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

NVIDIA published NeMo Switchyard on 11 August 2026 as a fully open-source model-routing proxy for LLM applications, licensed Apache 2.0.

Why it mattersIt is the first vendor-backed router that is trivial to drop in front of an existing agent, which is what turns model selection from a design-time choice into a runtime one.

nvidiatechn2026route · github2026nvidianem

Switchyard can route at the individual-task level or across phases of solving a single task; for multi-turn agent workloads NVIDIA states the router "may route each full request to a specific model or route at each step".

Why it mattersPer-step routing means a single user task can be served by two providers. Any control scoped to "the model we use" is scoped to the wrong unit.

nvidiatechn2026route

The Switchyard server accepts OpenAI Chat Completions, Anthropic Messages and OpenAI Responses API requests, and forwards to upstream targets including vLLM, NVIDIA NIM, Ollama or any OpenAI-compatible endpoint.

Why it mattersAdding a provider to the pool is a config edit, not an integration project. The technical cost of a new sub-processor has fallen to roughly zero.

nvidiatechn2026route · github2026nvidianem

Switchyard records the selected model, the decision rationale, token usage, latency and call outcomes for each routed call.

Why it mattersThe evidence a controller needs already exists at the router. Whether it is persisted next to the request, rather than aggregated into metrics, is a deployment decision.

nvidiatechn2026route

The Switchyard repository describes the project as pre-alpha experimental software, states it is not for production use, and warns of API changes before v1.0.

Why it mattersWorth quoting to anyone who wants to ship it this quarter — the vendor's own maturity statement is the cheapest risk argument available.

github2026nvidianem

LangChain's Switchyard benchmark ran 145 multi-step agentic tasks averaging 6.3 model calls each, routing between NVIDIA Nemotron 3.5 Lightning and Claude Opus 4.8, with Gemini 3.1 Flash Lite as judge.

Why it mattersSets the scope of every number quoted from it: 145 tasks with an LLM judge, not a production workload.

langchain2026how

Routing cut total cost by 74% against a frontier-only baseline — $0.026 per completed task against $0.092 for Claude Opus 4.8 alone.

Why it mattersThe comparator is what makes it a number. 74% against frontier-only is the whole commercial case for putting a router in the path.

langchain2026how · nvidiatechn2026route

Routed accuracy was 80.0% against 86.0% for Claude Opus 4.8 alone — a 6-point reduction.

Why it mattersThe accuracy of the deployed system belongs to the pool, not to either model. Neither vendor's card states 80.0%.

langchain2026how

7% of calls went to the frontier model and accounted for 68.4% of total spend, while Nemotron 3.5 Lightning handled 93% of calls for 10.4% of spend; the judge model consumed a further 21.2% of routed spend.

Why it mattersEscalated calls are ~10x the average token weight. The requests leaving for the frontier provider are the long, context-heavy ones — which is also where the sensitive payload usually is.

langchain2026how

LangChain notes its benchmark workload was relatively saturated, with just 8 points of variance between the two models, which limits how much value routing could demonstrate.

Why it mattersThe source's own caveat. It cuts both ways: on harder work the saving grows and so does the accuracy loss.

langchain2026how

NVIDIA's post reports that Cognition's Devin Desktop implementation routed between Opus 5 and Kimi K2.7, achieving 50.6% at a $3.11 mean cost — within 2.8 percentage points of Opus 5 accuracy at approximately 28% lower mean cost.

Why it mattersA shipped product routing across two providers in different jurisdictions. Attributed to NVIDIA's post, not to Cognition directly.

nvidiatechn2026route

DeepSeek published the DeepSeek-V4-Pro GA release note on 13 August 2026, adding native OpenAI Responses API support and one-click Codex integration.

Why it mattersResponses API compatibility is exactly the interface Switchyard forwards to, so a new provider became a drop-in upstream two days after the router shipped.

deepseekapi2026deepseekv

GDPR Article 28(2) provides that a processor shall not engage another processor without prior specific or general written authorisation of the controller, and Article 28(3)(a) binds processing to documented instructions including as to third-country transfers.

Why it mattersA routing pool is a set of sub-processors. Adding one is an authorisation event whether or not anyone treated it as one.

intersoftco2026processor

Under GDPR Chapter V, a transfer of personal data to a third country may take place only under the conditions of that Chapter; absent an adequacy decision under Article 45(3), Article 46 requires appropriate safeguards such as the Commission's standard contractual clauses.

Why it mattersThe transfer basis is per destination. One pool spanning several countries needs several bases, not one.

intersoftco2026transfers · intersoftco2026safeguard

GDPR Article 30(1) requires the record of processing to state the categories of recipients including those in third countries, and, where applicable, the transfers made to a third country and the identification of that country.

Why it mattersThe Article 30 record is the concrete Monday artefact: every model in the routing pool is a recipient, and each needs its country named.

intersoftco2026records