Question 1

What is liquid cooling and why do AI workloads need it?

Accepted Answer

Liquid carries roughly a thousand times more heat per unit volume than air, which is why racks above about 50 kW cannot be cooled by air at any sensible cost. Direct-to-chip cooling pipes coolant to cold plates on the processors; immersion submerges whole boards in dielectric fluid. Modern AI racks running 600 kW to 1 MW leave no practical alternative.

Question 2

What rack densities do GRECO campuses support?

Accepted Answer

Facilities are architected for the 600 kW to 1 MW class rather than retrofitted toward it, with direct-to-chip and immersion cooling and 800 VDC power distribution designed in from the start. Building for that envelope from day one avoids the stranded-capacity problem that afflicts air-cooled halls converted late in life.

Question 3

What is the difference between leasing compute capacity and a powered shell?

Accepted Answer

A powered shell is the building delivered powered, cooled and ready, where the tenant brings their own racks and hardware. Leasing capacity means GRECO fits it out and the tenant offtakes by the kilowatt. Powered shell suits operators with their own hardware standards; capacity suits those who want to skip the build entirely.

Question 4

How does liquid cooling make heat export possible?

Accepted Answer

Because it keeps the heat concentrated and at a useful temperature. Liquid returns from the chips at 45–48 °C, high enough for plate heat exchangers to lift it into the 55–65 °C district-energy range. Air cooling disperses the same energy into a large volume of warm air that is not worth capturing, which is why heat export and liquid cooling are effectively one decision.

Question 5

Can I offtake compute capacity without waiting for a grid connection?

Accepted Answer

That is the point of the model. Because GRECO generates power on-site behind the meter, campus capacity does not depend on a new utility interconnection, so an offtaker leases finished powered capacity in months rather than holding a shell through a five-to-seven-year queue.

Compute that arrives with its own power