Mercury Rack
A 48U composable rack platform with 72 double-wide PCIe Gen5 GPU slots, any vendor's accelerators on a vendor-neutral PCIe Gen5-over-optical fabric, up to 1.2 PB of composable NVMe per PRU, and 111 kW in a standard air-cooled cabinet.
One Rack, Every GPU, Yours to Compose
Every rack-scale AI platform shipping in 2026 asks you to make the same bet: pick one GPU vendor, accept the compute-to-memory-to-storage ratio someone else set at the factory, re-plumb your data center for 100% direct liquid cooling, and live with that decision for the life of the asset. The Mercury Rack takes the opposite position. It is a 48U composable platform with 72 double-wide PCIe Gen5 slots, and you decide what goes in them — including changing your mind later, in software, without a forklift.
Contact UsAny GPU. Any vendor. Any ratio.
The 72 slots take standard double-wide PCIe Gen5 accelerators, which means the shopping list is yours: RTX 6000, H200, MI350, MI210, R9700, Intel accelerators, and whatever ships next. NVIDIA and AMD silicon can sit in the same rack, on the same fabric, serving different workloads — with CUDA and ROCm available side by side rather than one excluding the other.
That is the part fixed monoliths structurally cannot do. A 72-GPU NVL72 is 72 B200s. A Helios rack is 72 MI455X. Mercury is 72 slots.
What is in the rack
Mercury is a full 48U cabinet, laid out to keep the GPU pool as large as possible while leaving room for the hosts and management that make it usable: 9x PRU 2500 shelves holding the accelerators and storage, 72 double-wide PCIe Gen5 slots forming the composable pool, 8U of host servers at an 8:1 GPU-to-CPU ratio, a 1U management switch, and 3U reserved for expansion. The result is up to 111 kW in a standard 19-inch rack — air-cooled, with in-rack direct liquid cooling available as an option rather than a prerequisite.
The fabric: PCIe Gen5 over optical
Composability is only interesting if the pooled devices perform like local ones. Mercury connects hosts and PRUs over a PCIe Gen5-over-optical fabric, giving full Gen5 bandwidth between any GPU pair in the rack — with no proprietary interconnect domain drawing a boundary around which GPUs can talk to which.
It is a vendor-neutral fabric built on the standard every accelerator already speaks, which is what makes mixing silicon possible in the first place.
The trade, stated plainly
Side by side against the 2026 rack-scale platforms, the trade is straightforward. Mercury gives up the proprietary scale-up interconnect and gains vendor choice, air cooling, mixed software stacks, and a lower entry price. If you are pre-training frontier models on a single stack, a monolith is built for you. For inference, fine-tuning, and everything downstream of it, the fabric that does not lock the rack is the one worth having.
Mercury vs. NVL72, Vera Rubin NVL144, and Helios MI455X
Storage that composes too
Each PRU 2500 can be populated with accelerators or with NVMe — up to 40 SSDs and 1.2 PB per PRU. Those drives are composable on the same fabric, which makes in-rack KV-cache offload and GPUDirect Storage practical without a separate storage rack sitting next to the AI rack.
For long-context inference, where the KV cache is the thing that actually runs out, that capacity is in the same cabinet as the GPUs consuming it.
Composer: The Rack Is Software
Corespan Composer is the control plane that turns the pool into shapes. It assigns devices to hosts at runtime — attach eight GPUs to one host for a fine-tune, break them apart into inference workers an hour later, add NVMe to a host that needs scratch space — without cabling changes, BIOS reboots, or downtime. It sits beneath Kubernetes and Slurm rather than replacing them, so the schedulers you already run keep working. This is the difference between a rack you buy and a rack you operate: the ratio of GPUs to hosts to storage stops being a purchasing decision made once and becomes a scheduling decision made continuously.
Contact UsKey Features
72 Double-Wide PCIe Gen5 Slots
A single 48U cabinet exposes 72 composable accelerator slots across 9x PRU 2500 shelves — the largest usable GPU pool in the rack, not 72 of one vendor's parts.
Vendor-Neutral Accelerator Pool
Populate the rack with NVIDIA, AMD, and Intel accelerators in any mix — RTX 6000, H200, MI350, MI210, R9700 and more — with CUDA and ROCm running side by side.
PCIe Gen5-Over-Optical Fabric
Full Gen5 bandwidth between any GPU pair in the rack, on the standard every accelerator already speaks, with no proprietary interconnect drawing a boundary around the pool.
Composable NVMe Storage
Fill any PRU 2500 with up to 40 SSDs and 1.2 PB of NVMe, composable on the same fabric for in-rack KV-cache offload and GPUDirect Storage.
Air-Cooled 48U Design
Up to 111 kW in a standard 19-inch rack with air cooling, keeping in-rack direct liquid cooling an option rather than a data-center prerequisite.
Composer Control Plane
Recompose GPUs, hosts, and storage at runtime with no downtime, beneath the Kubernetes and Slurm schedulers you already run.
Use Cases
Inference and Fine-Tuning Fleets
Attach GPUs into a fine-tune, break them into inference workers an hour later, and rebalance the pool as demand shifts across the day.
Agentic and RAG Serving
Pair pooled GPUs with in-rack composable NVMe for long-context and retrieval-heavy pipelines that keep the KV cache next to the compute.
Mixed-Silicon Research Clusters
Run NVIDIA and AMD accelerators on one fabric so teams can benchmark, port, and serve across CUDA and ROCm without separate clusters.
Neocloud GPU Services
Deliver GPU-as-a-Service with higher utilization and flexible tenant allocation from a single vendor-neutral pool.
Are you an SI/VAR interested in selling the Corespan Mercury Rack platform?
Partner Information