AI GlossaryㄹTechnical words in the news
routing algorithm
A set of rules that decides which AI model or server an incoming request should be sent to — essentially traffic control for requests.
In plain words
A routing algorithm works like a dispatcher at a call center. When a call (request) comes in, the dispatcher sends it to whichever agent is free right now, whichever agent knows that issue best, or to a specific agent based on a fixed order. In AI services, multiple models or servers sit in the place of those agents, and the routing algorithm decides which request goes where.
This kind of mechanism is needed because if a tool is built to always point at one single agent, it becomes hard to swap in a different agent behind the scenes. For example, you can keep your usual coding assistant running on the surface while, behind it, branching logic sends requests to a cheaper model depending on the situation, routes sensitive code only to an internal company server, or automatically switches to another model when responses keep failing. This branching logic itself is the routing algorithm, and it can range from something as simple as round-robin ordering to something as sophisticated as picking based on evaluated conditions.
How it shows up in the news
An article about NVIDIA's Switchyard introduces it as providing "typed routing algorithms," explaining that developers can write their own rules to split traffic across multiple models or to call different models at different stages based on signals. A common misunderstanding here is that the routing algorithm itself generates new answers — it doesn't. It only decides who the request gets sent to. The actual answer is always produced by whichever model was selected as the routing destination.
Try it yourself
You can get a feel for the concept without writing any code. Try asking your usual coding agent or chatbot: "Design a routing algorithm that splits requests across multiple AI models. Requirements: 1) automatically switch to another model when the failure rate rises 2) always send sensitive content to one specific model only 3) otherwise prefer the cheapest model. Show me pseudocode that handles these three conditions." This will give you a concrete sense of how routing rules are really just a combination of conditional statements.
See also
Stories using this term
- NVIDIA releases Switchyard, an LLM routing proxy for coding agentsAI · 2026.08.20
- Starcloud raises additional $250M for orbital data centersBusiness · 2026.08.22
- NVIDIA invests $1.5 billion in SoftBank-linked data center developerBusiness · 2026.08.18
- NVIDIA Confirms Official Guide for Connecting Two, Three, and Four DGX Spark UnitsAI · 2026.09.05
- NVIDIA pays startup Poolside $6 billion to license its AI modelBusiness · 2026.08.22
- Firebird Opens AI Factory in ArmeniaBusiness · 2026.08.09
