Nvidia has presented a clustering tool that lets you use idle Macs and PCs to run AI models. The idea of pooling dormant machines is twenty-five years old. What makes it relevant today is the convergence of three things we have been documenting separately for months.
We have covered local AI and models that fit on a single card. Here is the next step: several machines working together.
The principle
Many organisations have underused machines: workstations switched off at night, an empty computer room, previous-generation hardware that has not been thrown away.
A clustering tool makes them appear as a single resource. Depending on the task, you can either spread independent requests across several machines, or split a model too large to fit on a single one.
These two modes are very different in practice, and you need to know that. Spreading requests works very well: each machine works alone, and the link between them matters little. Splitting a model is far more demanding, because the machines have to exchange data constantly, and a home or office network is very slow compared with the internal links of a server.
Why now
Three developments are converging, and it is their conjunction that makes this tool interesting rather than nostalgic.
Useful models have become small. We have documented this all year, from Qwen 3.8 at 27 billion to distilled models. A model that fits on a consumer card changes what an ordinary machine is worth.
Confidentiality has become a purchasing criterion. We reported that half of the Fortune 500 would favour in-house hosting. An organisation that refuses to send its data elsewhere needs compute on its own premises.
Agents consume enormous amounts. An agent running in a loop produces a continuous bill. Running it on hardware that is already amortised changes the calculation completely, as we explained regarding chips dedicated to inference.
The network is the bottleneck. For anything that requires genuine collective work between machines, an ordinary local network is a hundred times slower than what a server uses internally. Gains therefore remain limited to parallelisable tasks.
Electricity is not free. Machines running continuously consume power, and the cost can exceed that of an API depending on local rates.
The attack surface grows. We were talking about this on Friday: tools that give models access to machines have become a target. Multiplying machines multiplies the doors.
Who it is genuinely interesting for
Three profiles, and no others.
A small organisation with hardware. A practice, a laboratory, a school that owns workstations and wants to process sensitive documents without sending them out.
A developer experimenting. Running agents in a loop for hours without counting, on hardware that is already paid for.
A use case with predictable volume. A repetitive, continuous task where amortising hardware beats a subscription.
For occasional use, an API remains simpler and cheaper. No one should buy machines for that.
What to take away
This tool is interesting less for what it enables than for what it signals: compute is no longer exclusively centralised.
For two years, the trajectory went in a single direction — ever larger models in ever larger data centers. What we have been observing for a few months is a second branch: compact models, accessible hardware, and now tools for pooling what you already have.
This second branch will not replace the first. It does, however, determine who can do what without asking anyone's permission, and that is a question that goes far beyond performance.