only on the controller's documented instructions, including for transfers outside the EU.
  • Confidentiality for everyone authorised to process the data.
  • The security measures of Article 32.
  • The conditions for engaging sub-processors.
  • Help with data subject requests.
  • Help with security, breach notification and impact assessments.
  • Deletion or return of the data when the service ends.
  • Information and audits that show compliance.

When you read a provider's DPA, check each point against its wording. Look hardest at the sub-processor clause, because that's where the chain grows.

The importance of a named sub-processor list

A clause that says a provider "may use affiliates and third parties" names nobody. Your client can't check it, and the EDPB expects names.

Ask each provider four things:

  1. The list: every company that can access the data, with name and location.
  2. The scope: which data each one gets, and for which part of the service.
  3. The notice: how many days' warning you get before a new one starts, and how you object.
  4. The way out: whether the service you buy can run without a given sub-processor.

Here is how we answer them. The list is not published on our site; you ask for the list and we send it. Which sub-processors apply depends on the service you choose, so your choice of service can rule one out. Each gets only the data its part of the service needs, under the same data protection duties we have, in writing. They sit mainly in the EU and the European Economic Area (EEA). Anything outside is covered by the GDPR's safeguards, such as standard contractual clauses (EU-approved contract terms for transfers).

We tell you before we add or replace one, and you can object within 10 business days. On 5 November 2026 we sent our customers that notice for three new sub-processors. The notice stated that our zero data retention policy, Schedule 1 of our DPA, applies in full to requests they process. It also stated that customers can always switch other providers off and run inference only on Lyceum.

Flow diagram of our sub-processor notice. We send notice before we add or replace a sub-processor. Then, highlighted in purple, you have 10 business days to object. Two arrows follow: accept, and the sub-processor works under the same duties in writing; or object, and choose a service that does not use it. Example: on 5 November 2026 we sent this notice for 3 new sub-processors.
Under our DPA you get notice before we add or replace a sub-processor and 10 business days to object, as with the 3 new sub-processors we announced on 5 November 2026.

Zero data retention and GPU memory caching

Procurement teams also ask what happens to a prompt in memory. To answer that, you need to know how a serving engine handles it.

Serving engines such as vLLM keep the model's working state for each request, the key-value (KV) cache, in GPU memory. PagedAttention splits that cache into small blocks, the way an operating system splits memory into pages. Blocks are freed and reused as requests finish.

Zero data retention means the provider never writes prompts or outputs to disk or a database. Here is what we keep at each stop of a serverless request, from our legal and data handling page:

StopWhat we seeWhat we keep
Our gatewayYour API key, the model and your prompt, in memoryA usage record: organisation, user, model, token counts and time
GPUYour prompt, in GPU memory in a European data centreRecent prompt prefixes in GPU memory, until newer requests replace them
ResponseThe answer as it streams backThe output token count, added to the usage record

The usage record never contains the prompt or the response. Nobody on our team can open your prompts, because there is nothing stored to open. For a full check, see how to verify zero data retention.

Resolving EU data residency and model routing

A DPA should match where the data actually goes. Ask where each model runs and where each sub-processor processes data, not only where the provider has its headquarters.

Our serverless models run in Paris and Finland. Dedicated endpoints and GPU virtual machines run in locations agreed per contract. Each model's page states where it runs, so check the model you plan to use. Our model-by-model guide to data residency shows how to read it.

Securing compliance & certifications for your clients

Your client's review will ask for documents, so have them ready before the first call.

  • The DPA: ours is Annex 1 of our terms, so it applies when you accept them. No separate signature is needed. You are the controller and we are the processor.
  • The TOMs: our technical and organisational measures, sent on request.
  • The sub-processor list: sent by email on request.
  • Certifications: our ISO 27001:2022 audit is in progress, and the engagement letter is open to read. The data centre operators hold ISO certifications at facility level.
  • Incidents: our DPA obliges us to tell you immediately about a breach involving your personal data. Service incidents appear on our status page.

For the steps before a proof of concept, see our DPA and customer-data checklist. When you're ready, ask for our sub-processor list.

Frequently asked questions

Is a data processing agreement required for AI inference?

Yes, when the provider processes personal data for you. Article 28(3) of the GDPR requires a binding contract with the points listed above.

Where can I get Lyceum's sub-processor list?

By email, on request. Our legal page has the request link, and the list is not published on the site.

Can my client rule out a specific sub-processor?

Often, yes. Which sub-processors apply depends on the service, and a new one can be refused within 10 business days of our notice.

Does the EDPB expect controllers to know every sub-processor?

Yes. Opinion 22/2024 says the identity of every processor and sub-processor should be readily available to the controller at all times.

Sources

  1. gdpr-info.eu
  2. gdpr-info.eu
  3. gdpr.eu
  4. edpb.europa.eu
  5. eur-lex.europa.eu
  6. dataprotection.ie
  7. arxiv.org
  8. docs.vllm.ai