100 kW-Class Racks

Overview#

For two decades, 'high-density' meant 15 kW per rack and 'extreme' meant 30 kW. Both numbers were set by what conventional perforated-tile air cooling can dissipate from a 600 mm × 1200 mm cabinet at reasonable supply-air conditions (ASHRAE A1/A2 envelope). Anything beyond 30 kW required heroic engineering — cold-aisle containment plus chimney rear doors plus tightly controlled bypass — and even then it topped out around 40-50 kW.

AI infrastructure broke those numbers. An HGX-H100 baseboard with eight 700 W GPUs dissipates 5.6 kW from the GPUs alone; the full 8U server lands around 10 kW; twelve such servers in a 48U rack already hit ~120 kW. NVIDIA's GB200 NVL72 — the rack-scale Blackwell platform — packages 72 GPUs and 36 Grace CPUs in a single rack drawing roughly 120 kW continuous. AMD MI300X and MI325X HGX-class boards land in the same band. 100 kW per rack is therefore not aspirational; it is what current frontier compute requires to fit in a single cabinet without sharding a tightly-coupled NVLink/Infinity Fabric domain across multiple racks.

Every facility subsystem changes at this density: power feeds move from 32 A cords to 250-400 A busbar tap-offs, cooling moves from CRAH-fed air to DLC + CDU, floors move from 1,000 kg/m² to 1,500-2,500 kg/m², and network moves from 25/100 G ToR to 400/800 G optics with overhead NVLink Switch trays. This entry helps you size a 100 kW rack programme — power, cooling, structural, network, and capex — and understand the reference rack design Yobitel NeoCloud uses across UK and EU regions for every GB200 NVL72 and high-density H100/H200 deployment.

Specifications#

The envelope below sits in the middle of published OCP Open Rack v3, NVIDIA GB200 NVL72, and the major server OEM (Supermicro, Dell, HPE, Lenovo) reference designs. Verify each parameter with the specific OEM datasheet before committing to a build.

Architecture#

A 100 kW rack is not a 15 kW rack with bigger pipes — it is a fundamentally different topology where power, liquid, and network all enter from above and the room itself becomes thermally neutral. The block diagram below maps the production envelope used by Yobitel NeoCloud and by hyperscale operators on the same density class.

• Power enters from overhead busbar — no floor PDU islands, no per-rack whip cables. Tap-off boxes can be moved or up-rated without re-cabling the row.
• Liquid enters from overhead headers, drops vertical manifolds at the rear of each rack, and connects to each server via dripless blind-mate quick disconnects (OCP ACS spec).
• Network enters from overhead trays. 400 G and 800 G optics dominate, with NVLink Switch fabric for GB200 racks confined to the rack itself (rack-scale switch domain).
• The room is no longer organised around hot and cold aisles — DLC captures 85-95 % of heat into the liquid loop, RDHx captures the chassis-air residual, and the data hall sees near-neutral exhaust.

100 kW rack — production topology
====================================================

Overhead (above row, 4.5-6 m clear ceiling):
  +----------------------------------------------+
  |  Busbar 400 V 3-phase, 800-1250 A run        |
  |    -> Tap-off box per rack (250-400 A)       |
  |  Liquid supply + return headers (DN50-DN80)  |
  |  Network trunks (400/800 G + NVLink Switch)  |
  +----------------------------------------------+
                 |          |          |
              power      liquid     network
                 v          v          v
  +----------------------------------------------+
  | Rack (47-52U EIA-310, or 1OU NVL72)          |
  |                                              |
  |  Top: rPDU A + rPDU B (busbar tap),          |
  |       quick-disconnects to vertical manifold,|
  |       ToR or NVLink Switch trays             |
  |                                              |
  |  Mid: 12-15 HGX servers (H100/H200/B200)     |
  |       or 1 GB200 NVL72 chassis stack         |
  |                                              |
  |  Bottom: in-rack CDU (40-80 kW, optional)    |
  |          or pass-through to in-row CDU       |
  +----------------------------------------------+

Adjacent to row:
  +----------------------------------------------+
  | In-row CDU (150-300 kW, N+1)                 |
  | Leak detection rope along base of every rack |
  | RDHx on rear of rack (for chassis-air        |
  |   residual, on same secondary loop)          |
  +----------------------------------------------+

Facility loop (out to dry cooler / adiabatic):
  Supply 30-35 C, return 40-55 C; free cooling
  year-round in temperate UK/EU climates.

