Corespan Systems builds composable AI infrastructure that disaggregates GPUs from fixed servers, letting operators pool and dynamically allocate accelerators, storage, and hosts through software. Its Mercury Rack offers 72 double-wide PCIe Gen5 slots in a standard air-cooled 48U cabinet, supporting mixed-vendor silicon like NVIDIA and AMD on a vendor-neutral photonic fabric. The Corespan Composer control plane reconfigures resources at runtime without reboots or cabling changes, working beneath Kubernetes and Slurm.
- Artificial Intelligence
- Hardware
- Internet of Things
- Semiconductor
- Software Only
Funding
Funding not disclosed
Founders
Product
Problem
Traditional rack-scale AI platforms lock customers into a single GPU vendor, a fixed compute-to-storage ratio, and proprietary scale-up interconnects, often requiring direct liquid cooling and making any future hardware change a costly, multi-year commitment. This rigidity strands GPU capacity, forces over-provisioning, and prevents organizations from mixing accelerators or adapting infrastructure to evolving workload demands.
Solution
Corespan Systems provides composable, photonic-native infrastructure that disaggregates GPUs, storage, and hosts into a shared resource pool managed entirely through software. The Corespan Composer control plane dynamically assigns devices to hosts at runtime—attaching eight GPUs for a fine-tune, breaking them into inference workers an hour later, or adding NVMe for scratch space—without cabling changes, BIOS reboots, or downtime. The Mercury Rack delivers 72 double-wide PCIe Gen5 slots across nine PRU 2500 shelves in a standard air-cooled 48U cabinet, supporting any vendor's accelerators side by side on a PCIe Gen5-over-optical fabric. The platform works beneath existing schedulers like Kubernetes and Slurm rather than replacing them, and it makes the GPU-to-host-to-storage ratio a continuous scheduling decision instead of a one-time purchasing choice.
Target Audience
Primary customers are neocloud providers, enterprise AI teams, and research clusters running inference, fine-tuning, agentic serving, RAG pipelines, or storage-heavy workloads who need flexible, multi-vendor GPU infrastructure without committing to a proprietary rack-scale platform.
Features
- 72 double-wide PCIe Gen5 slots across 9× PRU 2500 shelves in a standard 48U rack, with up to 111 kW air-cooled capacity and optional in-rack direct liquid cooling
- Vendor-neutral photonic fabric using PCIe Gen5-over-optical, enabling full Gen5 bandwidth between any GPU pair and mixing NVIDIA, AMD, and other accelerators on the same fabric
- Composable storage with up to 40 NVMe SSDs and 1.2 PB per PRU, supporting in-rack KV-cache offload and GPUDirect Storage without a separate storage rack
- Corespan Composer control plane that assigns devices to hosts at runtime with no reboots or cabling changes, operating beneath Kubernetes and Slurm
- 8:1 GPU-to-host ratio with 8U of host servers, eliminating the need for a dedicated CPU per GPU
- DynamicXcelerator software for GPU isolation, elastic allocation, interconnect-agnostic deployment, and time-based scheduling across mixed GPU fleets