A LocalLLaMA user is attempting to combine two computers into a single local large-model compute system: the primary machine has 94GB of DDR5 memory, an RTX 5070 Ti, and an RTX 3060, while the other has 64GB of DDR4 memory. The plan is to connect them directly over 5—10GbE (gigabit to 10-gigabit Ethernet) and use llama.cpp's RPC (remote procedure call) capability to coordinate inference across both machines and run a large MoE (Mixture of Experts) model. Our view is that this is not the release of a new product, but it accurately exposes the real bottleneck in local AI: as models keep growing, memory capacity is becoming the ceiling before compute does.
What this is
The core idea is to treat the GPUs (graphics processing units) and memory of multiple devices as one compute pool. The GPUs handle dense computation, while system memory can store some model weights. The two machines communicate over the network, with inference frameworks such as llama.cpp distributing requests and model layers between them. This does not literally merge the memory into one pool; it gives software the information it needs to understand where data resides and when it must be transferred between machines.
Industry view
What matters is that open inference frameworks already allow local compute to scale through RPC, giving individual users a path closer to a studio-grade setup and lowering the barrier to enterprise experiments with small models, local processing of sensitive data, and inference services. The risks are just as clear: network latency slows layer-by-layer inference, differences in operating systems and memory standards complicate scheduling, and added capacity does not mean added speed. If model weights must be transferred across machines frequently, a theoretical 158GB of memory may look impressive while delivering performance no better than that of a single high-memory machine.
Impact on regular people
For enterprise IT: Multi-node inference can help validate the capacity limits of local models, but production deployments must still account for networking, failover, and operations costs rather than comparing hardware prices alone.
For individual professionals: AI tooling is extending beyond local models running on a single PC toward coordinated use of multiple devices. Familiarity with inference deployment and basic networking will increasingly become a practical skill for technology roles.
For the consumer market: Beyond high-end GPUs, high-memory workstations and multi-machine interconnect solutions may attract more attention. In the near term, however, “combining two older PCs” remains an enthusiast experiment rather than a mass-market product.