For fifteen years a data-centre rack was a 5–10 kW problem. The first liquid-cooled GPU racks pushed that to 40 kW, then 80 kW. NVIDIA's GB300 NVL72 — the rack-scale Blackwell Ultra system shipping today — draws roughly 130–140 kW. That is already beyond what most existing Australian colocation halls can deliver to a single cabinet.
What the next two generations look like
| Platform | GPUs per rack | Rack power | Availability |
|---|---|---|---|
| NVIDIA GB300 NVL72 (Blackwell Ultra) | 72 | ~130–140 kW | Shipping now |
| NVIDIA Vera Rubin NVL72 | 72 | ~190–230 kW | H2 2026 |
| AMD Helios (Instinct MI455X) | 72 | ~200 kW class | Q4 2026 → H1 2027 |
| NVIDIA Rubin Ultra · Kyber NVL576 | 576 | up to ~600 kW | 2027–28 |
Two things move with each step. The obvious one is total power per rack. The less obvious one is how it is delivered: 415 V busway rather than cable whips, facility-water coolant distribution units rather than in-row units, floor loading measured in tonnes per cabinet, and structured cabling specified for 1.6 Tb/s ports rather than 800G.
Why 250 kW is the right design point for 2026–2029 builds
- It clears Rubin with margin. A 250 kW busway and CDU loop runs a Vera Rubin or Helios rack at 80–90 % of design, which is where you want steady-state operation.
- It is buildable with today's equipment. Direct-to-chip cooling at 250 kW per rack is a known engineering problem with multiple vendors; 600 kW is not yet.
- The 600 kW pathway is cheap to reserve and impossible to retrofit. Reserving busway routes, CDU plant space, pipe sizes and slab loading for 600 kW costs a few percent at design stage. Adding them to a live hall means shutting it down.
What it means for a customer
A hosting contract signed for B300 nodes today should be able to absorb the same customer's Rubin nodes in 2027 in the same row, without a facility upgrade. That is the difference between a hall that hosts a generation and a hall that hosts a customer.
QORINAI's Tier C halls are designed to a 250 kW rack with busway, CDU and structural pathways to 600 kW, and cabling and switch positions specified for 800G today and 1.6T next. Our hardware roadmap page sets out the platforms we design around.
Talk to the team that builds it.
Land and power, halls, hardware, GPU operations and the IntraLLM platform — one enquiry.