Samsung and AMD demonstrate AI servers borrowing memory from a shared CXL pool

We think the demonstration shows that a design works and tells a buyer little about what to order. The logic of the pitch is simple. A server can use only the memory installed inside it, so one machine can run short while another has memory sitting idle ^[7], and CXL lets a server running a heavy AI job borrow from a shared pool and return the capacity afterward ^[8]. We'd expect that to save money only when the heavy jobs land at different times. If they peak together, the pool has to be as large as the boxes it replaces.

That leaves three ways it can go. Buyers trim in-box memory and buy a smaller pool; buyers keep in-box memory and add a pool on top; or latency confines the pool to data that does not need speed. We'd expect the first, the one that would change how memory is bought, to need a pool that performs close to local memory. Samsung's exhibit leans toward the second: its KV-cache zone centered on CXL memory and enterprise SSDs ^[3], and KV cache is the temporary data a model keeps while generating a response ^[6].

The validation claim is Samsung's own ^[12], and the companies did not disclose pool size, latency or cost. Samsung is presenting pooling as one of three directions ^[4], so the demonstration does not signal a move away from HBM.

We'd be wrong if a server maker or cloud buyer published a configuration that cut per-server memory and added a pool.