The routing layer arrives.

Perplexity has shipped the first frontier-lab product built on the idea that local compute is a layer, not an answer. It proves the thesis. It does not settle the question sitting underneath it.

At Computex in Taipei this week, Aravind Srinivas walked onstage with Intel's Lip-Bu Tan and unveiled what Perplexity calls the world's first hybrid local-server inference orchestrator. The pitch fits on a slide. A piece of software acts as an air-traffic controller for AI work, deciding in real time, mid-task, which parts of a job stay on the machine in front of you and which get routed out to frontier models in the cloud. In the demo, local models on an Intel laptop worked through confidential deal materials and decided what could stay on the device and what needed the data centre. The platform is chip agnostic, Perplexity says, running across Intel and Nvidia silicon, though the capability shown on stage is on Intel for now. Behind it sits a company whose revenue has gone from $100 million to $500 million in five months, and a cost problem Srinivas put bluntly: the largest operators are now spending on the order of half a billion dollars a month on compute.

The reason this matters here is that the idea is not Perplexity's. Brightbeam set it out clearly only recently, and crucially before any major lab had a product to point at. Their case was that Ireland does not need a cluster of frontier token factories to stay useful in AI, because a great deal of everyday knowledge work can now run on local models, with the heavier jobs routed out to frontier systems only when they are genuinely needed. They called it sufficiency. It was a sharp read, made ahead of the evidence, and Perplexity has now built almost exactly the thing they described.

So the directional argument is sound, and it has just acquired product backing from a frontier lab. That is worth saying plainly, and it is to Brightbeam's credit that they said it first.

What the orchestrator changes, and what it leaves untouched, is where the real questions sit.

It changes cost, and it changes residency for the easy work. Offloading routine inference to the device genuinely lowers the cloud bill, and for an Irish-located workforce it means more of the daily grind can sit on Irish-located hardware. The privacy dividend is real too. A model that decides on the device which parts of a confidential document never leave it is a meaningful answer to the data-residency question regulated sectors have wrestled with for a decade. This is the part of Brightbeam's case the launch confirms.

It does not change where the complex reasoning runs. By design, the orchestrator sends the hard thinking to the cloud. And the hard thinking is exactly where Brightbeam's own tax argument bites. Under the OECD substance test, profit attaches to where the value-adding work physically happens. If the value-adding reasoning is routed to a server in Virginia, the value, and the tax, route with it. Sending the summarisation home does not move that. If anything it sharpens the split, by turning the local-versus-frontier boundary into an explicit, engineered decision rather than an accident of where the user happened to log in. Srinivas was candid that the product is built for people making, in his phrase, GDP-moving decisions. The GDP-moving part is the part that goes to the cloud.

And it does nothing for the physical constraints underneath. A chip-agnostic router still depends on the same supply chain. The frontier half of every routed workload lands on GPUs whose memory is allocated, whose packaging is sold out, and whose wafer slots are committed into 2028. The data centres they sit in draw on the same grids the regulators are now rationing. Sufficiency relieves pressure on the routine layer. It does nothing for the frontier layer, and the frontier layer is where the value, the cost, and the chokepoints all concentrate.

So the honest reading of Computex is this. The routing layer is real, it is shipping, and Brightbeam called it early. The value-share question, who captures the economic output of the complex reasoning and where it gets taxed, is exactly as open as it was the day before. A clever orchestrator decides where a workload runs. It does not decide who keeps what the workload produces.

Routing is a tactic. The value share is the strategy. That is the part Ireland should be working now.