PRECIPIA / TECHNOLOGY EXPLORER

AI is changing.
Infrastructure must move with it.

A visual guide to the compute, connectivity and thermal systems behind modern AI environments.

These are manufacturer technologies and engineering concepts we study to inform infrastructure planning. They are distinct from Precipia’s developing proprietary systems.

01 / RACK-SCALE COMPUTE

NVIDIA Vera Rubin.
A system beyond the chip.

NVIDIA’s platform integrates compute, networking, storage and rack infrastructure. Its DGX Vera Rubin NVL72 documentation describes a 72-GPU domain, with 18 compute trays and 9 NVLink switch trays. [1]

NVIDIA manufacturer illustration showing five rack-scale systems in the Vera Rubin platform
NVIDIA Vera Rubin rack-scale systems · manufacturer illustration. Image: NVIDIA. [2]

Why it matters for a facility

Accelerator selection is inseparable from power delivery, cooling, service access, network layout and commissioning. A rack is an integrated operational system, not simply a collection of servers.

What Precipia evaluates

Workload requirements, equipment-vendor constraints, phased power needs, physical logistics, thermal compatibility and the specialist teams required for delivery.

NVIDIA Vera Rubin NVL72 rack showing compute and NVLink switch trays
Vera Rubin NVL72 rack · NVIDIA. [2]
NVIDIA NVLink switch tray manufacturer illustration
NVLink switch tray · NVIDIA. [2]

02 / CONNECTIVITY

Move data.
Without losing the system.

Modern AI networks combine different physical layers. Copper can connect a rack-scale NVLink domain; optical networks carry traffic across wider fabrics. Installation and maintenance remain part of the design.

NVIDIA illustration of co-packaged optical networking technology
Co-packaged optics · NVIDIA manufacturer illustration. [3]
NVIDIA Spectrum-X Ethernet Photonics switches
Spectrum-X Ethernet Photonics switches · NVIDIA. [3]
SCALE UP

NVLink & copper

NVIDIA documents a passive copper cartridge backplane linking compute and switch trays within DGX Vera Rubin NVL72. [1]

SCALE OUT

Fiber & photonics

Co-packaged optics brings optical components close to switch silicon. NVIDIA’s portfolio includes Spectrum-X Ethernet and Quantum-X InfiniBand photonics. [3]

KEEP IT SERVICEABLE

Physical infrastructure

Plan fiber pathways, bend radii, connector access, labels, spare capacity and inspection procedures. High bandwidth still depends on disciplined installation.

A cable-free tray design does not mean a cable-free data center. Rack backplanes, external networking, power and facility connections remain relevant.

03 / LIQUID COOLING

Capture heat
where it is created.

Cold plates and liquid circuits move heat from chips into the wider facility system. NVIDIA’s MGX discussion includes warm-water cooling scenarios; the installed equipment’s validated temperature, pressure, flow and coolant requirements must govern the actual design. [2]

NVIDIA conceptual free-cooling illustration showing an external dry cooler, heat exchanger, facility loop and technology cooling loop
NVIDIA free-cooling scenario. Climate and equipment conditions determine applicability. Image: NVIDIA. [2]
AT THE CHIP

Cold plates & manifolds

Design distribution, isolation, monitoring and service access around equipment requirements and leak-management procedures.

BETWEEN LOOPS

Coolant distribution

Coordinate heat exchangers, pumping, controls, water quality and redundant capacity across the technology and facility circuits.

AT THE FACILITY

Reject or recover

Evaluate dry, evaporative and hybrid rejection alongside compatible heat customers. Closed loops do not establish zero facility water use. [4]

04 / USEFUL HEAT

The next connection
could be outside the building.

Compatible temperatures, nearby demand, reliable backup cooling and viable commercial terms determine whether recovered heat can become a useful local resource. [4]

THERMAL SYSTEM / CONCEPTCOMPUTE → CAPTURE → TRANSFER → USE
Conceptual data-center heat recovery: server cooling transfers heat through an isolated heat exchanger and heat pump to a compatible building. Independent heat rejection remains necessary.DATA CENTERHEAT EXCHANGERHEAT PUMPHEAT CUSTOMERIndependent cooling requiredIsolated circuitsTemperature lift + electricityCompatible demand
Recovered heat Cooler returnIllustrative pathway. No Newark installation is represented.

Finland: progress with phases.

Fortum’s May 2026 update reports operating heat plants using ambient-air heat pumps and electric boilers. Microsoft data-center heat integration is planned to begin in phases in 2027. [5]

Newark: a proposal to evaluate.

Precipia proposes screening a compatible anchor building and a short heat connection first. No Newark heat customer, installation or savings commitment is represented as secured.

Explore the Newark approach

PRIMARY SOURCES / IMAGE CREDITS

Evidence behind the visuals.

  1. NVIDIA DGX Vera Rubin NVL72 hardware documentation
  2. NVIDIA Vera Rubin POD technical overview
  3. NVIDIA Silicon Photonics
  4. U.S. DOE / NREL: Best Practices Guide for Energy-Efficient Data Center Design, July 2024
  5. Fortum heat-plant update, May 6, 2026

Source review: November 16, 2026. Hardware images are credited to NVIDIA; the homepage infrastructure image is AI-generated. Precipia diagrams are conceptual illustrations. Vendor capabilities and availability should be confirmed for each procurement.

Put the technology
in the context of your project.

Talk infrastructure