Compatible, not just free.
A fit respects GPU model, memory, MIG slices and fabric — and names the fragmentation that strands slices nobody can sell.
GPU and AI hosting
Every GPU, sold once.Every pod, powered and cooled.
For GPU cloud and AI hosting providers and their teams — operations, sales, facilities and customer management: Rymvard knows which GPU configurations you can still commit, where and from when, checks every pod against power, feeds and liquid cooling — and shows whether each customer uses the GPUs they pay for.
Read-only. New in version 0.26 — the screens show an illustrative example estate, not a customer.
A GPU is only free if the rack can carry it.
An eight-GPU node needs its model, its memory and its fabric — and a rack with the power, the feeds and the coolant flow to run it. Selling from the GPU count alone promises capacity the facility cannot deliver.
Power and cooling first.
Every pod is placed against rack power, independent feeds, liquid flow, CDU headroom and floor load — before it is promised.
Survive one failure.
Lose one feed or one CDU: does the pod still run, run degraded, or stop? Unknown when the topology is missing, never assumed.
What Rymvard shows
From the GPU to the coolant loop.
Devices, MIG slices, pods and their resilience — each figure traceable to the reading it came from.
What you can sell, by quarter.
Eight-GPU nodes, NVL72 racks and MIG profiles free now and per quarter, after holds and weighted demand — and the first quarter you run short.
Holds with an end.
A GPU hold for a customer expires on its own; contracts ending within six months are named before they end.
Used, not just allocated.
GPU use per customer from the NVIDIA DCGM exporter through Prometheus — GPUs allocated but idle become a finding.
A billing basis in GPU-hours.
GPU-hours per customer and month as counts for your own billing — not prices.
Read-only, from your GPU estate and your facility.
Rymvard reads devices, pods and the facility around them. It never books capacity or changes a customer workload.
- GPU devices, MIG partitions, pods and reservations from CSV templates or the REST API; utilisation from the NVIDIA DCGM exporter through a Prometheus preset.
- Rack power, feeds, liquid loops and CDUs from the facility sources Rymvard already reads — Redfish, SNMP, Modbus and BACnet.
- Placement suggestions never book capacity: a person approves every hold.
- Detailed records stay in their country; customer names are visible only to the people who need them.
New in 0.26 · Workspaces
Four homes for GPU and AI teams.
GPU operations, sales, AI facilities and customer management each open Rymvard on their own home — their question, their four figures and only the findings that concern them, with AI that explains every number and suggests the next step. See the workspaces
GPU Operations
Is every GPU healthy, used and cooled?
- GPUs failed or degraded
- GPUs idle but allocated
- Devices with stranded MIG slices
- Devices without recent GPU readings
GPU Sales
Which GPU configurations can I sell, where — and from when?
- 8-GPU nodes free now
- NVL72 racks free now
- First GPU short quarter
- GPU holds expiring in 3 days
AI Facilities
Do power, feeds and liquid cooling carry every pod — and survive losing one?
- Pods that do not survive one failure
- Pods short of independent feeds
- Loops or CDUs above 85%
- Pods with unknown topology
GPU Customer Management
Does every customer get the GPUs they pay for?
- Customers below 30% GPU use
- GPU contracts ending within 6 months
- Customers waiting on GPU demand
- GPU-hours this month
Screens from the running product on its example estate. Customer names appear only for planners and admins — findings never name a customer.
Across every solution
Recovery, together.
When a site fails, workloads from every solution land on the same recovery destination. Rymvard checks them together and sets each recovery time against the restore throughput your tests measured. Joint recovery
Know which GPUs you can commit.
One GPU estate, read-only: what you can sell and from when, which pods survive a failure, who uses what they pay for.