Borealis builds turnkey liquid-cooled AI data centers. Here is the short version of the questions buyers ask most.
Borealis, by AGT-USA, delivers turnkey liquid-cooled AI data centers: best-in-class liquid cooling engineered around NVIDIA-class compute, from a single vendor. Its products are the Borealis HALO, a complete 2 MW AI data center delivered on wheels; the Vega AIO, a self-contained 500 kW data center in one unit; and Borealis Retrofit, the complete cooling plant installed inside a hall you already own.
A complete 2 MW AI data center that arrives on wheels. Compute, cooling, power, and controls ship factory-integrated across three to four road-legal truckloads and reach first power in 10 to 15 working days on a prepared pad. It rejects heat 100 percent dry with zero process water, runs at a design PUE of 1.10 annualized, and carries 14 NVL72-class racks, 1,008 GPUs, on day one. It is fully reversible and scales by replication in 2 MW blocks to clusters of 20 MW and beyond.
The defining architecture of the Borealis HALO: the enclosure itself is the heat exchanger. The long walls are full-length coil banks behind louvered intakes and the roof carries a low-speed fan array, so air comes in through the walls and leaves through the roof with no recirculation. Large coil area means slow, quiet fans, and vertical discharge keeps hot air from short-circuiting back in. That is how 2 MW rejects dry at a 1.10 design PUE.
It cools at the rack with a closed liquid loop instead of cooling the room with air. Direct-to-chip cold plates and rear-door heat exchangers pull heat off the servers, a CDU pumps the isolated loop, a chiller makes the cold, and the heat is rejected dry to outside air: through a rooftop condenser, or in the HALO through the Condenser Envelope itself.
A modern NVIDIA AI rack draws about 132 kW. A CRAC unit tops out near 100 kW and cannot cool even one such rack, and it derates further in heat. Liquid carries heat about 25 times better than air, which is why high-density AI needs liquid cooling.
The Borealis HALO goes from pad to power in 10 to 15 working days. The self-contained Vega AIO runs in as little as a week on site. For Borealis Retrofit, the first wing is live in 14 days and a full 96-rack hall in 8 to 10 weeks, versus roughly 270 days for a conventional multi-vendor build.
A self-contained AI data center in one in-line unit: three rack bays built for NVIDIA-class compute, rear-door heat exchangers, a CDU, and an onboard chiller, carrying 500 kW, about five times the cooling density of a CRAC in the same footprint. Only the condenser line leaves the unit.
Borealis Retrofit drops the complete Borealis liquid cooling plant inside a building you already own and power. It installs in wings: each wing pairs an 829 kW CDU bank, an in-hall chiller, and a rooftop condenser with bays for six NVL72-class racks, rear-door heat exchangers on every one. First wing live in 14 days, a full hall in 8 to 10 weeks. You bring the building, the power, and the compute; one vendor owns everything that cools it.
Retrofit scales in wings: 12, 24, 48, or 96 racks as 2, 4, 8, or 16 wings. Each wing carries 829 kW of cooling for six NVL72-class racks. A full hall reaches 96 racks, 12.7 MW, and 6,912 GPUs.
No. The HALO rejects heat 100 percent dry in every mode: zero process water ever, no evaporative assist, and no Legionella program to run. Vega AIO and Retrofit reject through rooftop air-cooled condensers with sealed coolant loops, no cooling tower and no evaporation.
Yes. Heat rejection is air-cooled and the loops are sealed, so no water is consumed, and the systems are rated to 50 C ambient. The HALO adds a second thermal wing for 45 to 50 C sites. That suits hot, water-scarce regions including the Gulf, where evaporative cooling is a liability.
65 dBA at 10 meters, with a 55 dBA attenuated option. The huge coil face of the Condenser Envelope lets the roof fans run slow, and slow fans are quiet: quieter than most county ordinances allow.
No. It requires no permanent structure, only a prepared pad, power, and fiber. When the job is done the modules lift back out and the site is restored in days.
Borealis systems are engineered around NVIDIA reference compute: GB300 NVL72 racks (72 GPUs plus 36 Grace CPUs, 132 kW per rack), DGX B300, and DGX racks. The HALO carries 14 NVL72-class racks on day one and ships 800 VDC provisioned and two-phase ready, so the envelope outlives at least two GPU generations. Borealis sells the system that cools and carries the compute; the compute belongs to the customer and slides straight into the bays.
AGT-USA Inc., a precision thermal engineering company in Corona, California, with 35+ years of industrial cooling heritage and thousands of systems deployed. Borealis is its AI data center brand.
Borealis is turnkey liquid-cooled AI data centers from AGT-USA: best-in-class cooling engineered around NVIDIA-class compute, one vendor.
Explore the interactive site → Get a quote →