Loading today’s stories
Build4 distinct publishers3 min readPublished
PAIR proxies Ollama and LM Studio, so the agent keeps seeing one connection and no harness code changes. The adoption cost moves to disk, because a node is only eligible if it already has the exact model downloaded.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The 1.6x is the only performance number in the announcement, and it describes NVIDIA's workload. Five subagent requests, two nodes, the Qwen3.6 35B A3B model on a pair of RTX 5090 machines with 32 GB [12]. Because a request is assigned to one node and stays there for its lifetime [7], the placement problem is list scheduling. If the five calls cost roughly the same and the single-node baseline runs them one after another, the best possible split is three on one machine and two on the other: makespan falls from five units to three, which is 1.67x [18]. Measured result, about 1.6x [12]. That is roughly 96 percent of the bound the topology allows [19], which is a well-behaved scheduler.
The same arithmetic caps the upside. A third machine with that model on disk would take the makespan to two units, or 2.5x [21], and a fourth buys nothing at five requests. For the number to transfer, four things have to hold: the subagent calls are independent and similar in cost, there are more of them than nodes, every candidate node already has the exact model, and nobody has reclaimed the GPU for something interactive.
The third condition is the one with a price tag. PAIR routes only to a machine with the required engine enabled and the exact requested model available, and installing the same model on several machines is what gives the router options [9]. Eligibility is therefore a function of disk, not of idle silicon. It also does not pool VRAM or merge GPUs [8], so a model too large for any single node is still too large.
The design decision worth praising is the proxy. Rather than publishing a cluster API and asking every harness to integrate, PAIR sits behind the Ollama and LM Studio interfaces the agent already speaks, identifies the engine and model requirements, and picks a node while the caller keeps seeing one connection [3][6]. The New Stack frames the target as agents like NemoClaw, OpenClaw and Hermes running more subagents in parallel [17]. The supported pool is broad: GeForce RTX 20 Series and newer, RTX PRO workstation GPUs from Turing on, DGX Spark, and Apple M4 or newer silicon [10].
Home hardware is the awkward part, and NVIDIA says so. Nodes join when ready and drop away when they power down or hibernate [22], and when a user reclaims a GPU the local inference engine is stopped [14]. What neither writeup states is what happens to a request already executing on a node that leaves. Given that placement is for the request's lifetime, that is the first thing I would test.
One more figure, from The New Stack rather than NVIDIA: an RTX 5090 currently goes for about $5,000 against a $2,000 initial MSRP [13]. The reference configuration for a feature about using hardware you already own therefore holds roughly $10,000 of GPU [20].
Ranked by verification strength, evidence, and original report placement.
On Thursday, Nvidia launched the Nvidia Personal AI Router (PAIR), an open source software router for the home network that lets idle Macs and PCs run small models on demand and speed up agentic workflows with subagents.
The NVIDIA PAIR beta is available for supported Windows, macOS, and Linux systems through graphical and terminal interfaces.
PAIR routes each independent inference request to an available system on the home network and works with familiar local inference services including Ollama and LM Studio, so no agent harness changes are necessary.
PAIR is not a new inference engine; Ollama or LM Studio still runs the model on the selected machine.
PAIR discovers local systems with mDNS, pairs supported devices, tracks whether each system is ready for a request, schedules independent jobs, and returns each response to the originating application.
PAIR receives the request through its proxy, identifies its engine and model requirements, and selects one eligible node; that node executes the request from start to finish and sends the response back through PAIR, while the agent continues to see one connection.
Distinct publishers with included, body-backed reporting in this cluster.
2 articles · September 3, 2026
1 article · September 3, 2026
2 articles · September 3, 2026
1 article · September 3, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
Nvidia's agent sandbox holds; the model server underneath it answers to any browser tab1 distinct publisher
build
Apple's local-AI Mac Studio starts at $5,499, and the price that matters is still blank7 distinct publishers
product
Nvidia's PAIR spreads one agent's model calls across whichever home PCs are idle1 distinct publisher
security
NemoClaw's wide Ollama bind lets a webpage rewrite your model's chat template2 distinct publishers
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Mechanism documented, performance vendor-shaped
Every mechanical detail — mDNS pairing, the placement checklist, the rule that a request never leaves the node it lands on — traces to Nvidia's developer post and product page. Tom's Hardware and The New Stack relay it faithfully, but neither ran the software. And the one figure a buyer would act on doesn't survive comparison: Nvidia times three devices from 18 minutes to about 8, while The New Stack cites two RTX 5090 boxes at 1.6x. Solid on how it works, thin on how well.
Downloadable beta, no outside mileage
There is something to install, which beats a slide, but the entire usage record is one day old and belongs to the vendor. The only timed workload ran on a DGX Spark, an RTX Spark laptop and a 5090 — a bill of materials almost no household has. No third-party install, no community timings, no disclosed user counts, and The New Stack itself says it remains to be seen what a Mac Studio and a couple of Mac minis would deliver.
Cluster in the headline, scheduler in the body
Tom's Hardware's headline joins every GPU in your home into a cluster; Nvidia spends a paragraph insisting nothing is joined — no pooled VRAM, one request on one box, start to finish. That the seller is the one deflating the metaphor keeps this gap small. What pushes it positive is arithmetic nobody prints: with calls pinned to nodes, five subagents across two machines can never beat 1.67x, so the advertised 1.6x is a ceiling being touched, not a trend line. More speed means more machines, each holding its own copy of the model.
Two of four voices are the seller
The router costs nothing; what it costs is a second, third and fourth Nvidia-class machine on your LAN, each with the same weights on disk. Half of this reporting is Nvidia writing about Nvidia, and the independent half is enthusiast hardware press whose audience buys GPUs. Only The New Stack breaks frame long enough to mention that the demo's 5090 now goes for about $5,000 against a $2,000 MSRP — and to doubt the average household has several.
Sound mechanics, one-sided numbers
I would repeat the architecture to anyone without hedging: it is stated consistently in four places and the constraints are unusually candid. I would not repeat a speedup. The origin is single-vendor, the reporting is a day old, no one outside Nvidia has run it, and the two published performance accounts describe different hardware and different margins for the same workload.