Resources

Corespan's Software Defined Machine: software decides what machine exists, carving 16 to 32 GPUs per CPU from photonic GPU and SSD pools.

Airline Economics Need a Better Airplane

A GPU neocloud runs on airline economics, and the returns live in the last 25 points of utilization. Financing cannot reach them. Matching the machine to the workload can.

The Token-Cost Trap: One Rack, Every GPU, Yours to Compose

Disaggregated inference locks the attention-to-feed-forward ratio into the silicon layout. Mercury Rack turns that ratio into a Composer policy instead.

Corespan's Mercury Rack: Composable AI Infrastructure. A vendor-neutral, composable rack platform with 72 double-wide PCIe Gen5 GPU slots, choice of NVIDIA, AMD, or Intel accelerators, up to 1.2PB of composable NVMe SSDs per PRU 2500, in-rack KV cache offload, and up to 111 kW of rack power.

Mercury Rack: Build a Bigger Computer, Not Just a Bigger Cluster

Corespan's Mercury Rack turns pooled GPUs, xPUs, SSDs, and FPGAs into larger, right-sized computing blocks through the DynamicXcelerator architecture, native PCIe over optical, and Corespan Composer.

Corespan computing-evolution curve. Capacity per system falls from 1950s-60s mainframes through mid-range, workstation, server, PC and mobile eras, then rises again through 2, 4 and 8 GPUs per CPU, and on to Corespan's composable 16 and 32 GPUs per CPU with 64 marked as a 2027 roadmap goal — pooled GPUs and SSDs joined by a photonic fabric beyond the server boundary.

What Is Old Is New Again: The Modern Mainframe Is Composable

Time-sharing made one expensive machine serve many users. Composable infrastructure turns a shared pool into many right-sized machines — the same economics, without the fixed configuration.

In the Lab: 8x AMD Radeon AI PRO R9700, One Fabric

We ran eight AMD Radeon AI PRO R9700s as a single fabric in a PRU 2500. Every GPU-to-GPU pair holds ~111 GB/s bidirectional — roughly 2x what NVIDIA's own reference architecture targets for RTX PRO 6000 fleets.