MicroLink deploys high density liquid cooled compute inside energised industrial host facilities and returns the waste heat to the host. We intend to standardise a 1 GW programme, built as four 250 MW blocks, on Vertiv OneCore Reference Design 5.
Every line is taken from the Vertiv OneCore brochure (R12/26), the product catalogue, or Vertiv public reporting. Mark each one yes or no and leave a comment where we have it wrong.
The design, as printed
1,024 VR200 compute racks. Vera Rubin, not GB300. We read RD3 as the GB300 design, at 50 MW.
So RD5 at 250 MW is a campus of blocks, not one. 96 to 2,200+ racks per block. Where do the seams fall?
Eight compute halls against four, 2,208 racks against 1,104, plus one extra core hall of 3,588 m2 (38,627 sq ft).
Eight compute at 1,910 m2 (20,559 sq ft), 99.9%. Two core at 99.999%. Footprint 22,793 m2 (245,345 sq ft).
269,071 m2 (2,896,256 sq ft). What is the other 91%, and how much is heat rejection yard?
Chilled water, separate primary circuit for direct liquid cooling. RD3 runs 87 / 13, RD1 and RD2 run 80 / 20.
FIZ chiller for air, CoolLoop trim cooler for liquid. No host delivery loop in any of the five designs. Question one.
Electrical and thermal. Power smoothing at 0%. Standby generation sits outside OneCore scope entirely.
Racks at 400, 415 or 480 V three phase. 800 VDC fed racks supported, which is why we propose an 8 MW pilot.
Concurrently maintainable distributed redundancy. SFN N+1, PFN concurrently maintainable.
For RD3, RD4 and RD5. Only RD2 is marked Yes. How would a phased release structure work on your side?
237.5 MW of liquid against a 2,300 kW unit implies about 103 CDUs at N, before N+1. Please confirm the count.
Customer or GC provides the shell, Vertiv delivers the blocks. We deploy inside operating industrial facilities.
Engineered to order, which we understand. Our working assumption for Option 2 block scope is USD 1.3M to 1.8M per MW.
Rack power, derived not published
250 MW across 2,208 racks is 113 kW average per rack. Divided across the 1,024 compute racks alone it is 244 kW, but that assumes the 1,184 networking, storage and infrastructure racks draw nothing. Allowing them a realistic share puts VR200 compute racks in the region of 200 to 220 kW each. We could not find a published per rack figure and would like the real one.
Eleven questions MicroLink cannot answer alone. Mark each one yes or no and leave a comment.
The eleven questions
Will you co-engineer the third loop at the PFN boundary? OneCore ends at atmosphere. This is the one we need most.
What locks a slot, and what deposit schedule? We will pay early. Tell us the schedule and we work to it.
Four releases against a committed 1 GW. Our sites will not become ready on a fixed cadence. How does that sit with your loading?
Or only at block two? Two power architectures across a 1 GW estate is the worst outcome. We propose an 8 MW pilot.
Block one carries the NRE, priced honestly. How much comes back over blocks two, three and four, and what unlocks it?
Your method assumes a purpose built hall. We build inside operating facilities. What changes, and what does it cost?
How many CDU2300 in RD5, and the supply and return temperatures? Return temperature decides what the heat is worth.
Rack, CDU and skid loadings. We fit out on existing industrial floors, not new slab, so this is a real constraint.
Our working figure is 200 to 220 kW. Give us the real one. 244 kW only holds if the other 1,184 racks draw nothing.
We report ERE alongside PUE on a dry cooler first basis. We want your reporting and ours to reconcile.
Option 2 block scope, at or under USD 1.8M per MW, inside a USD 7.5M per MW facility envelope excluding compute.