The current platform, the unit of deployment, and how capacity is contracted.
Liquidata builds on NVIDIA's Vera Rubin platform and delivers it as dedicated, managed clusters on powered sites, with the network, storage, monitoring, security and operations around it. NVIDIA is the reference platform for the specifications and the calculator on this page. AMD Helios and custom accelerators are delivered the same way, at the same site specification for power and cooling, where a customer's stack requires them.
What a Vera Rubin NVL72 rack is, in the manufacturer's figures.
The NVL72 pairs 72 Rubin GPUs with 36 Vera CPUs in one rack and one NVLink domain, so that the 72 GPUs share memory and exchange data at 216 TB/s without crossing a network switch. That is the property that makes a single rack a training and inference unit rather than a collection of servers. The rack is cooled entirely by liquid at a 45 °C supply temperature, which allows cooling without chillers in most climates. NVIDIA reports the platform in full production since January 2026, with partner availability in the second half of 2026 [1]. The figures below are NVIDIA's published values; Liquidata's module and site designs are derived from them.
| Specification | Per NVL72 rack | Note |
|---|---|---|
| GPUs and CPUs | 72 Rubin GPUs, 36 Vera CPUs | One NVLink domain |
| NVFP4 inference | 3,600 PFLOPS | NVIDIA published figure |
| FP8 training | 1,260 PFLOPS | NVIDIA published figure |
| HBM4 memory | 20.7 TB, 1,400 TB/s | 288 GB per GPU |
| NVLink bandwidth | 216 TB/s | GPU to GPU |
| CPU memory | Up to 54 TB LPDDR5X | Vera CPUs |
| Cooling | 100% liquid, 45 °C inlet | Chiller-free in most climates |
| Rack power | About 190 kW | NVIDIA's standard setting; up to about 230 kW at maximum performance |
What a module requires from a site, derived from the rack figures.
Modules are contracted in racks and delivered against a site's power. The calculator applies NVIDIA's standard rack power of about 190 kW, a 5% allowance for scale-out network and storage, the design PUE chosen below and 8,760 hours a year. Moving the slider shows the IT load, the facility load, the annual electricity and the number of hours a 1,000 MWh storage station would carry the module [2]. At 50 racks the result is about 10 MW of IT load and 12.5 MW at the meter; a campus of ten such modules is about 100 MW of IT load and 125 MW at the meter.
190 kW per rack at NVIDIA's standard setting plus 5% for scale-out network and storage; 8,760 hours a year; 288 GB HBM4 per GPU.
What a module includes beyond the racks.
A rack is a component. What a customer contracts for is a cluster that has been designed to its workload, connected, accepted, secured and operated. The six items below are included in every module and are carried by Liquidata as the provider of record.
Cluster design and bring-up
Rack layout, power and cooling integration with the site, fabric topology, burn-in and acceptance testing against the customer's own workloads before the service start date.
High-speed fabric and storage
Scale-out networking and storage sized to the workload rather than to a catalogue, delivered with the module and operated as part of the service.
Monitoring, diagnostics and remote operations
Liquidata's own monitoring software reports cooling energy, availability, density and component health, and supports remote operations from the first day of service.
Security from site to tenant
Physical, network and tenant isolation designed into the module, for customers whose data and models cannot leave their jurisdiction or their own security boundary.
Round-the-clock operations
Operations, service levels and support carried by Liquidata as the provider of record, which is the party the customer's contract names.
Energy-aware scheduling
Workloads that tolerate scheduling, such as training runs and batch inference, can be placed around renewable output, storage and prices at sites that have them, without touching the service levels of the workloads that do not.
How Capacity is contracted, and why in this order.
Capacity is contracted before it is built. Customers commit to capacity for a defined term before the hardware for it is ordered; managed clusters are taken on commitments of 12 to 36 months, priced for the term and the module. Liquidata is the provider of record under every contract.
Hardware is ordered against signed contracts. No material hardware order is placed until the capacity is under contract and the site's energization dates are in a signed connection agreement, and no site is committed until power, permits and delivery dates are confirmed in signed agreements. The reason is that the equipment for a module is financed against the customer's contracted payments; an order placed without a signed customer and a dated connection would be a liability with nothing to fund it.
Delivery is timed to energization. Each module's delivery date is tied to its site's power delivery date, so customers contract against a date they can plan around, and the site owner is never asked to hold capacity for a module that has not been sold.
Generations. Rubin Ultra in the Kyber rack follows in 2027 at about 600 kW per rack. Contracts can be placed for either generation, and for AMD Helios or custom accelerators where a customer's stack requires them; the site specification is the same in each case.
Useful life. Operators now report useful lives of five to six years for accelerator hardware [3]. Commitment terms and the financing of each module are set within that life; the detailed terms are provided to customers under confidentiality and are not published.
What the figures on this page rest on.
Manufacturer specifications are as published; regulatory figures are as adopted or proposed on the dates given; estimates are identified as estimates. Where a figure is Liquidata's own design assumption, the text says so.
- NVIDIA, Vera Rubin NVL72 product specifications and production statements (2026).
- Calculator inputs: 190 kW per rack at NVIDIA's standard setting plus 5% for scale-out network and storage; design PUE of 1.15, 1.25 or 1.35; 8,760 hours a year; 288 GB HBM4 per GPU. The campus figures are ten times the 50-rack module.
- Public filings of listed GPU cloud operators (2025 to 2026): accelerator useful lives of five to six years.