Approach

Grid tie Generation Storage Shell

Site

Time to power Fleets Fig. 01 Contact

Powered shells for AI data centers

Power is the constraint. Everything else follows. Ultracomputation delivers energized, ready-for-fit-out shells in one to two months — behind the meter, in front of it, or both on one bus.

Contact Run time to power

Sized rack by rack for the hardware that is actually shipping

NVIDIAH100 HGX11.2 kW · 8 GPU node
NVIDIAB200 HGX13.6 kW · 8 GPU node
NVIDIAGB200 NVL72132 kW · 72 GPU rack
NVIDIAGB300 NVL72135 kW · 72 GPU rack
AMDMI300X11.6 kW · 8 GPU node
AMDMI355X16 kW · 8 GPU node
GoogleTPU v6e58 kW · 64 chip rack
Interconnection queues run in years, but the chips lose value by the month.
Ultracomputation energizes the shell in one to two months, so the fleet earns while the queue moves.
Grid tie

Take service on either side of the meter

A 138 kV utility service, revenue meter and main transformer when the utility can move. A behind-the-meter island when it cannot. Front of meter, behind the meter, or both landing on one 34.5 kV bus.

See the single line

quick turn to an energized shell

ServiceBTM · FTM · either side of the meter
Generation

Firm, dispatchable generation on site

Onsite generation behind the meter carries the load when the grid tie is late, curtailed, or not there yet. Firm capacity, dispatched to the rack, on the same bus as everything else.

Storage & shell

Ride through into white space

Storage on the bus rides through utility events and smooths generation, so the hall never sees the disturbance. Unit substations step down into a powered shell with white space ready for your racks — NVL72 at 135 kW today, whatever ships next.

Talk to us

uptime at the bus · the hall never sees the event

[Time to power]

What a month of delay costs

Pick the hardware, the site and the deal. The model sizes the fleet from the power, then prices every month the site sits dark — in today’s money, with nothing hard coded to make time win.

Three fleets, one model

Open the model

Rack power is quoted anywhere from 120 to 140 kW depending on vendor. That 17 percent spread moves the answer more than the power price does.

NVIDIA GB300 NVL72, 25 MW grid · $6.00/GPU-hr
Run this fleet GB300 · 135 kW

Stat

NVL72 racks energized from 25 MW of grid service at PUE 1.15

Stat

value destroyed by three months of delay, in today’s money

Deployed everywhere and quoted everywhere. The most reliable numbers in this list — and at $1.80 a GPU-hour, a fleet that is already underwater on its own opportunity cost.

NVIDIA H100 HGX, 25 MW grid · $1.80/GPU-hr
Run this fleet H100 · 11.2 kW node

Stat

GPUs on the floor from 25 MW of grid service at PUE 1.30

Stat

value destroyed by three months of delay, in today’s money

Google does not sell TPUs as hardware, so there is no merchant price for a rack of them and no published chip power. Every figure here is indicative. Replace them with your own before quoting anyone.

Google TPU v6e, 25 MW grid · $2.40/chip-hr
Run this fleet TPU v6e · 58 kW

Stat

chips energized from 25 MW of grid service at PUE 1.10

Stat

cash out while dark over three months: interest and rent

[Fig. 01]

One 34.5 kV bus. Utility service in front of the meter; generation and storage behind it; unit substations stepping down into the shell. Representative, not to scale, rev 04.

Run time to power
Fig. 01Representative single line
Grid tieGenerationStorageShell