Cruz Compute Controller, C3, manages the complete infrastructure lifecycle — from bare metal rack discovery to running AI workloads — across every GPU vendor and every scheduler, from a single control plane.
Every step from powering on a rack to monitoring a running AI training job, orchestrated through a single pipeline with autonomous agents at each stage.
C3 is architecturally split into two complementary products that cover the full Metal-to-Workload lifecycle.
C3 manages GPU infrastructure in two complementary modes for maximum flexibility and density.
A single rack BMC is the entry point. All GPUs are interconnected within the rack via NVLink Switch Trays or UALoE fabric — no inter-rack GPU networking required. Ideal for maximum-density AI training with tightly coupled GPU interconnects.
Multiple discrete GPU servers are networked together into composable platforms. Each server's BMC is discovered individually or through a K8s control plane. Inter-node fabric (InfiniBand, RoCEv2, Ethernet) is provisioned as part of the platform.
Purpose-built for the GPU datacenter era — a ground-up platform for multi-vendor, multi-scheduler GPU infrastructure.
Manage NVIDIA Blackwell, AMD MI450, and Intel accelerators from the same control plane. Your GPU vendor wins on silicon merit, not management lock-in.
SLURM, Run:ai, and KAI orchestrated through a unified abstraction layer. Mix schedulers across partitions within the same platform.
Break free from proprietary networking lock-in. Open Ethernet + C3 delivers equivalent performance at a fraction of InfiniBand-only BOM cost.
NVLink domain and UALoE fabric topology drives GPU-aware scheduling. Jobs land on physically optimal GPU groups, not random allocations.
Full MIG partition lifecycle — create, delete, query profiles. Run inference on GPU slices while training owns the rest.
Per-tenant, per-workload, per-GPU-hour cost metering. Know exactly what each team consumes and allocate GPU spend with precision.
Nine autonomous agents manage every stage of the Metal-to-Workload pipeline. They discover racks, form clusters, provision platforms, allocate GPUs, and submit workloads — continuously and without scripts.
Humans don't push buttons. They supervise. Every critical decision surfaces for approval. Agents learn from supervisory feedback through adaptive memory, getting smarter with every deployment cycle.
See Ada in ActionWalk through the full Metal-to-Workload pipeline on real GPU infrastructure — Aivres NVL72 racks, Celestica Helios with 72 AMD MI450 GPUs, and Supermicro HGX B300 clusters. Multi-vendor, multi-scheduler, one platform.
