The processor is now a runtime decision — model routing, sub-processors and GDPR Chapter V

NVIDIA open-sourced NeMo Switchyard on 11 August 2026, a routing proxy that selects which model — and therefore which provider — serves each call, per request or per step inside a single task. LangChain's benchmark on 145 multi-step agentic tasks cut total cost 74% against a frontier-only baseline ($0.026 per completed task against $0.092) for a 6-point accuracy loss, 86.0% to 80.0%, sending only 7% of calls to the frontier model — calls that carried 68.4% of the spend. The practitioner consequence is not the cost curve. A load balancer selects another replica of the same processor; a model router selects a different legal entity, in a different country, under a different contract. GDPR Article 28(2) requires the controller's authorisation before a processor engages another, and Chapter V requires a transfer basis per third country. A routing pool moves both decisions from procurement time to runtime, resolved by a cost heuristic — and the calls it escalates are the long, context-heavy ones.

Post

NVIDIA open-sourced a model router on 11 August. It moves the choice of which company processes your prompt from procurement to runtime.

NeMo Switchyard is a proxy. It speaks OpenAI Chat Completions, Anthropic Messages and the Responses API, and forwards upstream to vLLM, NIM, Ollama or anything OpenAI-compatible. It routes per request — or at each step inside a single task. Apache 2.0, and NVIDIA's own repo calls it pre-alpha and not for production use.

The numbers are why people will ship it anyway.

LangChain benchmarked it on 145 multi-step agentic tasks averaging 6.3 model calls each, routing between Nemotron 3.5 Lightning and Claude Opus 4.8. Total cost fell 74% against a frontier-only baseline — $0.026 per completed task against $0.092. Accuracy fell from 86.0% to 80.0%. Only 7% of calls reached the frontier model, and those 7% were 68.4% of the spend.

Now the European part.

A load balancer picks another replica of the same processor. A model router picks a different legal entity, in a different country, under a different contract. GDPR Article 28(2): a processor may not engage another processor without your authorisation. Chapter V: a transfer to a third country needs its own basis — adequacy, or standard clauses plus an assessment.

NVIDIA's own post reports Cognition routing Devin Desktop between Opus 5 and Kimi K2.7. Two providers, two jurisdictions. DeepSeek shipped native Responses API support on 13 August, which makes it a drop-in upstream too.

And look at which calls escalate. Seven percent of the calls carried 68.4% of the spend — the escalated ones are the long, context-heavy ones. Those are the prompts most likely to hold something you would rather keep on one continent.

What I'd do Monday:

  1. Persist the selected model beside the request ID. Switchyard records it — Prometheus is not a record of processing.
  2. Pin the pool. Deny by default. A new upstream is a DPA question, not a config change.
  3. Put every model in the pool into your Article 30 record and your sub-processor list, with its country.
  4. Quote accuracy per route. 80.0% belongs to neither model.

Routing is real engineering. It also turns a contract into a heuristic.

Which model answered your last support ticket, and where does it live?

#AgenticAI #GDPR #AIEngineering #AIGovernance #Berlin

AI disclosure: the narration voice, the cover art and the brand ident animation in this video are AI-generated. The script, the claims and the source checks are mine.

First comment

Sources, primary first —

NVIDIA Technical Blog, 11 August 2026 — Switchyard itself: the routing granularity ("route each full request to a specific model or route at each step"), the recorded fields, and the Cognition Devin Desktop result https://developer.nvidia.com/blog/route-ai-agent-workloads-across-models-with-nvidia-nemo-switchyard/

The repository, if you want to run it rather than read about it — Apache 2.0, and its own README says pre-alpha, not for production use https://github.com/NVIDIA-NeMo/Switchyard

LangChain, 11 August 2026 — the benchmark all the cost numbers come from: 145 tasks, 6.3 calls each, 74%, $0.026 vs $0.092, 86.0% to 80.0%, 7% of calls for 68.4% of spend https://www.langchain.com/blog/switchyard-agent-routing-benchmark

NVIDIA newsroom, 11 August 2026 — Nemotron 3.5 Lightning, the cheap end of that pairing https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/

DeepSeek API docs, 13 August 2026 — the GA note adding native Responses API support, which is what makes it a drop-in upstream https://api-docs.deepseek.com/news/news260813/

GDPR Article 28 — paragraph 2 is the sub-processor authorisation, paragraph 3(a) the documented-instructions duty that includes transfers https://gdpr-info.eu/art-28-gdpr/

GDPR Article 44 — the general principle for transfers to third countries https://gdpr-info.eu/art-44-gdpr/

GDPR Article 46 — appropriate safeguards where there is no adequacy decision https://gdpr-info.eu/art-46-gdpr/

GDPR Article 30 — the record of processing, including categories of recipients and third-country transfers https://gdpr-info.eu/art-30-gdpr/

One point cut for length: LangChain is honest about the limits of its own number. It notes the workload was relatively saturated, with just 8 points of variance between the two models, which caps how much routing could have shown. On a harder task set the cost saving and the accuracy loss would both be larger — and it is the accuracy loss that ends up in your Article 15 claim, not the saving.

Claims

Sources

Corrections

None.

AI disclosure

The video for this edition contains AI-generated media: the narration is a synthetic voice and the cover art is generated. The script, the claims and the source verification are human-authored and human-checked. Assets carry IPTC DigitalSourceType and C2PA Content Credentials.