The cloud-native world has long embraced separating compute from storage. But AI workloads are exposing a new bottleneck: the network itself.

Vendors respond by preprocessing data to cut bandwidth, or integrating RDMA and specialized hardware for faster transfers. The first still saturates the network during preprocessing. The second trades openness for hardware lock-in.

We took a third path: running user-defined WebAssembly code directly inside the storage cluster, as close to the NVMe devices as possible. Unlike SQL pushdown, computation is free-form. Unlike a compute cluster, data never leaves the storage nodes.
We will share how we used CSI and Gateway API on Kubernetes to expose this as a cloud-native, open-source capability.
Along with the dead ends we hit, the benchmarks we ran, and the security boundaries we have not yet crossed.

The project is open source. We want the community to tell us what comes next!