Form Factor, Power, and Thermal Envelope#

100 kW racks preserve the standard 19-inch EIA-310 service ergonomics (or OCP 21-inch for hyperscale density) but reshape every adjacent envelope. The table below contrasts what a 100 kW class rack demands against the legacy 15 kW baseline and the GB200 NVL72 reference design.

Vendor Ecosystem#

The 100 kW rack supply chain is mature in 2026 — multiple OEMs at every layer with broadly interoperable form factors. The map below names the active vendors at the rack, busbar, CDU, and cabling layers; cold plates and quick-disconnects are covered in the [Direct-to-chip](direct-to-chip-liquid-cooling) entry.

Sizing — Racks per Cluster, Racks per MW#

Sizing a 100 kW rack programme is top-down: start with the cluster scale (number of GPUs / NVLink domains), back out to rack count, then to MW of installed capacity, then to dry-cooler tonnage and utility-feed requirements. The shortcuts below match Yobitel NeoCloud's actual cluster planning method.

• Rack-to-MW ratio: 1 MW = 8-10 usable racks (not 10) — the gap accounts for power-conversion losses, cooling overhead, and non-IT loads. Use a planning factor of 1.10-1.15 (rack nameplate × racks × overhead = MW).
• Floor-plate area: a 100 kW rack with overhead busbar, CDU adjacency, and service aisles needs ~2.5-3.5 m² of data-hall floor including its share of aisle. A 5 MW build is therefore ~150-170 m² of densified white space.
• Utility-feed lead time: in the UK, 11 kV connections typically take 6-12 months from application to energisation; 33 kV is 12-24 months; 132 kV intake can be 24-48 months. Plan the substation work before the first server arrives.
• Heat-rejection sizing: dry coolers at warm-water DLC (W3/W4) need ~1.05-1.15× the IT load in nameplate cooling capacity to handle the 99 %-design wet bulb. Add adiabatic spray for sites that exceed the design wet bulb 1-2 weeks per year.
• Network sizing: each 100 kW rack typically lands 72-96 × 400 G or 800 G port equivalents into the leaf fabric. At 5 MW (40-48 racks), that is ~3,000-4,000 fabric ports — sized into Spectrum-X / Quantum-X reference designs.

Cost#

Capital cost ranges below are USD per installed kW (the figure that matters at AI scale) and per rack-position (the figure facility teams quote). They include the share of cooling plant, electrical distribution, network fabric, structural reinforcement, and project costs; they exclude the GPUs themselves and the long-haul fibre tail.

• Per-kW economics: 100 kW racks cost more per rack than 15 kW but materially less per kW of installed capacity. At AI scale (1+ MW), per-kW is the figure that matters because the cluster size is set by GPU count, not rack count.
• Capex split: roughly 35-45 % cooling (DLC + CDU + dry cooler), 25-30 % electrical (UPS, switchgear, busbar, transformer), 15-20 % structural (slab, raised floor, ceiling, BMS), 10-15 % network fabric (leaf/spine switches, optics, cabling).
• Opex baseline: at $0.18-0.32 per kWh UK/EU commercial electricity, a 100 kW rack at 80 % utilisation consumes ~$126k-$224k per year in power. PUE 1.15 vs 1.40 swings that by 22 %.
• Yobitel NeoCloud pricing reference: H100 SXM5 capacity in 100 kW DLC racks lands at the published NeoCloud per-GPU per-hour rate (Omniscient Compute exposes the live USD pricing per region) and absorbs the rack-level capex/opex into the per-GPU rate.

