Your GPU's memory isn't the ceiling anymore.
Physically separate RAM across networked machines becomes one address space, presented to a GPU as native VRAM. Zero modification required.
running across three consumer machines on a standard 1Gbps home network. The kind of hardware most teams already own, not a special cluster bought for the demo.
See the Proof →total hardware cost against the nearest alternative
to the GPU runtime, driver, inference app, or model artifact
cross-vendor pooling, architecture confirmed against live mesh data, physical hardware validation in progress
Every machine has memory sitting idle. MemWaxing makes it one machine's problem to have enough.
A model that needs more VRAM than any single GPU in the mesh has doesn't have to wait for bigger hardware. Physically separate RAM on other machines gets pooled into one address space and handed to the GPU as if it were local. The GPU never knows the difference, because nothing about how it reads memory has to change.
That's the actual patent claim: transparency. Not a faster way to move data between machines. A way to make the boundary between them disappear entirely, from the GPU's point of view. Building the future on the backbone of today's technology, not next year's hardware refresh.
Read the full mechanism →Pooling alone isn't the whole story. Meridian makes it fast by overcoming the latency.
A naive pool just moves data on demand, and pays a network round-trip every time. Meridian is MemWaxing's prefetch scheduler: it watches access patterns and stages the next likely block on the local node before the GPU asks for it. The mesh isn't just bigger. It's paying attention.
That's flipping the AI buildout script. Instead of buying more GPU to get more speed, the scheduling layer does the work an expensive card would otherwise have to.
MemWaxing applies anywhere a workload outgrows a single machine's memory. It's not about any one device. The machines have to be networked together for the pool to exist at all. Sovereign, self-hosted, no cloud dependency, no third-party inference service in the loop.
Discuss Licensing