THE NEXT AI BATTLEGROUND: MODEL ROUTING

If you've never (knowingly) used an AI model router - or if you don't know what one is - that's probably about to change.

Until the last three months, model routers were for the optimisers - that hardcore amongst us who wanted each query to be sent to the exactly right slice of digital intelligence. Whether to minimise the cost. Or to maximise the effectiveness.

But whilst tokens were in plentiful and cheap supply via subscriptions, there was no need for the rest of us to get involved. We slammed in our work to the best frontier models and got exactly the results we were looking for. All month long.

Now - as subscriptions become less generous and the cost of paying for frontier models via pay-per-use tokens is a reality for many - routers look set to become part of everyday operations.

And if Beijing decides to limit access to the best Chinese models, which appears to be under discussion, the pressure will only build.

All of which suggests the second tier of western players - including Mistral, Meta and xAI - might be reasonably expecting something of a resurgence.

And, perhaps, the prevalence of non-optimisers (economists call them 'satisficers') meant we all failed to spot where the real market power will lie.

Google won the internet by owning the layer that decided where search traffic went - and charging for the best positions. Perhaps model routers will capture ever-larger slices of agentic revenue the same way: by directing where the workloads go.

This turn of events also makes you think that perhaps the optimisers were on to something?

Maybe we should listen to them more often. They were, after all, the ones that gave us Swiss watches.