Begin with what request-level routing can arithmetically deliver. Five subagent calls spread over three nodes, with no single request ever split because PAIR assigns each call to one node and leaves it there for its lifetime [4], means the best available outcome is two waves of work instead of five, a hard ceiling of 2.5 times [3]. NVIDIA's run of Nous Research's Hermes Desktop [8] went from an average 18 minutes on one RTX Spark laptop [9] to 8 minutes 48 seconds on the three-box cluster [10], which is 2.05 times [1], or 82 per cent of that ceiling [3]. Adding two machines to the laptop bought roughly 0.52 times additional throughput each [4], and the fourth machine does worse than that unless the workload decomposes into more independent calls than the router has nodes.
PAIR is free and open source [1], which is the right price for software whose value to a household scales with the number of compatible boxes that household has already bought. Every node in the demonstration carried an NVIDIA badge: the RTX Spark laptop, a DGX Spark, an RTX 5090 desktop [6]. The router could have been built the other way, pooling VRAM so that two small cards hold a model neither fits alone, and NVIDIA says plainly that it does neither pooling nor sharding [4]. So the model ceiling is still whatever the smallest eligible node can run by itself [5], and the household that wants a bigger model still buys a bigger card.
The demonstration supports a few different readings, unequally likely. One possibility is that utilization becomes visible and the idle gaming rig earns its keep, with the effect that NVIDIA sells more silicon into homes that previously queued everything on one card [12]. Or the substitution runs the other way, and the household about to upgrade discovers the laptop it already owns absorbs the overflow, deferring the purchase. Or the per-node model ceiling [5] keeps local agent work in hobbyist territory. The composition of the demo tips me toward the first: a laptop, a DGX Spark and a 5090 [6] is a shopping list as much as a test rig, and 8 minutes 48 seconds [10] is the sort of number a retailer prints next to a second SKU.
What the release does not contain is a price. NVIDIA measured completion time and then disclaimed even that, calling the run unofficial and configuration-specific and naming workload parallelism, model, engine settings, hardware, network conditions and node availability as the variables [11]. What is tested here is 9.2 minutes, a 51 per cent reduction [2], on one simulated inbox, on hardware someone had already paid for; the proposition that multi-agent work no longer needs paid inference remains untested. The pull-through view breaks if the marginal node keeps returning about half a unit of throughput [4], because at that rate the third box is a rounding error against its own invoice, and PAIR is a latency feature for people who overbought.
