Liquidata designs, builds and operates GPU capacity on powered sites in Europe and the United States.
Liquidata is the provider of record for its customers: it signs the capacity contracts, procures and installs the accelerators, and operates the service. It deploys on sites that developers and energy companies have already brought to the point of deliverable power, because in 2026 the grid connection, not the hardware, sets the timetable for new capacity. The current platform is NVIDIA's Vera Rubin NVL72. Site designs are sized for the 600 kW racks NVIDIA has published for 2027, and AMD Helios or custom accelerators are deployed where a customer's software stack requires them.
Why each site is designed for the accelerator generation after the one being installed.
NVIDIA has moved to an annual product cadence: Blackwell in 2024 and 2025, Vera Rubin in 2026, Rubin Ultra in 2027, with a further architecture announced for 2028 [1]. Each step has raised the electrical and thermal load of a rack. Surveys by the Uptime Institute put the average rack in a conventional data center in the single-digit kilowatts [2]. A Vera Rubin NVL72 rack draws about 190 kW at NVIDIA's standard setting and up to about 230 kW at maximum performance [3]. The Kyber rack announced for Rubin Ultra is specified at about 600 kW, and NVIDIA has published an 800-volt direct-current distribution architecture for racks of that class, planned from 2027 [4].
A data hall, its power train and its cooling plant take roughly two to three years to design, permit and build, and are expected to serve for fifteen to twenty years. If they are sized for the racks being installed today, the next generation requires them to be rebuilt. Liquidata therefore specifies each module for the generation that follows the one it installs: floor loading, busways, switchgear and cooling capacity are sized for 600 kW racks, and halls are laid out so that one generation can replace another without rebuilding the site. The margin is carried mainly in busway and switchgear ratings, floor loading and cooling-plant capacity, whose cost is small relative to the building, the grid connection and the accelerators themselves; the engineering is described in the four areas below.
Accelerator generations
Modules of 50 to 100 racks are specified so that the next generation can replace the current one in place. The current platform is Vera Rubin NVL72: 72 GPUs in one NVLink domain, cooled by one liquid loop per rack. The design basis for 2027 is the Kyber rack at about 600 kW. AMD Helios (72 MI455X GPUs per rack, liquid cooled) and custom accelerators are deployed where a customer's stack requires them; the power and cooling specification of the site does not change [1][5].
Distribution rated for 600 kW racks
At 600 kW per rack, conventional low-voltage alternating-current distribution needs conductors and breakers that no longer fit beside the rack, which is why NVIDIA's published architecture moves to 800 V direct current from the facility to the rack [4]. Liquidata's power trains are specified for that voltage class, with busways, switchgear and short-duration storage sized for the step changes in load that synchronized training runs produce.
Heat rejected at density
A 190 kW rack cannot be cooled by air; the NVL72 is cooled entirely by liquid at a 45 °C supply temperature [3]. That temperature allows heat to be rejected to dry coolers without chillers for most hours of the year in most of Europe and North America, which is why a design PUE of 1.25 and near-zero water consumption follow from the choice of platform rather than from a separate program.
Buildings for the loads that follow
Prefabricated power and cooling blocks; halls engineered for the floor loading of racks that weigh more than a tonne each; and campuses with their own generation or storage where the grid cannot deliver the load on the customer's timetable. The site, not the rack, is the long-lived asset, so it carries the design margin.
Satellite spectrum as a second path for the control plane
Fiber carries the data. A campus of 100 MW exchanges data with its customers at rates measured in terabits per second, which no satellite link approaches. What a satellite link can carry is the control plane: monitoring, remote operations and customer access, which amount to a few megabits per second. Liquidata specifies that path as a second, independent route for the control plane when a fiber route fails, and as an interim link for sites that are energized before their second fiber route is complete.
Quantum technologies, limited to what is deployable
Two uses are current. NIST published its first post-quantum cryptography standards in August 2024, and the European Commission's coordinated roadmap of 2025 asks member states to begin the transition by the end of 2026 and to protect high-risk uses by 2030 [6]; Liquidata's platform adopts those standards ahead of the deadlines, and quantum-safe links are provided where a customer requires them. The site designs also leave room to host quantum hardware beside GPU clusters for the hybrid workloads now being developed. Claims that quantum methods will relieve the network's own scheduling and routing bottlenecks are not yet supported by deployed hardware; Liquidata follows that work and does not build on it.
What Liquidata sells, and in what units.
Liquidata is the provider of record: the customer's contract is with Liquidata, which procures the hardware, operates the service and is accountable for its service levels. Capacity is delivered in modules timed to each site's energization date, and modules are deployed in campuses, for the reasons given below.
Dedicated and managed clusters
Whole Vera Rubin NVL72 racks and multi-rack clusters for AI developers, enterprises and research organizations, on commitments of 12 to 36 months, delivered with the fabric, storage, monitoring and operations around them. The commitment length follows the customer's planning horizon. The hardware's useful life, which operators now report at five to six years, is longer than any single commitment, so a module serves several customers over its life [7].
Private deployments
Capacity that stays in-country on a partner site, inside the customer's security boundary and under its controls, for governments, research institutions and regulated industries. In the European Union, NIS2 and DORA require such organizations to control where their data and models reside and who operates the systems that hold them [8]; a private deployment is designed to meet those requirements rather than to approximate them.