Migration — From 30 kW Air-Cooled to 100 kW DLC#

Most operators arrive at 100 kW from either a 15-30 kW air-cooled brownfield or a clean-sheet greenfield. The migration path matters as much as the steady-state design — the table below captures the dominant patterns.

Pitfalls and Operational Notes#

Most 100 kW rack failures are not silicon failures — they are facility failures concentrated at the power-feed coordination, cooling commissioning, and network top-of-rack stages. The discipline below is what separates a first-time-installer schedule slip from a stable production cluster.

• Utility-feed lead time: getting 100 kW per rack ×40 racks (a 5 MW pod) requires utility coordination that takes 6-24 months in most UK and EU markets. Apply for the connection before the rack OEM PO is placed.
• Cooling commissioning: every cold plate, every quick-disconnect, every manifold connection must be pressure-tested at 1.5× working pressure for 24 hours and chemistry-sampled before GPUs energise. Skipping commissioning is the most common first-year incident cause (see Direct-to-chip entry for full leak-handling discipline).
• Hot-spot mismanagement in mixed-density rooms: when one rack is 100 kW DLC and adjacent racks are 15 kW air, the air-side thermal model can break in surprising ways — bypass airflow goes the wrong direction, return-air temperature spikes, and the air-cooled racks throttle. Segregate by density class or use CFD modelling.
• Network top-of-rack congestion: 72-96 × 400/800 G optics per rack produce real cabling complexity. Plan structured cabling, labelling, and slack management at design time; once the rack is populated, recabling is a multi-hour outage.
• Floor loading verification: a GB200 NVL72 arrives as one ~1,500 kg unit on castors. Verify slab loading along the entire delivery path (loading dock, corridor, data-hall doorway, raised-floor stringer) before the truck arrives. Re-routing a delivery costs days.
• Service ergonomics: services that were at the back of the rack now compete with manifolds, busbar tap-offs, and rear-door HX. Document service procedures for every common operation (server swap, cold-plate replacement, optic swap) before production.
• Power-feed phasing: balance the load across the three phases at the rack PDU level, not just at the row level. Unbalanced rack-level loading causes neutral-current heating that does not show up on aggregate metering.
• Acoustic environment: 100 kW racks are quieter than equivalent 15 kW racks (DLC removes most fans) but the row-level airflow noise from RDHx and in-row cooling still requires hearing protection at sustained exposure.
• First-deployment learning curve: a customer's first 100 kW rack deployment typically runs 15-25 % over benchmark on schedule and capex due to commissioning rework, runbook development, and ops-team training. Plan a 30-day shakedown period before scheduling production workloads.

Where 100 kW Racks Sit in the Yobitel Stack#

Yobitel NeoCloud's reference rack design is a 100 kW DLC-cooled position. Every GB200 NVL72 deployment in NeoCloud lands on this design natively; high-density H100/H200 pods (12-15 SXM5 servers per rack) use the same envelope. UK and EU sovereign regions deploy this rack class under NCSC OFFICIAL alignment, with the rack-level capex/opex absorbed into the per-GPU-hour rate published in Omniscient Compute.

For customers who want to consume the density without building it, NeoCloud is the lift-and-shift path. For customers building their own facility, our managed-ops team has commissioned 100 kW racks under both greenfield and brownfield-retrofit constraints and can hand over a complete commissioning runbook, leak-detection coverage map, and 30-day acceptance criteria as part of the Professional Services engagement.

References#

100 kW rack design crosses several authoritative reference sets. NVIDIA's GB200 NVL72 architecture brief defines the rack-scale platform; OCP Open Rack v3 standardises the 21-inch IT space and 48 V DC busbar; ASHRAE TC 9.9 governs the thermal envelope; Uptime Institute publishes high-density deployment guidance and Tier classification. A serious build cross-checks all four.