Ten modules, about 100 MW, contracted as one
A module is 50 to 100 racks, 3,600 to 7,200 GPUs and 10 to 20 MW of IT load. Liquidata deploys modules in campuses of ten, about 100 MW of IT load, because that is the scale on which the largest buyers of capacity plan. Each campus is contracted before it is built and delivered against confirmed power; the order in which that happens is explained on the platform page, and the terms of each engagement are provided to customers under confidentiality.
Why Liquidata deploys on other parties' powered sites, and what the site owner receives.
The International Energy Agency estimates that data centers consumed about 415 TWh of electricity in 2024 and projects about 945 TWh in 2030, with artificial intelligence the largest contributor to the increase [9]. Grid connections of the size an AI campus needs take years to obtain across the European Union and the United States, and most sites that already have one are committed. The developers and energy companies that hold such sites have generally sold or leased them to large operators for a fixed price, after which everything the compute earns accrues to the operator.
Liquidata offers those owners a second route: a compute business with its own customers, built and operated on the owner's site rather than bought out of it, so that the owner's involvement does not end on the day of a sale or lease. The four layers of a site are owned and financed separately and deal with each other at arm's length, so that neither party has to take on the other's business. The model is described on the Energy to Compute page; the terms of each site partnership are agreed site by site under confidentiality.
Why cooling determines the design of the data center at these densities.
Air cooling is limited by the heat that air can carry through a rack at a tolerable velocity, which in practice ends at a few tens of kilowatts per rack. The NVL72 rejects about 190 kW through direct-to-chip liquid cooling at a 45 °C supply temperature, which ASHRAE classifies as warm-water cooling [3][10]. Warm water matters for three reasons. First, heat at 45 °C or more can be rejected to the outside air through dry coolers for most of the year in most of Europe and North America, so chillers are needed for peak hours at most. Second, dry coolers consume no water in normal operation; where adiabatic assist is used on the hottest days, water use is confined to those hours. Third, water at that temperature can be raised by heat pumps to the temperatures district heating networks require, so the heat has a use.
Liquidata's work to date is in liquid-cooled, high-density infrastructure: thermal management, integrated systems and the monitoring software that reports cooling energy, availability, density and component health. The 2027 generation, at about 600 kW per rack, makes these the only workable choices rather than the preferred ones.
The sites under discussion, and the condition each must meet.
Liquidata's pipeline consists of powered sites held by developers and energy companies in the European Union and the United States, described here by type; sponsors, locations and capacities are shared under confidentiality. Sites are sequenced by the date on which power can be delivered and by where customers want capacity. A campus proceeds only where at least 100 MW of IT load can be energized within 24 months under a signed connection agreement with dates, because that is the scale and the horizon on which the largest buyers of capacity plan.
Technology-park sites with a grid supply assurance and a phased connection, and a regional expansion platform behind them.
400/220/110 kV
Brownfield industrial platforms with current grid approvals and room to expand, adjoining transmission substations, with on-site gas-to-power options and storage integration.
Gas-to-power options
A campus with on-site combined-cycle generation and substation, and powered land with as-of-right data-center zoning, power available now and a utility study for expansion.
Utility studies
An engineering-led developer of digital infrastructure, founded in 2023.
Liquidata, Inc. was founded in 2023 in Miami Beach, Florida. Its premise is that the data center is now the binding constraint on the deployment of artificial intelligence, and that it has to be engineered to the published roadmap of the hardware it houses rather than to the conventions of the previous decade. The senior team covers platform and AI engineering, data-center design and build, security, capital and structuring, public markets and product.
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, GTC keynotes and published product roadmap (March 2025 and March 2026): annual cadence of Blackwell, Blackwell Ultra, Vera Rubin, Rubin Ultra and a further architecture for 2028.
- Uptime Institute, Global Data Center Survey (2024 and 2025 editions): average rack density in surveyed facilities.
- NVIDIA, Vera Rubin NVL72 product specifications (2026): 72 Rubin GPUs, 36 Vera CPUs, 45 °C inlet, about 190 kW per rack at the standard setting and up to about 230 kW at maximum performance; full production stated from January 2026.
- NVIDIA, 800 VDC power architecture for next-generation AI factories (announced May 2025, planned from 2027) and the Kyber rack for Rubin Ultra at about 600 kW.
- AMD, Helios rack specifications (2026): 72 Instinct MI455X GPUs per rack, liquid cooled.
- NIST, FIPS 203, 204 and 205 (August 13, 2024); European Commission, coordinated implementation roadmap for the transition to post-quantum cryptography (June 2025).
- Public filings of listed GPU cloud operators (2025 to 2026): accelerator useful lives of five to six years.
- Directive (EU) 2022/2555 (NIS2) and Regulation (EU) 2022/2554 (DORA).
- International Energy Agency, Energy and AI (April 2025): data-center electricity consumption of about 415 TWh in 2024 and a base-case projection of about 945 TWh in 2030.
- ASHRAE, liquid cooling guidelines for data-center equipment (water classes W17 to W45+): classification of 45 °C supply as warm-water cooling. Air-cooling limits and dry-cooler operating hours are Liquidata's engineering estimates for European and North American climates.